April 4, 2024 • 38 mins
Jonathan Hurst is a professor at Oregon State University, and co-founder and chief robot officer at Agility Robotics. Jonathan's problem is this: How do you design a robot that can walk and do useful tasks that companies will pay for? The solution begins with trying to understand how birds walk.
See omnystudio.com/listener for privacy information.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:15):
Pushkin. So how long have you been trying to make
a robot walk?
Speaker 2 (00:25):
It's been my entire career, starting from why I went
to college in the first place.
Speaker 1 (00:29):
Why why that particular problem? Why is that your life's work?
Speaker 2 (00:35):
You know, there's few things more interesting and more dynamically
complex and more elegant than the way animals move in
the world. And to be able to get machines that
can move that way, they can interact physically with the
world the way humans and animals do. What a fun
and interesting thing to work on for a career.
Speaker 1 (00:57):
I'm Jacob Goldstein and this is What's Your Problem? The
show where I talk to people who are trying to
make technological progress. My guest today is Jonathan Hurst. He's
a professor at Oregon State University and founder and chief
robot officer at Agility Robotics. Agility Robotics has built a
robot called Digit. Digit looks kind of like a person.
(01:19):
It walks around on two legs. It's got this flat,
rectangular head, and it has two arms that it can
use to pick stuff up. Jonathan's problem is this, how
can you make a walking robot that can do useful
work and that companies will actually pay for. Jonathan says
that robot Digit is already being tested out in warehouses
(01:40):
in the real world.
Speaker 2 (01:43):
We are deploying robots with cost We have two announced.
We've announced a couple with Amazon and with GXO. You know,
you place an order and Digit handles that order as
part of the workflow that has happened is happening right now, and.
Speaker 1 (01:56):
So specifically, what are your robots doing well? First at Amazon, Yeah,
the first use case or the first class of use
cases that works for us is picking up these plastic
tote plastic bins and putting them somewhere else. And then
in warehouses, Yeah, but anywhere warehouses, in logistics, in manufacturing,
(02:18):
you know, the whole environments like that that are a
bit structured. They're kind of like there's these islands of
automation they call them, you know, where one machine is
putting things in a bin, another machine sorts bins and
sends them different parts of the warehouse. Right now, sometimes
a person will stand there, the robot will tell them
which bin to pick up, and then all they are
basically is a manipulator for the robots system. They pick
(02:41):
up the bin and put it on the conveyor belt
and wait for the robot to tell them the next
thing to do. And it's really hard to hire people
for that. There are a lot of open jobs in that,
and so it's kind of a perfect place for Digit
to walk in in this relatively structured first use case. Now,
digits of course going to evolve towards picking up boxes, depalletizing,
(03:01):
loading and unloading you know, tractor trailers and eventually getting
out to things like retail and stocking shelves and you
know sticks in hospitals carrying things around and eventually become
a consumer product. So that's where you are today and
where you want to get. I want to talk now
about how you got here. Sure, and there's this really
(03:24):
basic set of things you had to figure out just
around how locomotion works, right, how people walk, also interestingly
and sort of surprisingly, how how birds like ostriches walk.
So I know there is this series of robots that
you built on the way on the way to Digit, right,
(03:45):
they were two before and then Digit and it seems
like going through those and what you figured out the
sort of key insight on each one is a really
nice way to get a kind of deeper insight into
how it works and what you had to understand to
make a robot that can walk great, just being a
really hard problem, right, Like there's like lots of robots
(04:05):
that we don't even sort of think of as robots arms,
and you know, self driving cars are arguably a kind
of robot and whatever, but like getting a robot to
walk is clearly a very hard problem.
Speaker 2 (04:16):
Yeah, Okay, So where I started in trying to understand
how to make a machine work is on the biomechanics,
try to understand how animals work. And a lot of
biomechanics is about specific muscles and muscle groups and joints,
and we're instead looking at holistically from a big picture,
what is the center of mass of an animal doing?
What are the forces happening on the ground. I did
(04:38):
this collaboration with Monica Daily at the Royal Veterinary College
and we looked at guinea fowl and ostriches and turkeys
and you know, a whole bunch of different sizes.
Speaker 1 (04:47):
Why birds. It's really interesting to me that you did that.
Speaker 2 (04:51):
Why because humans are weird? How many other bipeds are there?
You know, all the evolved bipeds, all of the extant
existing bipeds in the world.
Speaker 1 (05:00):
Everything that walks on two legs. Yeah, huh.
Speaker 2 (05:03):
Those are all theropods. They're all kind of more like
a bird than they are like or a monkey. They
have evolved far longer than we have. You know, we've
been out of trees for a very short period of time.
Compared to all of the other existing bipeds in the world.
They can all run much faster than we can for
you know, very efficiently using less energy insul So.
Speaker 1 (05:27):
Like an ostrich is a better model for just abstract
walking on two legs. Oh sure, I mean, yeah, I
love that, But I want to be clear.
Speaker 2 (05:36):
What we're not trying to do is study how it
does an Ostrich run versus how a human runs. What
we're trying to do is study what is a fundamental
truth between all animals in how they run, so we
can try and weed out things that have to do
with the size of the animal, or things that have
to do with where exactly the ankle joint is or
how long this joint is or that one. We want
(05:56):
to know what is the same amongst all animals to
walk and run.
Speaker 1 (06:01):
It's like the Platonic ideal of bipedalism. The sort of
the fundrationing theory of it. Yeah, that's right.
Speaker 2 (06:07):
Yeah, And so biomechanists have been looking at this since
the seventies and thinking about this in terms of spring
mass models of locomotion. And it was only in like
the two thousands sometimes that I think it was Hart
McGuire and Andre Seifarth put together a paper that showed, hey,
this spring mass model reproduces all of the behavior we
see for walking and running and transitions between these gates.
Speaker 1 (06:27):
When you say, when you say spring mass model, that
sounds big and exciting, But just to be clear, what
do you mean when you say spring mass model.
Speaker 2 (06:35):
I mean a mathematical representation of a pogo stick. Go on, Okay,
So a pogo stick is basically the simplest thing that
can run, and a kangaroo looks a lot like a
pogo sticky. Now, if you just stick a pogo stick
on each leg, now you're bipedal running, you know.
Speaker 1 (06:52):
Okay.
Speaker 2 (06:53):
And then if you add a whole bunch of complexity
to it, you have heel toe and you know, knee
joints and all this other stuff. But if you really
boil it down and try and make it as simple
as possible, you get to some pretty basic math models
that do represent how the progression of the center of
mass of the animal moves, and how the ground reaction
forces progress and so on.
Speaker 1 (07:11):
Right, Okay, so if I picture just like a lump
of mass on top of two pogo sticks, you got it.
I'm kind of in the right.
Speaker 2 (07:18):
Okay, you're absolutely okay. So that's a roughly a math
model that at least gives you a very good concept
of how do all animals run, horses, ghost, crabs, humans, ostriches, whatever.
Speaker 1 (07:28):
Right, So this paper comes out, this paper that says,
think of a lump of mass on top of two
pogo sticks. You you know about it because you're in
this world? What do you What effect does it have
on you? What do you do as a result of this.
Speaker 2 (07:42):
What I was looking at? And here's the argument at
the time. Do these spring mass models simply seem to
describe the things we're observing, you know? Or is it
describing core physics of how it works? Like, in other words,
if you build a spring mass model and a policy
that works, is it going to make a robot stabilize?
Or is it simply like a picture that kind of
(08:04):
looks like what the animals are actually doing.
Speaker 1 (08:05):
So like if we actually do the lump of mass
on top of two pot sticks as a robot, will
it work?
Speaker 2 (08:11):
Yeah? And so look the question is, you know, how
do you control these things? And then how does that
translate into walking and running which had never really been
done before that way, you haven't done this continuous transition
between walking and running and changing the speed and everything else,
and it's unknown how to stabilize that over all kinds
of terrain. So that's why we built Atreus. That's why
we built the robot Atreus.
Speaker 1 (08:30):
And so what is Atreus?
Speaker 2 (08:32):
So Atreus is it is a bipedal robot.
Speaker 1 (08:35):
Okay, what does Atrius look like?
Speaker 2 (08:38):
So Atrius was on the Late Night with Stephen Colbert
and he described it as a microwave on stilts.
Speaker 1 (08:44):
Okay, that's good. So it doesn't it doesn't look humanoid
at all. It looks maybe like a dancing alien or
like a moon lander or something, but not humanoid.
Speaker 2 (08:58):
Yeah, absolutely not. So it doesn't look like an animal
in any way. But it's designed entirely to be the
math model of what we see an animal running. And
the name Atrius is an acronym for assume the robot
is a sphere, right. The whole idea is the robot
is this simple math model.
Speaker 1 (09:14):
Where it's just some mass in the middle and then
some very light springy legs.
Speaker 2 (09:18):
That's it. And so what we ultimately showed with Atreus
it's the first robot ever to reproduce human walking gate dynamics.
The robot walks across the force plate, a graduate student
walks across the forest plate. Looking at the data, you
can't actually tell a difference.
Speaker 1 (09:32):
Huh. So, like from the plate's point of view, it
feels the same whether you're robot or a human is
walking across.
Speaker 2 (09:38):
That's correct.
Speaker 1 (09:39):
So it's not true for earlier robots that looked like
they were walking like humans.
Speaker 2 (09:43):
They would look the same. But if you looked at
the dynamics of it, if you looked at the ground
reaction forces, they differed quite a lot.
Speaker 1 (09:49):
Why is that important? Why is that important?
Speaker 2 (09:52):
It's just one symptom, right to show that, hey, we've
actually captured the physics here. But the other symptom that's
important is we are able to walk and run outdoors
over all kinds of terrain without any sort of perception.
The robot can handle amazing obstacles and it would just
soak them up, you know, going over potholes, going from
grass to pavement, going over big pieces of plywood we
(10:13):
would throw in its way.
Speaker 1 (10:15):
And you're saying it didn't do this because it had
like a clever brain. You're saying it was just the
physics of the brainless machine was able to deal.
Speaker 2 (10:23):
No cameras on it, It had no awareness of the environment.
It was a very simple spring mass model, very very
simple control that did nothing but try to balance that.
And it was able to just absorb all these kinds
of disturbances and just keep on going.
Speaker 1 (10:36):
That sounds like a big deal. I agree.
Speaker 2 (10:39):
I'm very excited about this, right. That is the point
that we decided we were going to found Agility Robotics.
We said, you know, this was a mission for like
years and years and years. That is why I became
a professor, is to say, my goal here in academia
is to show that this spring mass physics is real
and really make sure we understand that well. And if
(11:02):
we can show that and prove that and with this
ATREUS project, we then can take the next steps. Right,
So that's what we did. It was a scientific kind
of breakthrough. But the machine could only walk and run.
It couldn't stand, it couldn't turn. You know, it breaks often.
If it ever fell, it would just be completely destroyed.
(11:23):
So it's not a productive, useful machine. It's a science demonstrator.
Speaker 1 (11:27):
Okay, So you have a TRIOS that is academically it
kind of intellectually a breakthrough, right, but nobody's going to
buy it to do anything. It's not useful in a
practical sense. What do you do next?
Speaker 2 (11:44):
Okay? So at that point we say, hey, you know,
we understand a really significant portion of what does it
take to make a robot that can go where people go?
And what an opportunity? This is? What are all the
things that robots are going to be able to do
when they can be coexisting with humans? Right? And the
barrier to moving forward on that now is more about
(12:06):
execution on engineering and execution on building a use case
in a business around it.
Speaker 1 (12:12):
Because you feel like you've solved the sort of core
threshold technical problem of a robot that can walk in
unfamiliar environments and not fall down.
Speaker 2 (12:21):
Correct, we have the foundation layer of leged locomotion. Now,
sure we'd like to add on perception to it so
that it can handle stairs effectively and things like that
and intentionally handle obstacles. We need to do the engineering
of the electrical system and the hardware system that still
captures the same physics but can take a beating, can stand,
can steer, can start to do things right. So we
know we're going to need to build a robot with legs,
(12:43):
with manipulators, with sensors. Okay, because so we're kind of
going down a path now where we want to take
the exact same first principles approach to how do we
build a machine that can manipulate things in a human
world and get around in it and interact with people.
Build a human centric machine. So the first step to
(13:03):
doing that was designing a robot that we could sell
to other researchers to continue the work on the leg
leggs as we then worked towards you know, arms and manipulators,
and that was Cassie, our first robot that Agilia Robotics sold.
Cassie added the ability to stand in place because it
had ankles, and it had the ability to steer because
you could turn the legs. But more importantly, it was
(13:26):
much more compact and extremely robust, so this robot can
fall hard on concrete and you just pick it back
up and it can get going again.
Speaker 1 (13:34):
And so, Cassie, I'm looking at a picture of it now.
It basically looks like a pair of Ostrich legs. Yeah,
it looks like an Ostrich. Does an Ostrich have a wasist?
I don't know an ostrich from the waist down, No.
Speaker 2 (13:47):
It doesn't know. Yeah, the pelvis is stationary and a
bird fixed rather than a human pelvis, which is mobile.
Speaker 1 (13:54):
So I know it's not technically correct to say that
an Ostrich legs bend backwards, but it looks that way, right.
Speaker 2 (14:00):
Yeah, Their thigh is very short, their knee is up
next to their body, and what you perceive as their
knee is actually their ankle.
Speaker 1 (14:07):
In designing this robot, how do you get to legs
that look like Ostrich legs? Like, it's like convergent evolution,
you know, convergent evolution maybe hopefully. Isn't it the case
that like cephalopods that like whatever squids have eyes like
our eyes, but they evolve totally independently. Is this like that?
Speaker 2 (14:26):
There's a lot of examples of convergent evolution. We can
only guess, right because we don't necessarily know, and a
scientists only hypothesize the evolutionary pressures that cause animals to
be the shape that they are.
Speaker 1 (14:36):
But the pressure that led you you didn't say let's
make legs that look like Ostrich eggs. You just did
a bunch of math and you wound up with legs
that look like Ostrich legs.
Speaker 2 (14:44):
That is correct, and there are a bunch of features
on our robot that have gone down a similar path.
And I actually love that because when we end up
you know, blank sheet, pursue all of the different configuration
options and say okay, here's what we think is the
optimal choice, and we say, wow, that looks just like
a person, or that just looks like a bird or something.
It's actually really good. It means we're probably on the
(15:04):
right path.
Speaker 1 (15:05):
Yeah, it's it's exciting in a way, right, like like
you do a bunch of bath and then suddenly you
look up and you see an Ostrich.
Speaker 2 (15:13):
But it won't always be that way, right because we're
not using muscle and bone. We're using aluminum and you know, electricity,
it's a whole different thing.
Speaker 1 (15:19):
In a way, it's surprising that it is right, like
you would expect that it wouldn't look at all.
Speaker 2 (15:24):
Familiar, but there are clear differences.
Speaker 1 (15:26):
I suppose I'm projecting this. It's like, whatever is the
Ostrich version of anthropomorphizing, right, I'm Ostrich pomorphizing?
Speaker 2 (15:34):
Yeah, you got it. And like you said, it's like
a cartoon version of an Ostrich leg maybe.
Speaker 1 (15:38):
Yeah. Okay, So you've got this robot that is looks
to my little brain like a pair of Ostrich legs.
It's just a coincidence because it just turns out to
be the best way to build a couple legs. And
do you sell it to other academics? What do you
do it?
Speaker 2 (15:52):
Yeah? We sold it to some of the top universities
in the country and the kind in the world.
Speaker 1 (15:59):
So Cassie is a robot that academics can experiment with
and learn from. But it doesn't have arms. It isn't
built to do useful tasks. In a minute, Nathan and
his colleagues build a robot that does have farms, that
can do useful tasks, and that companies like Amazon are
testing out in the real world right now. The latest
(16:30):
robot from Jonathan's company is called Digit. It has two legs,
two arms, It walks around, it picks stuff up, and
it looks kind of like a person. But Jonathan says
he and his colleagues didn't set out to build a
robot that looks like a person.
Speaker 2 (16:44):
So yeah, so like, I'll take you down one of
these thought processes that ended up looking like a person. Okay,
We said, okay, we need to do some sort of
inertial control of this thing because the robot can't turn
very well. It's got little feet, and so when you
try to turn aggressively, its skids right, okay, and if
you look at any other bipads, this is one of
(17:06):
the reasons wings evolved. It's because they're running in the stick,
going out to catch the air to help them in
maneuvering and turn it. You can go in a straight line,
but figuring out how to maneuver quickly as hard when
you've only got a little foot on the ground. You know,
udropeds can really plant all four feet and twist and
apply big torchs on the ground, and a biped not
so much. You've got one foot at a time. How
do you change your orientation?
Speaker 1 (17:27):
Right?
Speaker 2 (17:28):
So we looked at like putting a gyro on board
reaction wheels or tails or things like that, because you know,
we ruled out reaction wheels because that's just a big
thing of brass that has to be mass you don't
want in a robot like this. We thought about tails,
and you look at any animals with tails, bipeds in particular.
Typically that's to control the pitch. In other words, you'd
(17:48):
leap off the ground, and then you want to reorient
your body and your feet so that your feet land
forward rather than you know, just tumbling in the air. Okay,
but we don't want to control pitch. We want to
control y'all. We want to steer the robot. So what
we kind of came up with is that the best
way to do that would be a pair of tails
that are symmetrical on the front of the back or
(18:10):
the side of the side of the robot, so that
when you swing those tails, you're controlling exactly down the
center line of your yaw, and so that just happens
to be where our arms and shoulders are. You know,
this bilaterally symmetrical pair of tails that can inertially actuate
you around the center and allow you to steer.
Speaker 1 (18:27):
So that's the sort of intellectually elegant version of how
you get to a robot that looks like it has arms.
You're saying, in fact, in terms of the way you
thought of it, it's a pair of tails that happened
to sit where our arms sit. I mean, presumably there's
a simpler version, which is, you want to build a
robot that can do stuff in a world that is
built for humans, and having arms would be useful in
(18:50):
that context as well. Or no, was it true?
Speaker 2 (18:53):
Member? Yeah, what we're not trying to do is make
a humanoid robot that looks like a person. What we're
trying to do is on first principles, understand exactly why
we do each thing.
Speaker 1 (19:01):
But are you really just adding the arm so that
it can turn when it's walking, Like I feel like
that's what you're saying, and I'm skeptical.
Speaker 2 (19:08):
Yeah. It's also that there were three other reasons why
the arms should be there that were all aligned. They
were not compromises where you know, putting the arms here
is better for one thing or another.
Speaker 1 (19:18):
Huh.
Speaker 2 (19:19):
It's that. Hey, if you go on just the just
the path of I want to improve my locomotion capability,
you land at the solution of where the arms are.
If you go down the path of this robot's going
to fall, and we know that it can't just fall
and land on its torso it'll break things. How do
we put manipulator's arms something on it so that it's
going to be able to catch itself when.
Speaker 1 (19:38):
It falls oriented? Okay.
Speaker 2 (19:41):
And then the third one, of course, is picking up
things right manipulation in the world and being by manual
in your manipulation so that you can basically a giant
pincher grass. That's how you pick up boxes and tots
and all these things you want to move. That's also
the best place for them. So basically, you just set
out to build a machine that could go where humans
go and pick up things that are the size that
(20:03):
humans pick up, and from first principles, with your eyes closed,
you wound up with a thing that looks like a guy.
Speaker 1 (20:10):
Absolutely so okay, so this is how you get to
digit the robot that you're now building and selling to people.
Speaker 2 (20:17):
That's right. So now we're taking this transition right now
as a company from that very intellectual and first principles
approach that I shared with you to now working with
the customer understanding what their use case is, writing down
the sets of requirements, like you know, the temperature ranges,
the you know, weights of all the things that you're
going to be able to pick up, the safety requirements,
(20:39):
you know. And it's a massive list, hundreds of things
in a list to write down the requirements documents so
that we can engineer a system that is a product. Yeah,
very different from designing a robot that can do cool things.
We're engineering a product, and that's the pivot that our
company is in right the second.
Speaker 1 (20:54):
And like I imagine for you personally, that must be
a significant shift, right if you spent whatever twenty some
years in the kind of abstract academic world of like
let's build the thing that works and know think deep
thoughts to like let's mass produce a product that people
will pay us for. That's quite different.
Speaker 2 (21:15):
Oh, it's fundamentally different. It's a whole different way of thinking.
In fact, I changed my title to chief robot Officer, right,
and we hired Melanie Wise as our new chief technology officer.
She comes out of FET Robotics. She was the founder
there and recently sold that company and they were deploying
thousands of robots in logistics squarehouses. And she is an
(21:36):
absolute expert on understanding customers and product and creating a product.
And what we've done is we've shifted our organization. So
you know, Melanie is in charge span of that whole
product side of the organization and the engineering to make
a product. I'm now leading the innovation team.
Speaker 1 (21:52):
So you get to keep doing kind of the stuff
you've been doing.
Speaker 2 (21:56):
The things I'm good at.
Speaker 1 (21:57):
Yeah, so what is the frontier on the innovation side.
What are you trying to figure out next?
Speaker 2 (22:02):
It is fundamentally hardware that enables the kinds of physics
that we want to achieve, powered by some of these
new AI tools. You know, we're getting to a point
now where some of these tools will allow us to
create behaviors and create things that as an engineer we
don't know how to model. Huh, And that's super interesting.
(22:24):
So instead you're describing the symptoms of it, and then
the system, the learning system, figures out how to make
that happen.
Speaker 1 (22:31):
Amazingly different than what you've been describing. You've been describing
of like, let's think of you know, first principles, just
the physics of the universe, and from that build a machine.
And now you're talking about you know, an era when
possibly you'll be able to ignore all of that and
say to the AI, you figure it out, here's what
I want to do.
Speaker 2 (22:50):
Well, let me put some caveats on that.
Speaker 1 (22:52):
Yeah, that sounds ridiculous when I say it that way.
Speaker 2 (22:55):
Well, remember that the AI has to operate on a
piece of hardware. Yeah right, And so that piece of
hardware we still have to engineer and design to be
able to achieve the physics that we want to achieve.
Speaker 1 (23:05):
Though you could say to an AI, here's what we
want to do. What should the hardware look like? Maybe
in one hundred years you think one hundred, one hundred,
who knows one hundred.
Speaker 2 (23:14):
It's fine, fine, three years, you know, future, Not today.
Speaker 1 (23:18):
Not certainly not. So what specifically are you doing today?
Are you taking this robot that you have digit and
seeing if you sort of put an AI layer in it,
on it near it? What can you do?
Speaker 2 (23:27):
Is that what's happening all of the above, So, you know,
on the hardware side, So there's a lot on the
hardware you do just to make it even possible for
the AI to learn. We're building then a whole architecture
of a digital twin so that you can learn things
in simulation first, and then you know, transfer from sim
to reel.
Speaker 1 (23:43):
A digital twin is basically making a version of the
robot that exists as software that exists virtually.
Speaker 2 (23:49):
Version of the environment as well that the robot operates in,
so that everything can be done, you know, decades of
experience on the robot can be done in hours of
time through the you know, et cetera.
Speaker 1 (24:00):
So the digital twin is allowing you to try and
generate data to train the AI. Is that that's what's the.
Speaker 2 (24:06):
Source of the data, right, A lot of language models
and things like that are based on data from the internet. Well,
I don't think that that's feasible for robot control because
the physics of the hardware is so unique. So even
if you're trying to teleoperate this thing, you've got this
weird translation between what a person would do to then
the robot trying to mimic that, which is probably not
the dime of the right.
Speaker 1 (24:25):
Dynamic robot doesn't actually walk like a person, even if
it looks like it's.
Speaker 2 (24:29):
Right, that's right, everything's sort of different internally about it.
How it would control itself.
Speaker 1 (24:34):
What what what are you worried about? Like you what
could go wrong and how are you trying to get
it not to go wrong?
Speaker 2 (24:43):
So before I talk about what I'm worried about, let
me tell you what I'm excited about. Fair We had
ten of these Cassie robots out in the world, and
so researchers all over the place for for years and
years are working on various kinds of controllers. When we
were able to successfully get a learned policy working on Cassie,
were to run a five k across campus, we were
able to the world record in the one hundred meter dash.
(25:04):
We were able to.
Speaker 1 (25:05):
Do for a robot. Yes, and when you say learned,
you mean developed with machine learning as opposed to in
the old way. Is that what you mean?
Speaker 2 (25:12):
Correct? It's an entirely machine learned policy that was learned
in simulation and then put on the thing.
Speaker 1 (25:18):
What's a definition of a policy? As you're using the word, I.
Speaker 2 (25:22):
Mean, a policy is just a bunch of math that
takes as input all of the sensors and then spits
out numbers that describe the torques you should apply to
the motors.
Speaker 1 (25:34):
Okay, so it's sort of if this happens in the world, robot,
you should do this set of things.
Speaker 2 (25:40):
Yeah, you're based on you know, but you know it's
like the input from thirty encoders or one hundred different sensors,
all of that complex input.
Speaker 1 (25:47):
But if this is complicated and the then that are complicated, Yeah.
Speaker 2 (25:51):
No, you're rreat you're right about that. Yeah, it's an equation.
Speaker 1 (25:53):
Good. So the machine learning basically made the robot work
way better.
Speaker 2 (25:58):
It made the robot work much better, and even more importantly,
once the pipeline was in place, we can learn new
skills and learn new policies much faster with much less
engineer time. How we can get there faster and have
higher performance using learning approaches to control.
Speaker 1 (26:16):
It's like a productivity like supercharger. It just makes everything
go much faster and more efficiently.
Speaker 2 (26:21):
Absolutely, it's a new tool. It's amazing, that's right. Okay,
So what am I worried about? Right? What am I
worried about? I'm worried that this is one of those
kind of black Swan events. Right, this is one of
those things that changes everything, and everybody doesn't exactly know
what all the implications are yet and what are the
(26:41):
you know, the right paths forward, and so everybody's trying everything.
Speaker 1 (26:45):
And this being basically a useful bipedal robot like it,
it could be hugely important in ways that we don't understand,
and there could be unintended consequences.
Speaker 2 (26:55):
What I actually mean is that this new realization that
we can use learning policies to control dynamic robots and machines, yeah,
means that the entire way it all robotics controls people
have been doing robot controls before is not as relevant.
(27:17):
And this new tool that nobody really understands that well
yet is clearly the future of how it's going to work.
Speaker 1 (27:25):
So what are I mean? I understand that if it's
really a black swan, you don't know what's going to happen,
because if you knew, it wouldn't be a black swan,
But like, what are you thinking of? What could it mean?
You know, plainly, bipedal robots are a very powerful tool,
and you could imagine malevolent uses of them, right, I guess, So,
(27:46):
I guess we've already got drones. Right, It doesn't matter
whether the robot that kills you looks like a dude, right,
those things already exist.
Speaker 2 (27:52):
In fact, a humanoid robot is probably the least effective
way to do that I think, honestly, my biggest worry
about AI in general has a lot more to do
with its ability to influence people, its ability to model
people's feelings and that kind of thing.
Speaker 1 (28:06):
So that's more like using large language model for kind
of personalized misinformation or that.
Speaker 2 (28:11):
Stat it that's the biggest threat building robots that are
going to like take people. I don't know, I just
don't see it fair.
Speaker 1 (28:19):
I mean, I'm kind of tired of talking about technological
unemployment because the robot looks like a person. It makes
me feel like we should touch on it. Do you
want to just speak for a moment to the prospect
of technological unemployment?
Speaker 2 (28:30):
Sure, I'll say our entire business model is centered around
the number of unfilled roles that exist in the logistics environment.
It's not centered around how it's going to be less
expensive than human Labor's centered around how they actually in
geographic locations do not have enough people to provide the
service that they're providing, you know, the way they're doing
(28:51):
it now, there's no way forward for improved logistics and
getting your things in one day and you know, all
the stuff that people really really want. There's no path
forward to do that with more human labor doing it,
it must be automated in a significant way. So you know,
that's our whole business model is based on unfilled roles.
Speaker 1 (29:11):
What is your like happy version of the future in
whatever number of years seems like the right number of.
Speaker 2 (29:17):
Man These robots are actually safe and smart enough in
order to do a lot of useful things in the world.
And the relationship with people is kind of like a
service animal. And you know, these robots are everywhere and
delivering all the packages to your door, and you know,
being a telepresence device that you can easily log into
and keep in touch with people, and you know, in
(29:38):
a lot of warehousing and stocking and you know, doing
the dull, dirty, dangerous to classic three d's of robotics.
It's we've always wanted that. It's all about improving the
quality of life. It's all about letting enabling humans to
be more human, letting people do the things that are
that they want to do that involve the social interaction
and the creativity and the variety that people are so
(30:00):
good at, and having robots that can pick up all
of the tasks that we'd rather not do, and having
robots that can be in environments that are designed around us.
Is a really important step to being able to achieve
that in a really great way.
Speaker 1 (30:14):
Is there anything else you want to talk about?
Speaker 2 (30:17):
MM. One thing I want to make sure that we
get across is kind of the clear argument that the
fastest path to general purpose machine that does use full
work in human spaces is to do one thing first
and then the second thing, and to do it for
customers and to get it deployed and to figure out
(30:38):
the reliability and the safety and so on. That's the path.
There is no good way to just like jump to
the answer.
Speaker 1 (30:44):
So basically you're saying, you can't just build a robot
that does everything. You have to build a robot that
does one thing and then figure out how to make
it to a second thing.
Speaker 2 (30:52):
Yeah, but you want to stay on that vision. You
want to have your guess of what the everything robot
looks like, but you know you're going to be wrong,
and so then you start on what's the match to
the first thing that it should be able to do,
and then you keep on revising and iterating on that path.
So I'll give an example. This This is through our
entire est then of building the function first and the
physics first and the first principles approach to figuring out
(31:13):
things like the legs. Right, yeah, I want to point out,
like hands, how do you produce a dextrous manipulator? By dexterous,
I mean something that can pick up pens and pick
up objects and do useful things with the hand, open doorknobs,
all the stuff. Right, I don't care how it looks,
I care what it does. So what we're doing for
our first use case is picking up totes that can
(31:35):
have a twenty kilogram bowling ball rolling around in it.
That's a very hard thing to pick up. And so
now our grippers are these big graspers that can grasp
the side of this tote and pick it up. A
lot of groups are sort of have these five figured
hands on their robots which look like a human's hand,
but certainly couldn't pick up those twenty kilogram totes.
Speaker 1 (31:55):
What do you have just done? Pincer sort of basically yeah.
Speaker 2 (31:58):
You know, effectively it's a big sort of four fingered
pincher thing that can have leverage on it and grasp
the sides of the tote.
Speaker 1 (32:05):
I mean does that mean that it can't do sort
of fine motor things. The trade off that you're making this.
Speaker 2 (32:10):
I would say the current pincher design, Yeah, it's not
designed to pick up small objects. But you know, we
see a vision where that's actually a tool, not a hand.
And just like people use tools, robots are gonna use tools.
But there's such an opportunity to have the tool be
attached at the wrist or the fingers or the elbow
or the forearm or wherever. And we don't know exactly
(32:31):
how that should be, and nobody does. But one thing
I can say with confidence is that these five fingered
hands cannot do dexterous manipulation. It's not just a controls problem.
They're making the same mistake of creating something that has
the same morphology. It looks like a person. It looks
like a person, but that doesn't mean it can apply
(32:52):
the right forces, or have the right kinematics, or have
the right dynamics or the right compliance or anything that.
We don't know how to do that. It's one of
the grand challenges in robotics. So the fastest path to
get there is to start manipulating things and do stuff
that has metrics right measures, a.
Speaker 1 (33:09):
Job that someone is actually willing to pay for that's right.
So right now it's moving tots around as basically, what's
the second job?
Speaker 2 (33:16):
Boxes? Cardboard boxes, huh? And then the third job is
starting we want to start doing each picking, you know,
picking up things and putting them in the boxes and
in the So those are much.
Speaker 1 (33:25):
Smaller things, different than picking up a big punt.
Speaker 2 (33:30):
That's right. And so the question is, and nobody knows
the answer to this, is that two different tools or
is that one general purpose manipulator that can do both
of those things.
Speaker 1 (33:38):
They need to put a different hand on the robot
to pick up a little thing versus pick up a
big sure, or is it optimal?
Speaker 2 (33:44):
And you know this will be the this gets into
customer requirements question rather than fundamental science question. Yeah, that's
your that's for your product person to figure out. Yeah,
but that's how this is going to evolve. That's how
we're going to get to dexterous manipulation in the world
in a way that really is meaningful.
Speaker 1 (33:59):
Investors, one job at a ton, Yeah.
Speaker 2 (34:01):
Figure out if you've got the first like you know,
the first four or five jobs, and now you've got
your requirements, yours of requirements and your measurements and your
metrics and now engineers are going to be able to
iterate on that really make something work. But just trying
to copy a five figure in hand and say now
now it's an AI problem completely false, that's just not
going to happen.
Speaker 1 (34:24):
We'll be back in a minute with the lightning round.
Let's finish with the lightning round. I've read that you
like to jog at night for work, but like, are
(34:45):
there specific things that have happened to you or that
you've done to try and sort of you know, put
yourself in hard walking or running settings and where you've
actually had a thing happened and thought, oh, we need
to make sure the robot doesn't fall over when X happens.
Speaker 2 (35:00):
Honestly, it's more like when you're taking a hike, you know,
and you kind of get into the mode of you're
just your long day and a long hike and you know,
watching and being mindful, I guess, and thinking about what
are my feet doing and how is the contact progressing
with the ground and how does that feel and why
is that happening? And sort of daydreaming while you're thinking
(35:21):
that through is a really good way then to start
to recognize connections to research papers and connections to things
that people have found scientifically, and then start to pull
together hypotheses about how you would implement something like this
on a robot, or what is necessary because we're human,
and what is necessary because it's fundamental to look emotion
(35:43):
like why do we have feet? That's very different from
bird feet atreus didn't have feet, you know, And then
what exactly do we get anyway? Thinking through all that
kind of stuff.
Speaker 1 (35:56):
C three po or R two D two.
Speaker 2 (36:01):
Why that is a tough one. I'm going to say
R two D two.
Speaker 1 (36:05):
Not the bipedal one, not because that walks on two legs.
Speaker 2 (36:10):
C three, Well, no, I'm going to change my mind.
Speaker 1 (36:13):
I buyased it. Wait, why were you going to say
or two I was going to say to.
Speaker 2 (36:16):
You because C three pos a protocol droid And if
it's just about language, why do you have legs? Was
On the other hand, you.
Speaker 1 (36:24):
Just you don't need a robot at all, right, you
just need a little phone or whatever.
Speaker 2 (36:27):
On the other hand, right, it's a human centric robot.
It's a robot that's meant to be a translator and
be existing with people in the room and so on,
and so actually making something that's more of a humanoid
makes more sense if it's basically a social robot.
Speaker 1 (36:41):
Yeah, okay, Uh, what's something you wish that a robot
could do for you, you know, outside of work.
Speaker 2 (36:50):
Honestly, what I want is for robots to make everything cheaper.
Speaker 1 (36:55):
Uh huh. I love it when technology makes things cheaper. Well,
and that's what's been happening for the past couple hundred years,
and all of automation has effectively made everybody richer effectively
and improved everybody's quality of life because everything is cheaper.
And I want to be able to book my vacation
to the moon. I want to be able to not
really worry about you know, as it is, I don't
(37:17):
really worry about how much my phone costs or whatever.
If it breaks, it's not an expensive phone, it's fine,
it works, and all my software to transfers over. I
want everything to be like that. I want my lifestyle
to be supported by things that don't really cost a lot. Yes, well,
and I mean in the very long run, the sort
of technology driven productivity gains lift people out of poverty. Right,
(37:40):
everybody cares about our phones, but there are a lot
of people who have three dollars a day right now,
and it would be great if they could get to
three hundred dollars a day.
Speaker 2 (37:48):
Or beyond that. It's just that, you know, all the
things that we need become very easy to acquire. Food
is no longer, labor on farms is done in an
automated way, Labor in terms of logistics and transporting things
has done in an automated way, and so all of
this stuff becomes so affordable that it's easy to uplift
the quality of life of everybody on earth. That is
(38:08):
what a lot from Robotics.
Speaker 1 (38:15):
Jonathan Hurst is the co founder and chief robot officer
at Agility Robotics. Today's show was produced by Gabriel Hunter Chang.
It was edited by Lydia Jane Kott and engineered by
Sarah Bruguer. You can email us at problem at Pushkin
dot FM. I'm Jacob Goldstein, and we'll be back next
week with another episode of What's Your Problem
Pushkin. So how long have you been trying to make
a robot walk?
Speaker 2 (00:25):
It's been my entire career, starting from why I went
to college in the first place.
Speaker 1 (00:29):
Why why that particular problem? Why is that your life's work?
Speaker 2 (00:35):
You know, there's few things more interesting and more dynamically
complex and more elegant than the way animals move in
the world. And to be able to get machines that
can move that way, they can interact physically with the
world the way humans and animals do. What a fun
and interesting thing to work on for a career.
Speaker 1 (00:57):
I'm Jacob Goldstein and this is What's Your Problem? The
show where I talk to people who are trying to
make technological progress. My guest today is Jonathan Hurst. He's
a professor at Oregon State University and founder and chief
robot officer at Agility Robotics. Agility Robotics has built a
robot called Digit. Digit looks kind of like a person.
(01:19):
It walks around on two legs. It's got this flat,
rectangular head, and it has two arms that it can
use to pick stuff up. Jonathan's problem is this, how
can you make a walking robot that can do useful
work and that companies will actually pay for. Jonathan says
that robot Digit is already being tested out in warehouses
(01:40):
in the real world.
Speaker 2 (01:43):
We are deploying robots with cost We have two announced.
We've announced a couple with Amazon and with GXO. You know,
you place an order and Digit handles that order as
part of the workflow that has happened is happening right now, and.
Speaker 1 (01:56):
So specifically, what are your robots doing well? First at Amazon, Yeah,
the first use case or the first class of use
cases that works for us is picking up these plastic
tote plastic bins and putting them somewhere else. And then
in warehouses, Yeah, but anywhere warehouses, in logistics, in manufacturing,
(02:18):
you know, the whole environments like that that are a
bit structured. They're kind of like there's these islands of
automation they call them, you know, where one machine is
putting things in a bin, another machine sorts bins and
sends them different parts of the warehouse. Right now, sometimes
a person will stand there, the robot will tell them
which bin to pick up, and then all they are
basically is a manipulator for the robots system. They pick
(02:41):
up the bin and put it on the conveyor belt
and wait for the robot to tell them the next
thing to do. And it's really hard to hire people
for that. There are a lot of open jobs in that,
and so it's kind of a perfect place for Digit
to walk in in this relatively structured first use case. Now,
digits of course going to evolve towards picking up boxes, depalletizing,
(03:01):
loading and unloading you know, tractor trailers and eventually getting
out to things like retail and stocking shelves and you
know sticks in hospitals carrying things around and eventually become
a consumer product. So that's where you are today and
where you want to get. I want to talk now
about how you got here. Sure, and there's this really
(03:24):
basic set of things you had to figure out just
around how locomotion works, right, how people walk, also interestingly
and sort of surprisingly, how how birds like ostriches walk.
So I know there is this series of robots that
you built on the way on the way to Digit, right,
(03:45):
they were two before and then Digit and it seems
like going through those and what you figured out the
sort of key insight on each one is a really
nice way to get a kind of deeper insight into
how it works and what you had to understand to
make a robot that can walk great, just being a
really hard problem, right, Like there's like lots of robots
(04:05):
that we don't even sort of think of as robots arms,
and you know, self driving cars are arguably a kind
of robot and whatever, but like getting a robot to
walk is clearly a very hard problem.
Speaker 2 (04:16):
Yeah, Okay, So where I started in trying to understand
how to make a machine work is on the biomechanics,
try to understand how animals work. And a lot of
biomechanics is about specific muscles and muscle groups and joints,
and we're instead looking at holistically from a big picture,
what is the center of mass of an animal doing?
What are the forces happening on the ground. I did
(04:38):
this collaboration with Monica Daily at the Royal Veterinary College
and we looked at guinea fowl and ostriches and turkeys
and you know, a whole bunch of different sizes.
Speaker 1 (04:47):
Why birds. It's really interesting to me that you did that.
Speaker 2 (04:51):
Why because humans are weird? How many other bipeds are there?
You know, all the evolved bipeds, all of the extant
existing bipeds in the world.
Speaker 1 (05:00):
Everything that walks on two legs. Yeah, huh.
Speaker 2 (05:03):
Those are all theropods. They're all kind of more like
a bird than they are like or a monkey. They
have evolved far longer than we have. You know, we've
been out of trees for a very short period of time.
Compared to all of the other existing bipeds in the world.
They can all run much faster than we can for
you know, very efficiently using less energy insul So.
Speaker 1 (05:27):
Like an ostrich is a better model for just abstract
walking on two legs. Oh sure, I mean, yeah, I
love that, But I want to be clear.
Speaker 2 (05:36):
What we're not trying to do is study how it
does an Ostrich run versus how a human runs. What
we're trying to do is study what is a fundamental
truth between all animals in how they run, so we
can try and weed out things that have to do
with the size of the animal, or things that have
to do with where exactly the ankle joint is or
how long this joint is or that one. We want
(05:56):
to know what is the same amongst all animals to
walk and run.
Speaker 1 (06:01):
It's like the Platonic ideal of bipedalism. The sort of
the fundrationing theory of it. Yeah, that's right.
Speaker 2 (06:07):
Yeah, And so biomechanists have been looking at this since
the seventies and thinking about this in terms of spring
mass models of locomotion. And it was only in like
the two thousands sometimes that I think it was Hart
McGuire and Andre Seifarth put together a paper that showed, hey,
this spring mass model reproduces all of the behavior we
see for walking and running and transitions between these gates.
Speaker 1 (06:27):
When you say, when you say spring mass model, that
sounds big and exciting, But just to be clear, what
do you mean when you say spring mass model.
Speaker 2 (06:35):
I mean a mathematical representation of a pogo stick. Go on, Okay,
So a pogo stick is basically the simplest thing that
can run, and a kangaroo looks a lot like a
pogo sticky. Now, if you just stick a pogo stick
on each leg, now you're bipedal running, you know.
Speaker 1 (06:52):
Okay.
Speaker 2 (06:53):
And then if you add a whole bunch of complexity
to it, you have heel toe and you know, knee
joints and all this other stuff. But if you really
boil it down and try and make it as simple
as possible, you get to some pretty basic math models
that do represent how the progression of the center of
mass of the animal moves, and how the ground reaction
forces progress and so on.
Speaker 1 (07:11):
Right, Okay, so if I picture just like a lump
of mass on top of two pogo sticks, you got it.
I'm kind of in the right.
Speaker 2 (07:18):
Okay, you're absolutely okay. So that's a roughly a math
model that at least gives you a very good concept
of how do all animals run, horses, ghost, crabs, humans, ostriches, whatever.
Speaker 1 (07:28):
Right, So this paper comes out, this paper that says,
think of a lump of mass on top of two
pogo sticks. You you know about it because you're in
this world? What do you What effect does it have
on you? What do you do as a result of this.
Speaker 2 (07:42):
What I was looking at? And here's the argument at
the time. Do these spring mass models simply seem to
describe the things we're observing, you know? Or is it
describing core physics of how it works? Like, in other words,
if you build a spring mass model and a policy
that works, is it going to make a robot stabilize?
Or is it simply like a picture that kind of
(08:04):
looks like what the animals are actually doing.
Speaker 1 (08:05):
So like if we actually do the lump of mass
on top of two pot sticks as a robot, will
it work?
Speaker 2 (08:11):
Yeah? And so look the question is, you know, how
do you control these things? And then how does that
translate into walking and running which had never really been
done before that way, you haven't done this continuous transition
between walking and running and changing the speed and everything else,
and it's unknown how to stabilize that over all kinds
of terrain. So that's why we built Atreus. That's why
we built the robot Atreus.
Speaker 1 (08:30):
And so what is Atreus?
Speaker 2 (08:32):
So Atreus is it is a bipedal robot.
Speaker 1 (08:35):
Okay, what does Atrius look like?
Speaker 2 (08:38):
So Atrius was on the Late Night with Stephen Colbert
and he described it as a microwave on stilts.
Speaker 1 (08:44):
Okay, that's good. So it doesn't it doesn't look humanoid
at all. It looks maybe like a dancing alien or
like a moon lander or something, but not humanoid.
Speaker 2 (08:58):
Yeah, absolutely not. So it doesn't look like an animal
in any way. But it's designed entirely to be the
math model of what we see an animal running. And
the name Atrius is an acronym for assume the robot
is a sphere, right. The whole idea is the robot
is this simple math model.
Speaker 1 (09:14):
Where it's just some mass in the middle and then
some very light springy legs.
Speaker 2 (09:18):
That's it. And so what we ultimately showed with Atreus
it's the first robot ever to reproduce human walking gate dynamics.
The robot walks across the force plate, a graduate student
walks across the forest plate. Looking at the data, you
can't actually tell a difference.
Speaker 1 (09:32):
Huh. So, like from the plate's point of view, it
feels the same whether you're robot or a human is
walking across.
Speaker 2 (09:38):
That's correct.
Speaker 1 (09:39):
So it's not true for earlier robots that looked like
they were walking like humans.
Speaker 2 (09:43):
They would look the same. But if you looked at
the dynamics of it, if you looked at the ground
reaction forces, they differed quite a lot.
Speaker 1 (09:49):
Why is that important? Why is that important?
Speaker 2 (09:52):
It's just one symptom, right to show that, hey, we've
actually captured the physics here. But the other symptom that's
important is we are able to walk and run outdoors
over all kinds of terrain without any sort of perception.
The robot can handle amazing obstacles and it would just
soak them up, you know, going over potholes, going from
grass to pavement, going over big pieces of plywood we
(10:13):
would throw in its way.
Speaker 1 (10:15):
And you're saying it didn't do this because it had
like a clever brain. You're saying it was just the
physics of the brainless machine was able to deal.
Speaker 2 (10:23):
No cameras on it, It had no awareness of the environment.
It was a very simple spring mass model, very very
simple control that did nothing but try to balance that.
And it was able to just absorb all these kinds
of disturbances and just keep on going.
Speaker 1 (10:36):
That sounds like a big deal. I agree.
Speaker 2 (10:39):
I'm very excited about this, right. That is the point
that we decided we were going to found Agility Robotics.
We said, you know, this was a mission for like
years and years and years. That is why I became
a professor, is to say, my goal here in academia
is to show that this spring mass physics is real
and really make sure we understand that well. And if
(11:02):
we can show that and prove that and with this
ATREUS project, we then can take the next steps. Right,
So that's what we did. It was a scientific kind
of breakthrough. But the machine could only walk and run.
It couldn't stand, it couldn't turn. You know, it breaks often.
If it ever fell, it would just be completely destroyed.
(11:23):
So it's not a productive, useful machine. It's a science demonstrator.
Speaker 1 (11:27):
Okay, So you have a TRIOS that is academically it
kind of intellectually a breakthrough, right, but nobody's going to
buy it to do anything. It's not useful in a
practical sense. What do you do next?
Speaker 2 (11:44):
Okay? So at that point we say, hey, you know,
we understand a really significant portion of what does it
take to make a robot that can go where people go?
And what an opportunity? This is? What are all the
things that robots are going to be able to do
when they can be coexisting with humans? Right? And the
barrier to moving forward on that now is more about
(12:06):
execution on engineering and execution on building a use case
in a business around it.
Speaker 1 (12:12):
Because you feel like you've solved the sort of core
threshold technical problem of a robot that can walk in
unfamiliar environments and not fall down.
Speaker 2 (12:21):
Correct, we have the foundation layer of leged locomotion. Now,
sure we'd like to add on perception to it so
that it can handle stairs effectively and things like that
and intentionally handle obstacles. We need to do the engineering
of the electrical system and the hardware system that still
captures the same physics but can take a beating, can stand,
can steer, can start to do things right. So we
know we're going to need to build a robot with legs,
(12:43):
with manipulators, with sensors. Okay, because so we're kind of
going down a path now where we want to take
the exact same first principles approach to how do we
build a machine that can manipulate things in a human
world and get around in it and interact with people.
Build a human centric machine. So the first step to
(13:03):
doing that was designing a robot that we could sell
to other researchers to continue the work on the leg
leggs as we then worked towards you know, arms and manipulators,
and that was Cassie, our first robot that Agilia Robotics sold.
Cassie added the ability to stand in place because it
had ankles, and it had the ability to steer because
you could turn the legs. But more importantly, it was
(13:26):
much more compact and extremely robust, so this robot can
fall hard on concrete and you just pick it back
up and it can get going again.
Speaker 1 (13:34):
And so, Cassie, I'm looking at a picture of it now.
It basically looks like a pair of Ostrich legs. Yeah,
it looks like an Ostrich. Does an Ostrich have a wasist?
I don't know an ostrich from the waist down, No.
Speaker 2 (13:47):
It doesn't know. Yeah, the pelvis is stationary and a
bird fixed rather than a human pelvis, which is mobile.
Speaker 1 (13:54):
So I know it's not technically correct to say that
an Ostrich legs bend backwards, but it looks that way, right.
Speaker 2 (14:00):
Yeah, Their thigh is very short, their knee is up
next to their body, and what you perceive as their
knee is actually their ankle.
Speaker 1 (14:07):
In designing this robot, how do you get to legs
that look like Ostrich legs? Like, it's like convergent evolution,
you know, convergent evolution maybe hopefully. Isn't it the case
that like cephalopods that like whatever squids have eyes like
our eyes, but they evolve totally independently. Is this like that?
Speaker 2 (14:26):
There's a lot of examples of convergent evolution. We can
only guess, right because we don't necessarily know, and a
scientists only hypothesize the evolutionary pressures that cause animals to
be the shape that they are.
Speaker 1 (14:36):
But the pressure that led you you didn't say let's
make legs that look like Ostrich eggs. You just did
a bunch of math and you wound up with legs
that look like Ostrich legs.
Speaker 2 (14:44):
That is correct, and there are a bunch of features
on our robot that have gone down a similar path.
And I actually love that because when we end up
you know, blank sheet, pursue all of the different configuration
options and say okay, here's what we think is the
optimal choice, and we say, wow, that looks just like
a person, or that just looks like a bird or something.
It's actually really good. It means we're probably on the
(15:04):
right path.
Speaker 1 (15:05):
Yeah, it's it's exciting in a way, right, like like
you do a bunch of bath and then suddenly you
look up and you see an Ostrich.
Speaker 2 (15:13):
But it won't always be that way, right because we're
not using muscle and bone. We're using aluminum and you know, electricity,
it's a whole different thing.
Speaker 1 (15:19):
In a way, it's surprising that it is right, like
you would expect that it wouldn't look at all.
Speaker 2 (15:24):
Familiar, but there are clear differences.
Speaker 1 (15:26):
I suppose I'm projecting this. It's like, whatever is the
Ostrich version of anthropomorphizing, right, I'm Ostrich pomorphizing?
Speaker 2 (15:34):
Yeah, you got it. And like you said, it's like
a cartoon version of an Ostrich leg maybe.
Speaker 1 (15:38):
Yeah. Okay, So you've got this robot that is looks
to my little brain like a pair of Ostrich legs.
It's just a coincidence because it just turns out to
be the best way to build a couple legs. And
do you sell it to other academics? What do you
do it?
Speaker 2 (15:52):
Yeah? We sold it to some of the top universities
in the country and the kind in the world.
Speaker 1 (15:59):
So Cassie is a robot that academics can experiment with
and learn from. But it doesn't have arms. It isn't
built to do useful tasks. In a minute, Nathan and
his colleagues build a robot that does have farms, that
can do useful tasks, and that companies like Amazon are
testing out in the real world right now. The latest
(16:30):
robot from Jonathan's company is called Digit. It has two legs,
two arms, It walks around, it picks stuff up, and
it looks kind of like a person. But Jonathan says
he and his colleagues didn't set out to build a
robot that looks like a person.
Speaker 2 (16:44):
So yeah, so like, I'll take you down one of
these thought processes that ended up looking like a person. Okay,
We said, okay, we need to do some sort of
inertial control of this thing because the robot can't turn
very well. It's got little feet, and so when you
try to turn aggressively, its skids right, okay, and if
you look at any other bipads, this is one of
(17:06):
the reasons wings evolved. It's because they're running in the stick,
going out to catch the air to help them in
maneuvering and turn it. You can go in a straight line,
but figuring out how to maneuver quickly as hard when
you've only got a little foot on the ground. You know,
udropeds can really plant all four feet and twist and
apply big torchs on the ground, and a biped not
so much. You've got one foot at a time. How
do you change your orientation?
Speaker 1 (17:27):
Right?
Speaker 2 (17:28):
So we looked at like putting a gyro on board
reaction wheels or tails or things like that, because you know,
we ruled out reaction wheels because that's just a big
thing of brass that has to be mass you don't
want in a robot like this. We thought about tails,
and you look at any animals with tails, bipeds in particular.
Typically that's to control the pitch. In other words, you'd
(17:48):
leap off the ground, and then you want to reorient
your body and your feet so that your feet land
forward rather than you know, just tumbling in the air. Okay,
but we don't want to control pitch. We want to
control y'all. We want to steer the robot. So what
we kind of came up with is that the best
way to do that would be a pair of tails
that are symmetrical on the front of the back or
(18:10):
the side of the side of the robot, so that
when you swing those tails, you're controlling exactly down the
center line of your yaw, and so that just happens
to be where our arms and shoulders are. You know,
this bilaterally symmetrical pair of tails that can inertially actuate
you around the center and allow you to steer.
Speaker 1 (18:27):
So that's the sort of intellectually elegant version of how
you get to a robot that looks like it has arms.
You're saying, in fact, in terms of the way you
thought of it, it's a pair of tails that happened
to sit where our arms sit. I mean, presumably there's
a simpler version, which is, you want to build a
robot that can do stuff in a world that is
built for humans, and having arms would be useful in
(18:50):
that context as well. Or no, was it true?
Speaker 2 (18:53):
Member? Yeah, what we're not trying to do is make
a humanoid robot that looks like a person. What we're
trying to do is on first principles, understand exactly why
we do each thing.
Speaker 1 (19:01):
But are you really just adding the arm so that
it can turn when it's walking, Like I feel like
that's what you're saying, and I'm skeptical.
Speaker 2 (19:08):
Yeah. It's also that there were three other reasons why
the arms should be there that were all aligned. They
were not compromises where you know, putting the arms here
is better for one thing or another.
Speaker 1 (19:18):
Huh.
Speaker 2 (19:19):
It's that. Hey, if you go on just the just
the path of I want to improve my locomotion capability,
you land at the solution of where the arms are.
If you go down the path of this robot's going
to fall, and we know that it can't just fall
and land on its torso it'll break things. How do
we put manipulator's arms something on it so that it's
going to be able to catch itself when.
Speaker 1 (19:38):
It falls oriented? Okay.
Speaker 2 (19:41):
And then the third one, of course, is picking up
things right manipulation in the world and being by manual
in your manipulation so that you can basically a giant
pincher grass. That's how you pick up boxes and tots
and all these things you want to move. That's also
the best place for them. So basically, you just set
out to build a machine that could go where humans
go and pick up things that are the size that
(20:03):
humans pick up, and from first principles, with your eyes closed,
you wound up with a thing that looks like a guy.
Speaker 1 (20:10):
Absolutely so okay, so this is how you get to
digit the robot that you're now building and selling to people.
Speaker 2 (20:17):
That's right. So now we're taking this transition right now
as a company from that very intellectual and first principles
approach that I shared with you to now working with
the customer understanding what their use case is, writing down
the sets of requirements, like you know, the temperature ranges,
the you know, weights of all the things that you're
going to be able to pick up, the safety requirements,
(20:39):
you know. And it's a massive list, hundreds of things
in a list to write down the requirements documents so
that we can engineer a system that is a product. Yeah,
very different from designing a robot that can do cool things.
We're engineering a product, and that's the pivot that our
company is in right the second.
Speaker 1 (20:54):
And like I imagine for you personally, that must be
a significant shift, right if you spent whatever twenty some
years in the kind of abstract academic world of like
let's build the thing that works and know think deep
thoughts to like let's mass produce a product that people
will pay us for. That's quite different.
Speaker 2 (21:15):
Oh, it's fundamentally different. It's a whole different way of thinking.
In fact, I changed my title to chief robot Officer, right,
and we hired Melanie Wise as our new chief technology officer.
She comes out of FET Robotics. She was the founder
there and recently sold that company and they were deploying
thousands of robots in logistics squarehouses. And she is an
(21:36):
absolute expert on understanding customers and product and creating a product.
And what we've done is we've shifted our organization. So
you know, Melanie is in charge span of that whole
product side of the organization and the engineering to make
a product. I'm now leading the innovation team.
Speaker 1 (21:52):
So you get to keep doing kind of the stuff
you've been doing.
Speaker 2 (21:56):
The things I'm good at.
Speaker 1 (21:57):
Yeah, so what is the frontier on the innovation side.
What are you trying to figure out next?
Speaker 2 (22:02):
It is fundamentally hardware that enables the kinds of physics
that we want to achieve, powered by some of these
new AI tools. You know, we're getting to a point
now where some of these tools will allow us to
create behaviors and create things that as an engineer we
don't know how to model. Huh, And that's super interesting.
(22:24):
So instead you're describing the symptoms of it, and then
the system, the learning system, figures out how to make
that happen.
Speaker 1 (22:31):
Amazingly different than what you've been describing. You've been describing
of like, let's think of you know, first principles, just
the physics of the universe, and from that build a machine.
And now you're talking about you know, an era when
possibly you'll be able to ignore all of that and
say to the AI, you figure it out, here's what
I want to do.
Speaker 2 (22:50):
Well, let me put some caveats on that.
Speaker 1 (22:52):
Yeah, that sounds ridiculous when I say it that way.
Speaker 2 (22:55):
Well, remember that the AI has to operate on a
piece of hardware. Yeah right, And so that piece of
hardware we still have to engineer and design to be
able to achieve the physics that we want to achieve.
Speaker 1 (23:05):
Though you could say to an AI, here's what we
want to do. What should the hardware look like? Maybe
in one hundred years you think one hundred, one hundred,
who knows one hundred.
Speaker 2 (23:14):
It's fine, fine, three years, you know, future, Not today.
Speaker 1 (23:18):
Not certainly not. So what specifically are you doing today?
Are you taking this robot that you have digit and
seeing if you sort of put an AI layer in it,
on it near it? What can you do?
Speaker 2 (23:27):
Is that what's happening all of the above, So, you know,
on the hardware side, So there's a lot on the
hardware you do just to make it even possible for
the AI to learn. We're building then a whole architecture
of a digital twin so that you can learn things
in simulation first, and then you know, transfer from sim
to reel.
Speaker 1 (23:43):
A digital twin is basically making a version of the
robot that exists as software that exists virtually.
Speaker 2 (23:49):
Version of the environment as well that the robot operates in,
so that everything can be done, you know, decades of
experience on the robot can be done in hours of
time through the you know, et cetera.
Speaker 1 (24:00):
So the digital twin is allowing you to try and
generate data to train the AI. Is that that's what's the.
Speaker 2 (24:06):
Source of the data, right, A lot of language models
and things like that are based on data from the internet. Well,
I don't think that that's feasible for robot control because
the physics of the hardware is so unique. So even
if you're trying to teleoperate this thing, you've got this
weird translation between what a person would do to then
the robot trying to mimic that, which is probably not
the dime of the right.
Speaker 1 (24:25):
Dynamic robot doesn't actually walk like a person, even if
it looks like it's.
Speaker 2 (24:29):
Right, that's right, everything's sort of different internally about it.
How it would control itself.
Speaker 1 (24:34):
What what what are you worried about? Like you what
could go wrong and how are you trying to get
it not to go wrong?
Speaker 2 (24:43):
So before I talk about what I'm worried about, let
me tell you what I'm excited about. Fair We had
ten of these Cassie robots out in the world, and
so researchers all over the place for for years and
years are working on various kinds of controllers. When we
were able to successfully get a learned policy working on Cassie,
were to run a five k across campus, we were
able to the world record in the one hundred meter dash.
(25:04):
We were able to.
Speaker 1 (25:05):
Do for a robot. Yes, and when you say learned,
you mean developed with machine learning as opposed to in
the old way. Is that what you mean?
Speaker 2 (25:12):
Correct? It's an entirely machine learned policy that was learned
in simulation and then put on the thing.
Speaker 1 (25:18):
What's a definition of a policy? As you're using the word, I.
Speaker 2 (25:22):
Mean, a policy is just a bunch of math that
takes as input all of the sensors and then spits
out numbers that describe the torques you should apply to
the motors.
Speaker 1 (25:34):
Okay, so it's sort of if this happens in the world, robot,
you should do this set of things.
Speaker 2 (25:40):
Yeah, you're based on you know, but you know it's
like the input from thirty encoders or one hundred different sensors,
all of that complex input.
Speaker 1 (25:47):
But if this is complicated and the then that are complicated, Yeah.
Speaker 2 (25:51):
No, you're rreat you're right about that. Yeah, it's an equation.
Speaker 1 (25:53):
Good. So the machine learning basically made the robot work
way better.
Speaker 2 (25:58):
It made the robot work much better, and even more importantly,
once the pipeline was in place, we can learn new
skills and learn new policies much faster with much less
engineer time. How we can get there faster and have
higher performance using learning approaches to control.
Speaker 1 (26:16):
It's like a productivity like supercharger. It just makes everything
go much faster and more efficiently.
Speaker 2 (26:21):
Absolutely, it's a new tool. It's amazing, that's right. Okay,
So what am I worried about? Right? What am I
worried about? I'm worried that this is one of those
kind of black Swan events. Right, this is one of
those things that changes everything, and everybody doesn't exactly know
what all the implications are yet and what are the
(26:41):
you know, the right paths forward, and so everybody's trying everything.
Speaker 1 (26:45):
And this being basically a useful bipedal robot like it,
it could be hugely important in ways that we don't understand,
and there could be unintended consequences.
Speaker 2 (26:55):
What I actually mean is that this new realization that
we can use learning policies to control dynamic robots and machines, yeah,
means that the entire way it all robotics controls people
have been doing robot controls before is not as relevant.
(27:17):
And this new tool that nobody really understands that well
yet is clearly the future of how it's going to work.
Speaker 1 (27:25):
So what are I mean? I understand that if it's
really a black swan, you don't know what's going to happen,
because if you knew, it wouldn't be a black swan,
But like, what are you thinking of? What could it mean?
You know, plainly, bipedal robots are a very powerful tool,
and you could imagine malevolent uses of them, right, I guess, So,
(27:46):
I guess we've already got drones. Right, It doesn't matter
whether the robot that kills you looks like a dude, right,
those things already exist.
Speaker 2 (27:52):
In fact, a humanoid robot is probably the least effective
way to do that I think, honestly, my biggest worry
about AI in general has a lot more to do
with its ability to influence people, its ability to model
people's feelings and that kind of thing.
Speaker 1 (28:06):
So that's more like using large language model for kind
of personalized misinformation or that.
Speaker 2 (28:11):
Stat it that's the biggest threat building robots that are
going to like take people. I don't know, I just
don't see it fair.
Speaker 1 (28:19):
I mean, I'm kind of tired of talking about technological
unemployment because the robot looks like a person. It makes
me feel like we should touch on it. Do you
want to just speak for a moment to the prospect
of technological unemployment?
Speaker 2 (28:30):
Sure, I'll say our entire business model is centered around
the number of unfilled roles that exist in the logistics environment.
It's not centered around how it's going to be less
expensive than human Labor's centered around how they actually in
geographic locations do not have enough people to provide the
service that they're providing, you know, the way they're doing
(28:51):
it now, there's no way forward for improved logistics and
getting your things in one day and you know, all
the stuff that people really really want. There's no path
forward to do that with more human labor doing it,
it must be automated in a significant way. So you know,
that's our whole business model is based on unfilled roles.
Speaker 1 (29:11):
What is your like happy version of the future in
whatever number of years seems like the right number of.
Speaker 2 (29:17):
Man These robots are actually safe and smart enough in
order to do a lot of useful things in the world.
And the relationship with people is kind of like a
service animal. And you know, these robots are everywhere and
delivering all the packages to your door, and you know,
being a telepresence device that you can easily log into
and keep in touch with people, and you know, in
(29:38):
a lot of warehousing and stocking and you know, doing
the dull, dirty, dangerous to classic three d's of robotics.
It's we've always wanted that. It's all about improving the
quality of life. It's all about letting enabling humans to
be more human, letting people do the things that are
that they want to do that involve the social interaction
and the creativity and the variety that people are so
(30:00):
good at, and having robots that can pick up all
of the tasks that we'd rather not do, and having
robots that can be in environments that are designed around us.
Is a really important step to being able to achieve
that in a really great way.
Speaker 1 (30:14):
Is there anything else you want to talk about?
Speaker 2 (30:17):
MM. One thing I want to make sure that we
get across is kind of the clear argument that the
fastest path to general purpose machine that does use full
work in human spaces is to do one thing first
and then the second thing, and to do it for
customers and to get it deployed and to figure out
(30:38):
the reliability and the safety and so on. That's the path.
There is no good way to just like jump to
the answer.
Speaker 1 (30:44):
So basically you're saying, you can't just build a robot
that does everything. You have to build a robot that
does one thing and then figure out how to make
it to a second thing.
Speaker 2 (30:52):
Yeah, but you want to stay on that vision. You
want to have your guess of what the everything robot
looks like, but you know you're going to be wrong,
and so then you start on what's the match to
the first thing that it should be able to do,
and then you keep on revising and iterating on that path.
So I'll give an example. This This is through our
entire est then of building the function first and the
physics first and the first principles approach to figuring out
(31:13):
things like the legs. Right, yeah, I want to point out,
like hands, how do you produce a dextrous manipulator? By dexterous,
I mean something that can pick up pens and pick
up objects and do useful things with the hand, open doorknobs,
all the stuff. Right, I don't care how it looks,
I care what it does. So what we're doing for
our first use case is picking up totes that can
(31:35):
have a twenty kilogram bowling ball rolling around in it.
That's a very hard thing to pick up. And so
now our grippers are these big graspers that can grasp
the side of this tote and pick it up. A
lot of groups are sort of have these five figured
hands on their robots which look like a human's hand,
but certainly couldn't pick up those twenty kilogram totes.
Speaker 1 (31:55):
What do you have just done? Pincer sort of basically yeah.
Speaker 2 (31:58):
You know, effectively it's a big sort of four fingered
pincher thing that can have leverage on it and grasp
the sides of the tote.
Speaker 1 (32:05):
I mean does that mean that it can't do sort
of fine motor things. The trade off that you're making this.
Speaker 2 (32:10):
I would say the current pincher design, Yeah, it's not
designed to pick up small objects. But you know, we
see a vision where that's actually a tool, not a hand.
And just like people use tools, robots are gonna use tools.
But there's such an opportunity to have the tool be
attached at the wrist or the fingers or the elbow
or the forearm or wherever. And we don't know exactly
(32:31):
how that should be, and nobody does. But one thing
I can say with confidence is that these five fingered
hands cannot do dexterous manipulation. It's not just a controls problem.
They're making the same mistake of creating something that has
the same morphology. It looks like a person. It looks
like a person, but that doesn't mean it can apply
(32:52):
the right forces, or have the right kinematics, or have
the right dynamics or the right compliance or anything that.
We don't know how to do that. It's one of
the grand challenges in robotics. So the fastest path to
get there is to start manipulating things and do stuff
that has metrics right measures, a.
Speaker 1 (33:09):
Job that someone is actually willing to pay for that's right.
So right now it's moving tots around as basically, what's
the second job?
Speaker 2 (33:16):
Boxes? Cardboard boxes, huh? And then the third job is
starting we want to start doing each picking, you know,
picking up things and putting them in the boxes and
in the So those are much.
Speaker 1 (33:25):
Smaller things, different than picking up a big punt.
Speaker 2 (33:30):
That's right. And so the question is, and nobody knows
the answer to this, is that two different tools or
is that one general purpose manipulator that can do both
of those things.
Speaker 1 (33:38):
They need to put a different hand on the robot
to pick up a little thing versus pick up a
big sure, or is it optimal?
Speaker 2 (33:44):
And you know this will be the this gets into
customer requirements question rather than fundamental science question. Yeah, that's
your that's for your product person to figure out. Yeah,
but that's how this is going to evolve. That's how
we're going to get to dexterous manipulation in the world
in a way that really is meaningful.
Speaker 1 (33:59):
Investors, one job at a ton, Yeah.
Speaker 2 (34:01):
Figure out if you've got the first like you know,
the first four or five jobs, and now you've got
your requirements, yours of requirements and your measurements and your
metrics and now engineers are going to be able to
iterate on that really make something work. But just trying
to copy a five figure in hand and say now
now it's an AI problem completely false, that's just not
going to happen.
Speaker 1 (34:24):
We'll be back in a minute with the lightning round.
Let's finish with the lightning round. I've read that you
like to jog at night for work, but like, are
(34:45):
there specific things that have happened to you or that
you've done to try and sort of you know, put
yourself in hard walking or running settings and where you've
actually had a thing happened and thought, oh, we need
to make sure the robot doesn't fall over when X happens.
Speaker 2 (35:00):
Honestly, it's more like when you're taking a hike, you know,
and you kind of get into the mode of you're
just your long day and a long hike and you know,
watching and being mindful, I guess, and thinking about what
are my feet doing and how is the contact progressing
with the ground and how does that feel and why
is that happening? And sort of daydreaming while you're thinking
(35:21):
that through is a really good way then to start
to recognize connections to research papers and connections to things
that people have found scientifically, and then start to pull
together hypotheses about how you would implement something like this
on a robot, or what is necessary because we're human,
and what is necessary because it's fundamental to look emotion
(35:43):
like why do we have feet? That's very different from
bird feet atreus didn't have feet, you know, And then
what exactly do we get anyway? Thinking through all that
kind of stuff.
Speaker 1 (35:56):
C three po or R two D two.
Speaker 2 (36:01):
Why that is a tough one. I'm going to say
R two D two.
Speaker 1 (36:05):
Not the bipedal one, not because that walks on two legs.
Speaker 2 (36:10):
C three, Well, no, I'm going to change my mind.
Speaker 1 (36:13):
I buyased it. Wait, why were you going to say
or two I was going to say to.
Speaker 2 (36:16):
You because C three pos a protocol droid And if
it's just about language, why do you have legs? Was
On the other hand, you.
Speaker 1 (36:24):
Just you don't need a robot at all, right, you
just need a little phone or whatever.
Speaker 2 (36:27):
On the other hand, right, it's a human centric robot.
It's a robot that's meant to be a translator and
be existing with people in the room and so on,
and so actually making something that's more of a humanoid
makes more sense if it's basically a social robot.
Speaker 1 (36:41):
Yeah, okay, Uh, what's something you wish that a robot
could do for you, you know, outside of work.
Speaker 2 (36:50):
Honestly, what I want is for robots to make everything cheaper.
Speaker 1 (36:55):
Uh huh. I love it when technology makes things cheaper. Well,
and that's what's been happening for the past couple hundred years,
and all of automation has effectively made everybody richer effectively
and improved everybody's quality of life because everything is cheaper.
And I want to be able to book my vacation
to the moon. I want to be able to not
really worry about you know, as it is, I don't
(37:17):
really worry about how much my phone costs or whatever.
If it breaks, it's not an expensive phone, it's fine,
it works, and all my software to transfers over. I
want everything to be like that. I want my lifestyle
to be supported by things that don't really cost a lot. Yes, well,
and I mean in the very long run, the sort
of technology driven productivity gains lift people out of poverty. Right,
(37:40):
everybody cares about our phones, but there are a lot
of people who have three dollars a day right now,
and it would be great if they could get to
three hundred dollars a day.
Speaker 2 (37:48):
Or beyond that. It's just that, you know, all the
things that we need become very easy to acquire. Food
is no longer, labor on farms is done in an
automated way, Labor in terms of logistics and transporting things
has done in an automated way, and so all of
this stuff becomes so affordable that it's easy to uplift
the quality of life of everybody on earth. That is
(38:08):
what a lot from Robotics.
Speaker 1 (38:15):
Jonathan Hurst is the co founder and chief robot officer
at Agility Robotics. Today's show was produced by Gabriel Hunter Chang.
It was edited by Lydia Jane Kott and engineered by
Sarah Bruguer. You can email us at problem at Pushkin
dot FM. I'm Jacob Goldstein, and we'll be back next
week with another episode of What's Your Problem
Popular Podcasts
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations.
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
The Nikki Glaser Podcast
Every week comedian and infamous roaster Nikki Glaser provides a fun, fast-paced, and brutally honest look into current pop-culture and her own personal life.