April 30, 2026 • 40 mins
Aaron Edsinger left his job as director of robotics at Google to start a company called Hello Robot. Aaron’s problem is this: How do you build an affordable robot that people can use to solve real problems at home? The result is a robot that looks nothing like a person. In fact, it’s closer to a Roomba with an arm.
In this episode, Aaron explains:
- Why home robots have barely progressed since the Roomba
- Why simpler robots can be more useful than complex humanoids
- Why the physical world is so much harder for AI than language
- How robot safety is fundamentally a physics problem
Connect with us:
- Follow Jacob Goldstein on LinkedIn, X and Instagram
- Email us at problem@pushkin.fm
- Follow Pushkin on Instagram, LinkedIn or X
- Listen to Jacob’s other show, Business History
- To listen to the show early and ad free, sign up for Pushkin+
See omnystudio.com/listener for privacy information.
Listen
Watch
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:15):
Pushkin. If I close my eyes and picture a robot,
I see basically a metal guy arms, legs ahead, maybe
a couple of lights for eyes. And there's a reason
I see this. This is the way people have been
dreaming of robots for something like a hundred years now,
(00:38):
and today people are doing more than dreaming. They're spending
billions of dollars to build robots like this, to build humanoids,
to build robots that look like people. But it's really
hard to build a robot that looks like a guy.
The physics are complicated. You need a lot of moving parts,
which is why despite the amazing demos we keep seeing
(01:01):
on social media, and despite all the smart people and
all the money trying to make robots happen, almost nobody
has a humanoid robot at their house. Maybe another approach
would make more sense. Maybe instead of starting with the
human body and saying how can we make a robot
that looks like this, we should ask how can we
build an affordable robot to help people solve real problems
(01:24):
at home? Right now, I'm Jacob Goldstein, this is what's
your problem? And my guest today is Aaron Edzinger. Aaron
has been building robots for decades. A while back, he
(01:46):
sold a company to Google, then he became the head
of robotics at Google, and then he left to start
a company called Hello Robot, where he is now the CEO.
Aaron's problem is this, how do you build a robot
that is useful and affordable for at least some people
to buy and use at home. Hello Robot hasn't really
(02:07):
solved that problem yet, but they're working on it. They
have a robot called Stretch three. They've sold hundreds of them,
and the robots are proving truly useful to some people
in very difficult circumstances. Also, the robots don't look anything
like metal people. They're not humanoids. In our conversation, Aaron
(02:28):
and I talk about why home robots have barely progressed
since the roomba, what it actually takes to get a
robot to feed someone, and why the physical world is
so much harder for AI than language. One of the
interesting things to me about your approach about Stretch three,
the robot that you have in the world now is, frankly,
the things that it's not right. It's not autonomous, it's
(02:51):
not a humanoid, it's a robot. Arm that's operated by
remote control. And you know you've been in the field
for a long time. You were the director of robotics
at Google. So why did you leave Google to build this?
Speaker 2 (03:06):
Oh it's it yeah, complicated answer, it's a good question.
In the end. You know, with envisioned robots being something
different than what they were becoming at that time, one
part of that is just about simplicity, right, So you know,
more and more the robots that we're building at Google
(03:29):
were becoming more complex, and you see that with the
humanoids today. There's this push towards kind of human level
dexterity capability. And our premise was, you know what, you
could do a lot today if you push on simplifying
the design, the approach. So the whole premise with Heller
robot is how do we practically deliver a value with
(03:52):
a robot to help people? And forget about should it
look like a human Let's really focus on just that
core question. And from that came to Toutch three, which
is a very minimalist robot. We like to think of
it almost like robotic cubism, right, So if you take
a human form and decompose it, put it back together,
in some other way. If you squint, it's recognizable as
(04:15):
a human. Maybe it can do eighty percent of what
a human or a humanoid could do. But it's much
more practical, deployable from a business perspective, but also just
from a usability and also autonomy. So the robot is autonomous.
We don't highlight a lot of that, but the autonomy
gets easier too when the robot gets simpler. So that's
(04:37):
the basic premise of starting the company.
Speaker 1 (04:40):
Yeah, I mean it's interesting as you were saying that,
I was thinking of sort of the classic software thing
is the minimum viable product, right, Like, what's the simplest
thing we can get out into the world and then
learn what's going on with it, iterate, etc. It's like
you're building the minimum viable robot exactly exactly, So you
have built a robot and it is in the world.
(05:03):
The current version is called Stretch three. Tell me about
Stretch three, so.
Speaker 2 (05:08):
S Touch three a third generation. It is fundamentally think
of it like a roomba with a big mast on it, right.
Comes from that perspective, right, and then you add to
that mast a telescoping arm, you know, telescoping the way
a car Antenta might telescope a turned sideways and you
put on the end of that telescoping arm a hand, so.
Speaker 1 (05:31):
It's basically an arm. There's a room b with a
vertical pole going up what three feet or something about
four feet yep, four feet, and then basically a robot
arm what we think of as a robot arm on
top of that pole. Exactly a room but that can
drive a robot arm around.
Speaker 2 (05:44):
Your house, drive around, and it can reach up to
the countertop underneath the couch. That type of thing. Very
very simple design, and what we saw early on is
just by remote controlling it, you can do quite a
few things around a house that you care about. And
so our first launch video showed it picking up toys
tele operated, but picking up toys, wiping down countertops, putting
(06:07):
laundry in a laundry machine with a very simple device.
You know, I just come from Google, where we're building
a very complex device and to do the same type
of thing. And I just thought, you know, if these
robots are ever really going to get out to the
world as real products, it's going to be and look
and feel more like this.
Speaker 1 (06:25):
How much does it cost?
Speaker 2 (06:27):
Right now? It's twenty about twenty five thousand dollars. And
that's that's because we run the company off of sales.
We're not it's trying to you know, it's low volume.
Speaker 1 (06:37):
So does that mean your your marginal cost is about
twenty five thousand dollars, you're selling it for about what
it costs.
Speaker 2 (06:43):
No, so the opposite. So what I'm saying is it's
actually can be very low cost, but because we're funding
R and D and engineering and production.
Speaker 1 (06:51):
You have a margin on it.
Speaker 2 (06:53):
We have a real, operational, sustainable business selling robots today.
Speaker 1 (06:57):
So you're a real company.
Speaker 2 (06:58):
We are a real company selling robots to people. Yes,
And that was part of the strategy from the get gos.
You know, I've been to robotics for twenty five years
now and seen a lot of companies over promise what's
possible with the technology, and then you know, their capital,
(07:19):
their investors getting patient, and they end up closing up shop.
And we see this time and time again.
Speaker 1 (07:24):
We're selling to Google and a happier version.
Speaker 2 (07:27):
Yeah, well you could do that too, which yeah, but
you know, really we wanted to have a viable business,
built a product that people liked and loved, and then
from there then you have permission to take the time
that it needs to really to do it correctly.
Speaker 1 (07:43):
How many have you sold?
Speaker 2 (07:45):
Yes, we have about three hundred robots in the world
across twenty three countries, and about forty five to different institutions,
from top tier research labs to small startups, to people
looking at vertical farming, to people looking at healthcare, you know,
a whole range of things.
Speaker 1 (08:04):
I saw a video about a man named Henry Vince
who uses one of your robots, and I found it
quite compelling. Yeah, tell me about Henry Evans, right.
Speaker 2 (08:18):
Henry and his care partner and wife, Jane. You know,
they're in California, not too far from us. They've collaborated
with my co founder, Charlie camp for at least I
think fifteen years now, so it's been a long relationship.
Henry is quadriplegic from a stroke. Jane is the primary caregiver.
(08:39):
But Henry is, you know, an early adopter. He's really
excited to find ways that technology can give him agency,
not just that, but make life easier on chain and
so you know, he's been looking at robotics as a
solution partial solution for that for quite some time. We've
collaborated with him, and really when we found it, Hello robot,
(09:01):
you know, he was one of the key well i'd
say inspirations and users end users that we had in mind.
So through some some work that was funded by the NIH,
we've spent it last three years working with him and
some other people in similar situations looking at how Stretch
can can benefit him. And so, you know, Henry isn't
(09:22):
able to communicate, to talk except through I gays, so
he uses I gaze on a letter board to spill
out words and then he can also use I gays
to control a computer mouse, and through that he's able
to control Stretch to do simple tasks, things that might
seem you know, trivial, are you not so exciting to
(09:43):
other people, but for him, it's it's it's it's really
life changing. And so a really really nice story with
with Henry is, you know, we came in thinking like, well,
you know, Stretch will do very practical tasks very quickly.
What we learned is there's a whole range of things
that are much more social, emotional about connection that we
(10:05):
just didn't anticipate with the device like this. And so
one example is his granddaughter at the time was probably
about three or four. She was a little bit scared
of him because you know, he's immobile, he's in a
powerchair in bed, can't talk, and so they didn't really
have a relationship. And what we found is through Stretch,
(10:27):
because he could control Stretch, he could start to play
with her, interact with her, and in the physical world.
She would decorate stretch, put stickers on it. They would
have a tablet on it, so you know, his face
would appear, and soon they started to play. So they'd
played basketball, pick up, a little nerve, basketball, play different games,
(10:49):
and you know, she started to really kind of see
him in a different light and former relationship with him,
and so now she calls him Papa Wheeli because he's
in his chair and and she, you know, can be
excited to see him. So it really bridge that gap
between the two of them, and and the type of
thing that with robot that there's so much potential and
(11:12):
I think we really don't quite know how far we
can go on the sort of the social side of it.
You know, really this can be this empathetic technology that
brings so much to people, and that's that's kind of
what we're excited to explore.
Speaker 1 (11:29):
As you were talking about the robot playing with a
you know, if you're a young child, the question of
safety crossed my mind, Like what do you have to
worry about in that setting? How do you deal with it?
Speaker 2 (11:41):
Like?
Speaker 1 (11:41):
How much is the robot weigh? Yeah, it's a great question.
Speaker 2 (11:44):
So this robot weighs about fifty pounds, a little over
fifty fifty pounds most that weighs down in the base, right,
And it's really it's an important question because there's so
much happening in robotics right now where we're seeing humanoids
in the home doing dishes right, and these are scary machines.
(12:05):
And really the fundamental thing you have to think about
is just that physics of it. Right, when you have
a humanoid that's balancing, particularly a human scale one, the
amount of weight that's up high, the potential energy that's
there is significant, and when things go.
Speaker 1 (12:23):
Wrong, if it falls over, you could smush things. That's fundamentally,
that's fundamentally it.
Speaker 2 (12:28):
And so just the pure physics physics of that are
fundamentally you know, we can't cheat it, right, you can't,
and things things will fail. So really, with from the
get go for for hello, robot and stretch. We thought
about safety from just the physical properties of it. How
do we make it so it's very hard for this
(12:49):
device to cause harm? So first off, it should be
light weight, that potential energy, the mass should be down
low so that if if it falls over, it's not
gonna you know, smash and break something. And then you
think about the robot joints themselves.
Speaker 1 (13:06):
And is there is there a pincer We haven't actually
talked about the sort of yes, gripper is it? Is
it like a two piece pincer.
Speaker 2 (13:12):
It's a it's basically like a grabber that you buy
an Amazon, you know, like to pick up trash. Yeah,
it's one of those that we've attached to motor to essentially.
But if you think about, you know, the robot kind
of motors. If if you have to hold your arm
out against gravity all day long, you know you're gonna
need a big motor there, right, because that's a lot
(13:33):
of energy to support that. If you don't, if you
can design the robot in a minimal way where you're
not spending all of your time fighting energy or first off,
you're more energy efficient. Second you have a smaller motor
which will be less dangerous. And so the arm moves
in a way, it is designed in a way that
it's n have to fight gravity, and that just makes
(13:53):
it inherently safer. Right, And so when you're working around
older adults, people with disabilities, kids, pets, these things really
really matter. And so that's that's just the starting point
from the design. The intrinsic safety has to be there.
The physics of it has has to be right.
Speaker 1 (14:12):
So people with quadriplegia, people with extremely limited mobility are
an obvious early use case. What else, like, what else
are people using this for? What else do you have
in mind for early use case?
Speaker 2 (14:26):
Yeah, one of the most interesting use cases is in
the home around caregiving. It's a very challenging one. It
is further out, and so you know, there's a lot
of use cases nearer term which I think will be
in service, you know, back of retail logistics. Uh, And
we're seeing you know, data centers, a lot of stuff
(14:47):
like that that we are seeing interest in and people exploring.
Speaker 1 (14:50):
Like picking and pecking. I mean there are specialized robots
for those things.
Speaker 2 (14:54):
Right. Well, yeah, there's some places where there's not yet, right, So,
for example, micro fulfillment little fulfillment, you know, like even
like a you know, stocking a seven to eleven. There's
lots of things where a robot like this could play
a role.
Speaker 1 (15:09):
Options And just to bit clear, my understanding is that
your robot is mainly not autonomous, that it's mainly remote
controlled tteley operated. Is that right? And what does that
mean for that use case?
Speaker 2 (15:21):
Right? It can be autonomous. There are autonomous it's completely
open source as a whole ecosystem of open source software
people have built for it, including autonomy, and the generation
of robot we're working on now will really be autonomy powered, right,
so in terms of the sensing of the software. But
(15:42):
you know, back to that sort of really minimal approach. Yeah,
we're trying to say what can we do today that
has value? And for someone like Henry Evans, he controls
the robot, it has autonomy mixed in there, so he's
not controlling every motion necessarily, and so what we see
is as a steady progression from direct teleoperation to mixed
(16:06):
autonomy to full autonomy, but doing that really deployed, right,
So this is not something in a research lab. It's
actually in the product. And you know, we also see
a path from people with really severe mobility impairments like Henry.
They're interesting because they have a high motivation to use
(16:27):
the device, and it's been the time to control it, right,
They've got time and motivation. And then you can layer
in the economy over time, and as you do that,
now you can broaden the set of people who can
who benefit from it. I think from people like Henry.
The next step in terms of in home care is
(16:49):
a remote operation by a family member a care provider.
You know, for example, I'm in New Orleans right now,
my mother is back in California, and you know, I'd
love to hop on and check in, and you know,
of course you can zoom, but sometimes you need to
go into the other room, open the cabinet door, and
you know, find the medication or something like that. And
(17:12):
so again there's a human in the loop, but they're
not going to be controlling every motion. They're going to
be clicking on a map and saying go to the bedroom, open,
open the drawer, that type of thing. Yeah, And then
from there, of course, you can bring in more autonomy,
and suddenly you have in home autonomous care and eventually
in home robots that are doing the laundry and dishes
(17:34):
and everything that you know, people are excited about. But
for us, again, we're taking this very pragmatic approach about
how do we stair step, how do we kind of
climb our way up to that future?
Speaker 1 (17:47):
Stairstep is an interesting choice of words for a robot
on wheels. But you know, one of the things people
always talk about in the context of robots in the
last few years is getting the data to train the
model for autonomy, right, And the obvious analog is large
(18:08):
language models got so good because the Internet is full
of text, right, and there is no Internet of the
physical world. There is no giant data set that you
can just train a model on so that robots can
understand how physics works in the real world. Right, how
are you dealing with that?
Speaker 2 (18:28):
You know, I would say, in terms of AI physical
AI as it's called around robotics, we're taking, you know,
an approach that's really about taking things that are mature
and integrating them into our autonomy, versus what's called end
to end learning, where you know, you might collect a
huge data set, right, you might train the robot tele
(18:51):
operated to open the dishwasher a thousand times, collect that
data and train a deep learned model to do that task,
and you see very exciting results from that, and I
do think that will continue to progress, but in terms
of being ready to be deployed, it'll be take some time.
And the amount of data that's required beyond just a dishwasher, Well,
(19:15):
now it's a different dishwasher. Now it's a dishwashering different
lighting conditions. Oh there's something in front of it. That
complexity gets so large so quickly.
Speaker 1 (19:23):
And it doesn't generalize, like there are not models where
you can teach it one dishwasher and it can make
inferences about other dishwashers.
Speaker 2 (19:31):
There's inklings of that happening, right, but we're not there yet,
and resourcers are still really exploring what's possible. You know,
some are saying it's going to learn from YouTube, some
are saying it's going to be all on simulation. Some
are our hiring teams of operators to collect that data. Right.
Our view is, you know that that world is progressing
(19:52):
quickly and it'll sort itself out soon. You know, we
will be able to train the robot to open that dishwasher.
I think what's more interesting is how do you get
it to do you know the five or six important
tasks you care about and actually do it in a
robot that people feel is trusted in the home, it's
safe in the home, is affordable, you know, all the
(20:15):
things that actually to have a business.
Speaker 1 (20:17):
Yeah, so you're basically it's again the like the physical
intelligence is the bright shiny object that you are not
chasing so that you can build a robot that works exactly.
Speaker 2 (20:27):
And and you know, think about uh, you know, you know,
is Apple building a big AI team? Well, no, it's
partnering with the people who are doing the AI. And
I think you know, for us, we're more on that
side of the fence.
Speaker 1 (20:42):
Are some of those AI gains that you are sort
of piggybacking on coming from autonomous vehicles? I mean when
you talk about saying go to the other room and
have it go there, I think of that, That's where
my mind goes, right that those are the robots out
in the world right now driving around San Francisco.
Speaker 2 (20:58):
Yeah, I mean it's slightly different in terms of the
requirements the sensors, but it's certainly certainly related. I mean,
we've all benefited from he's wrapping gains and sensing and
compute and and you know even ll MS is they're
being used to help their robots plan and think and replan. Right.
(21:19):
So you know, when I was started in robotics twenty
twenty years ago, I couldn't imagine being able to say
to the robot, hey, go to the other room and
find the blue mug that's next to the microwave. Right,
Like the context of that was just like, how are
we going to solve that? Are we going to write
every rule into the robot and that's.
Speaker 1 (21:38):
Like commodified like those are That's that's the easy part
now at some.
Speaker 2 (21:43):
Level, yes, I mean certainly there's more to be done.
I'd say, you know, the real challenges today are going
to be around desk to manipulation and being able to
do things that are really in dynamic, you know, kind
of home type environments.
Speaker 1 (22:00):
So when you when you talk about you know, dexter
sort of fine motor skill type things, what are you
thinking of? What do you want the robot to be
able to do in the relatively near term that it
can't do now?
Speaker 2 (22:11):
I think just broaden the range of tasks. Right. So
one thing that's really exciting to this user group is feeding, right,
I think feeding. Feeding is one of the things where
right now, so your partner care partner has to come
and feed you, and that there's a number of reasons
why that's just they would prefer to do it themselves.
Of course.
Speaker 1 (22:30):
The basic it's a basic function of adult life is
you feed yourself.
Speaker 2 (22:36):
You feed yourself, and depending on someone changes that relationship
with that person, right, And and so feeding is one
that we've looked at a fair amount. It is challenging.
There are safety considerations, so we're careful about what utensils
and what type of food. But I think if it
can feed, if you can feed yourself, if you can
(22:58):
get yourself a glass of water, you can do basic
you know, hygiene maintenance things for yourself. I think that's
kind of the base line of regaining a sense of agency, right,
And what do you have.
Speaker 1 (23:08):
To figure out to make feeding really work that you
haven't figured out yet?
Speaker 2 (23:12):
Yeah, it's a mixture of we have. Food is kind
of messy, mushy, right, it's not like a rigid object.
Speaker 1 (23:20):
Well, and it's it's super heterogeneous, right, Food is many
different kinds of.
Speaker 2 (23:25):
Ways exactly, and so but there is you know, one
thing we've seen is you know, the caregiver is willing
to prepare the food, maybe put it in a special tray,
certain types of food that we agree it can work with.
So but of course you'd love that to get broader.
But you know, food in itself is challenging. Then there's
the safety challenges. But also, you know, a really interesting
(23:46):
part of this, which kind of relates to you know,
where AI is in general, is understanding the person, right,
are they ready for another bite? You know, are they
you know, what side would they like to be fed from.
That's all these little things that are about that human
interaction that make the product useful, exciting, acceptable in the home.
(24:11):
And so much of what's happening in ai AI right
now doesn't really get at the human side of the story.
And I think that's one thing where we would love
to see a advancements is really having better understanding of
these sort of more subtle cues and interactions that the
robot will need in order to be accepted.
Speaker 1 (24:35):
We'll be back in just a minute. So I'm curious
because you have been working on robots for more than
twenty years, right, I mean, so much has happened in software, yeah, right,
(24:58):
and even in other kinds of hardware, Like you were
working on robots before there was an iPhone, right, and
so so much has changed, and it feels like, at
least as a as a layperson, certainly the home like
robots are not around the way iPhones are around. And
even now large language models are around. And so has
less happened with robots than you thought.
Speaker 2 (25:20):
I would say less in terms of deployed robots, right,
Like the rumba is still one of the most successful robots,
but there aren't many other examples like it, right, and
so why is that?
Speaker 1 (25:32):
Yes?
Speaker 2 (25:32):
What?
Speaker 1 (25:33):
Like that's a great way to ask it, and that
I mean, you worked with Rodney Brooks, right, one of
the rumbook guys at MIT, and that's like the only
robot anyone can name, and there are lots of failed
companies that were rumba.
Speaker 2 (25:45):
Like right, yeah, yeah, and so yeah, So I did
my PhD with rod Brooks at MIT and at the
time he was spinning up a robot, inventing the rumba.
And around the time I finished, he had just sold
his million rumba and he's he inspired me to go
off to my own company.
Speaker 1 (26:03):
I mean, it must have seemed like they're here home
robots are here, like you would think at that moment,
it's like, oh, hey, here we go.
Speaker 2 (26:10):
Yeah yeah, well but things seemed very hard, right, so
it's like, okay, a vacuum cleaner that bounces around off walls,
like we can do that, right, Yeah. And I think
the insight from Rumba that I still carry is, you know,
they thought really carefully about the price, the complexity of
the device. They weren't trying to There was a competitor
(26:32):
call the Electrial Lucks at the time that was about
three thousand dollars had all sorts of cool features and
computer and sensing. But with the Rumbo, what they did
is they went to the mall and they asked people,
how much can you spend on something without having to
ask your significant other for permission?
Speaker 1 (26:49):
Oh?
Speaker 2 (26:50):
And the answer was about two hundred dollars.
Speaker 1 (26:53):
I was just talking to a friend who basically does
a B to B business. That's business a business, and
apparently that's a sales strategy in business as well. Like
the person who you're selling to they have some spending
authority that they can spend on their own. Yes, of course,
it's about a thing times greater. Right, It's like they
can spend one hundred thousand dollars or whatever without asking
(27:13):
somebody else, So you really want to price it like that,
otherwise it's much harder.
Speaker 2 (27:17):
So anyway, yeah, exactly, But you know, I think the
room to success in part was they made a minimal
product within that spending, you know, authority and kind of
made it all work. And I think the reason we
haven't seen a lot beyond that is it's hard. It's
been hard traditionally to find the things that match all
(27:37):
those constraints. That technology can be there, the use case
can be there, right clearly in home care, there's so
much need for help in the home, but can we
do it at a price? And is it technology there?
You know, five thousand dollars would be a lot of
money for most families.
Speaker 1 (27:54):
Yes, although for a caregiver, right like, if you narrow
it not to like doing the dishes.
Speaker 2 (27:59):
But yeah, I.
Speaker 1 (28:01):
Mean, you know, many people have been bankrupted by just
needing to hire a simple caregiver, right Like, people could
definitely figure out how to spend five thousand dollars on
a robot that could significantly replace a caregiver. One would think,
I guess that's your bet, that's your.
Speaker 2 (28:20):
That's partial partially our bet. Yeah, and you know, I
think it's not replacing, it's it's augmenting, you know, I
think think what makes sense, yes, because we're not going
to do everything, and nor should we, you know, but
there's just aren't enough people ours caregiving hours in the
in the world today to take care of everyone who
needs care. That's just the truth of it.
Speaker 1 (28:39):
Yeah, zooming out from from Hello robot, from your company
to the field more generally, like what's the landscape look
to you? How does the landscape look to you right now?
Speaker 2 (28:48):
Yeah, it's fascinating there. There's a lot of first off,
a lot of capital flowing into it. Into two areas really,
one is robot foundation models. You know, this is sort
of deep learning the alms of robots. And then the
other would be humanoids, right, and sometimes there's both, right,
but there's a lot of excitement around that, and would
(29:09):
say generally they both have a little ways to go
before they get deployed and really applied.
Speaker 1 (29:15):
I mean the foundation model one, you know, an ll
M for robots sounds good to me, It seems interesting
to me. Tell me about what's happening there.
Speaker 2 (29:24):
Yeah, Well, it's first off, it's remarkable what you see
the robots doing with these these models, right, so peeling
an egg or tying shoelaces or folding laundry. And they're
getting better, right, They're they're getting they're having long I think.
Speaker 1 (29:41):
The peeling an egg was teleoperated.
Speaker 2 (29:43):
Okay, fair enough, but they are doing things like that
that you think, wow, I just could not imagine this.
And and they're getting better, and that they're having longer
memory so they can do stuff over a longer period
of time. They're getting better at recovering from from issues.
But also they haven't yet learned to generalize. Right, So
(30:03):
if you think about what they're learning, that's such as
we were saying, such richness to diversity in the environments
they meant to operate, and they're learning on one kind
of little island and the whole sea of data that
they need access to to really be generally useful. And
so that's kind of a real challenge that that field
(30:23):
is trying to work through, is how do you get
to that generality and that robustness to that diversity. And
it's very much an unsolved problem. I think they're making progress,
but it also means to me that this is not
something that we're going to be deploying into a robot
product anytime soon.
Speaker 1 (30:42):
You know, it's interesting because it seems like one of
the frankly big surprises about large language models is how
good they are at generalizing, right, how they often are
much better than specialist models.
Speaker 2 (30:57):
Right.
Speaker 1 (30:57):
Just take a frontier model and throw anything at it
and it will be really good. Not quite, but kind of.
And so it's interesting to hear that that's not at
all the case with robot foundation models. Why do you
pose that.
Speaker 2 (31:09):
Is, Well, you know, I think one thing is just
the amount of data that's available, Right, So if you
think of a ll M, one simple way to think
of it, it's going to predict the next word that
you say, and it has the whole corpus of the
Internet to make that prediction. Right, for a robot to
predict the next thing that's going to happen when it's
(31:30):
peeling the egg, there's not a lot of data out
there for it to understand that. And some people are saying, well,
I can do that in simulation, but very hard to
imagine simulating a lot of that kind of complexity that's
actually in the world, just.
Speaker 1 (31:47):
Because like the world is so diverse. What do you
even simulate or how.
Speaker 2 (31:50):
Do you even simulate an egg being peeled? And so,
you know, I think it's a challenge because robots are
in the physical world.
Speaker 1 (32:00):
Is there something as I understand it, the basic model,
the transformer model that powers all of the frontier elements
today was developed based on the nature of language itself, right, Like,
is there a thought that that's maybe just not the
best fundamental model for understanding the physical world for training robots.
(32:23):
Might it be the case that some very different approach
ends up working much better.
Speaker 2 (32:28):
Yeah, I know less about that. I do know there
is a large interest in world models today that can
kind of infer the physics of the world. And I
think one thing that that is beneficial is there is
so much structure in the world, both from physics.
Speaker 1 (32:44):
Like laws of physics, laws of physics, yep, but also.
Speaker 2 (32:47):
In the way that we design our homes, our appliances
are you know, there is a lot of structure to
be gained and learned, and so you can imagine these
models become very good at predicting what your home is
going to look like because it's seen a lot of
homes and they don't vary that much fundamentally, so so
that there is there is an enough structure to learn
(33:10):
from what that model and how that transpires. I don't know,
but people are exploring that.
Speaker 1 (33:16):
I mean, one of the one of the obvious challenges
is we're just making the robots come live in our world, right,
And I could imagine, for a case like someone with quadriplegia,
it might I could imagine a universe where it's like, well,
if you set up your room this way, not you
have to remodel your house, but you have to do
some set of things, then our robot can do a lot.
(33:36):
And I can imagine that being worthwhile in many settings.
Is that something you are doing or thinking about? Sort
of a compromised solution like that?
Speaker 2 (33:44):
Absolutely? And you know, as an example, a Stretch three
as it is today, has a hard time opening some refrigerators, right,
It's just the magnetic force is too high. But you
can attach a little you know, grabbing thing that breaks
the seal and makes it easy to open it. Right,
And so it's a three dollars apart, you stick it
onto your fridge. Now Stretch can open their fridge for
(34:07):
you and for for these users, you know, those accommodations
make a lot of sense, and uh, you know, I
do think you know, maybe over time you'll have special
plates at the robot nose it cannot you know, grab
and not break, and you know, there will be accommodations
that people make for it. But you know, for for now,
(34:29):
we are using and counting on those types of accommodations.
Speaker 1 (34:33):
So I want to do two kind of forward looking
things first, the sad one and the happy one. Basically okay,
first is how might it not work out? And then
the second is what does it look like if it
does work out for the low robots, So like what
might go wrong for for your company in a you know,
significant way, right.
Speaker 2 (34:57):
Well, there's there's always the risk of just the business
doesn't work out. I think, I think from from the
technology and product side, it's really about acceptance, right, So
I think implicit to bringing this technology and your life
into your home, there has to be trust, right, and
people have to trust and want this technology around them,
and so I think, you know, there there's always a
(35:19):
risk that people like you know, what we're going to
we're going to find different ways to addrust the caregiving problem,
you know, which you know, I should. There's lots of
going to be lots of solutions to that, but robots
may may not turn out to be part of that story.
We're bullish that it will, but there it's a real
kind of complex dynamic between the people and the technology.
(35:41):
So that's that's sort of the the bare case. The
bowl case would be, you know, ten years out. Call
it a five thousand dollars device that's in your home
that can provide caregiving support in the times when there's
not care present. So basically it's just kind of facilitating
quality of life at home, agency at home as you
(36:02):
want to kind of age independently without having so when
you're when your grandkids come over, you're not they're not
doing chores for you. You're interacting, right because of the
robots helping fill that gap.
Speaker 1 (36:17):
We'll be back in a minute with the lightning round.
Let's finish with the lightning round.
Speaker 2 (36:35):
Okay, do you.
Speaker 1 (36:37):
Still make art?
Speaker 2 (36:38):
Oh? I not for a while, But that is how
I got into robotics, was doing robotic sculpture, and then
that led one thing led to another. But yeah, really
fascinated by well, love art but also love you know,
kind of robotics and how it relates to art.
Speaker 1 (36:58):
What's the least practical robot you've ever built?
Speaker 2 (37:02):
Wow? Uh? So let's see in the in the nineteen
nineties in San Francisco, we had a robotic art performance
troupe and it was a theatrical multimedia performance of robots
on stage acting performing. You know, this was a kind
of robotic gallery of kind of characters from the from
(37:25):
the local neighborhood. So these were a street musician and
you know, someone who was got of causing trouble and
so they were kind of robotic denizens of the Tenderloin
of nineteen nineties, San Francisco put onto stage.
Speaker 1 (37:43):
What was the first robot you ever built?
Speaker 2 (37:46):
It was actually the Robotic Street Preacher. So it was
a little robot that would go around and big spinning
head of a speaker and it would be on stage
and perform and we would sit there and tell operate it.
And it had, you know, no real autonomy. This is
you know, everything with the hands salted, handbuilt and their
(38:06):
computers were super tiny, are super a small capability, And
we put that on stage and performed it.
Speaker 1 (38:15):
Did you ever try and build robots when you were
a kid?
Speaker 2 (38:19):
No, you know, Strangely, I never had a fascination with
robusts as a kid. I came out of grew up
in Washington State on a farm, and uh, you know,
I think my interest in robots came from having to
do farm work and learning about automation.
Speaker 1 (38:36):
Like Henry Ford. You know, Henry Ford was the same
story about a farm and thought that was too much
labor exactly, and I just thought, you know, there's got
to be a better way than this.
Speaker 2 (38:45):
My my mother's family, which is Mexican, had a history
and kind of uh kind of working in the fields
and kind of migrant labor, and so kind of the
relationship to labor and work kind of inspired my interest
in automation and then roproducts.
Speaker 1 (39:02):
Did you ever try and build a machine to do
your chores for you?
Speaker 2 (39:05):
I absolutely did. It never worked, but it was fun,
you know.
Speaker 1 (39:10):
I mean, I guess that's still what you're working on
at some level.
Speaker 2 (39:13):
Exactly.
Speaker 1 (39:15):
What's one thing you've learned working with Frank Gary?
Speaker 2 (39:17):
Oh? Right, So we had a project looking at this
conceptual but a large robotic kind of botanical garden and
this sole thing in Singapore, so that we were looking
at developing that with him and Uh, you know, I'd say, actually,
the thing that was most interesting is just he really
(39:41):
put a lot of trust into the young people in
the studio and really was generous in that way. And
you know, it's been an inspiration for me.
Speaker 1 (39:50):
That's a lovely surprise one. This is I guess some
bias that I have. I imagine a famous architect as
not being like that.
Speaker 2 (39:58):
So to hear that so lovely, very very generous in
a way that can be inspiring to to the to
people in their career. So I would try to carry
that forward in my own work.
Speaker 1 (40:09):
Uh, thank you for your time.
Speaker 2 (40:11):
Yeah, it's been a pleasure. Thank you.
Speaker 1 (40:20):
Aaron Edzinger is the co founder and CEO of Hello Robot.
Today's show was produced by Gabriel Hunter Chang and edited
by Lydia Jane Kott. Our engineer this week was Hansdale Sheet.
We're always looking for ideas for who to talk to
and what to cover on the show. You can email
us at problem at Pushkin dot fm. You can find
(40:42):
me on x at Jacob Goldstein. You can find me
on LinkedIn. I'm Jacob Goldstein. Thank you very much for
listening to the show, and we'll be back next week
with another episode
Pushkin. If I close my eyes and picture a robot,
I see basically a metal guy arms, legs ahead, maybe
a couple of lights for eyes. And there's a reason
I see this. This is the way people have been
dreaming of robots for something like a hundred years now,
(00:38):
and today people are doing more than dreaming. They're spending
billions of dollars to build robots like this, to build humanoids,
to build robots that look like people. But it's really
hard to build a robot that looks like a guy.
The physics are complicated. You need a lot of moving parts,
which is why despite the amazing demos we keep seeing
(01:01):
on social media, and despite all the smart people and
all the money trying to make robots happen, almost nobody
has a humanoid robot at their house. Maybe another approach
would make more sense. Maybe instead of starting with the
human body and saying how can we make a robot
that looks like this, we should ask how can we
build an affordable robot to help people solve real problems
(01:24):
at home? Right now, I'm Jacob Goldstein, this is what's
your problem? And my guest today is Aaron Edzinger. Aaron
has been building robots for decades. A while back, he
(01:46):
sold a company to Google, then he became the head
of robotics at Google, and then he left to start
a company called Hello Robot, where he is now the CEO.
Aaron's problem is this, how do you build a robot
that is useful and affordable for at least some people
to buy and use at home. Hello Robot hasn't really
(02:07):
solved that problem yet, but they're working on it. They
have a robot called Stretch three. They've sold hundreds of them,
and the robots are proving truly useful to some people
in very difficult circumstances. Also, the robots don't look anything
like metal people. They're not humanoids. In our conversation, Aaron
(02:28):
and I talk about why home robots have barely progressed
since the roomba, what it actually takes to get a
robot to feed someone, and why the physical world is
so much harder for AI than language. One of the
interesting things to me about your approach about Stretch three,
the robot that you have in the world now is, frankly,
the things that it's not right. It's not autonomous, it's
(02:51):
not a humanoid, it's a robot. Arm that's operated by
remote control. And you know you've been in the field
for a long time. You were the director of robotics
at Google. So why did you leave Google to build this?
Speaker 2 (03:06):
Oh it's it yeah, complicated answer, it's a good question.
In the end. You know, with envisioned robots being something
different than what they were becoming at that time, one
part of that is just about simplicity, right, So you know,
more and more the robots that we're building at Google
(03:29):
were becoming more complex, and you see that with the
humanoids today. There's this push towards kind of human level
dexterity capability. And our premise was, you know what, you
could do a lot today if you push on simplifying
the design, the approach. So the whole premise with Heller
robot is how do we practically deliver a value with
(03:52):
a robot to help people? And forget about should it
look like a human Let's really focus on just that
core question. And from that came to Toutch three, which
is a very minimalist robot. We like to think of
it almost like robotic cubism, right, So if you take
a human form and decompose it, put it back together,
in some other way. If you squint, it's recognizable as
(04:15):
a human. Maybe it can do eighty percent of what
a human or a humanoid could do. But it's much
more practical, deployable from a business perspective, but also just
from a usability and also autonomy. So the robot is autonomous.
We don't highlight a lot of that, but the autonomy
gets easier too when the robot gets simpler. So that's
(04:37):
the basic premise of starting the company.
Speaker 1 (04:40):
Yeah, I mean it's interesting as you were saying that,
I was thinking of sort of the classic software thing
is the minimum viable product, right, Like, what's the simplest
thing we can get out into the world and then
learn what's going on with it, iterate, etc. It's like
you're building the minimum viable robot exactly exactly, So you
have built a robot and it is in the world.
(05:03):
The current version is called Stretch three. Tell me about
Stretch three, so.
Speaker 2 (05:08):
S Touch three a third generation. It is fundamentally think
of it like a roomba with a big mast on it, right.
Comes from that perspective, right, and then you add to
that mast a telescoping arm, you know, telescoping the way
a car Antenta might telescope a turned sideways and you
put on the end of that telescoping arm a hand, so.
Speaker 1 (05:31):
It's basically an arm. There's a room b with a
vertical pole going up what three feet or something about
four feet yep, four feet, and then basically a robot
arm what we think of as a robot arm on
top of that pole. Exactly a room but that can
drive a robot arm around.
Speaker 2 (05:44):
Your house, drive around, and it can reach up to
the countertop underneath the couch. That type of thing. Very
very simple design, and what we saw early on is
just by remote controlling it, you can do quite a
few things around a house that you care about. And
so our first launch video showed it picking up toys
tele operated, but picking up toys, wiping down countertops, putting
(06:07):
laundry in a laundry machine with a very simple device.
You know, I just come from Google, where we're building
a very complex device and to do the same type
of thing. And I just thought, you know, if these
robots are ever really going to get out to the
world as real products, it's going to be and look
and feel more like this.
Speaker 1 (06:25):
How much does it cost?
Speaker 2 (06:27):
Right now? It's twenty about twenty five thousand dollars. And
that's that's because we run the company off of sales.
We're not it's trying to you know, it's low volume.
Speaker 1 (06:37):
So does that mean your your marginal cost is about
twenty five thousand dollars, you're selling it for about what
it costs.
Speaker 2 (06:43):
No, so the opposite. So what I'm saying is it's
actually can be very low cost, but because we're funding
R and D and engineering and production.
Speaker 1 (06:51):
You have a margin on it.
Speaker 2 (06:53):
We have a real, operational, sustainable business selling robots today.
Speaker 1 (06:57):
So you're a real company.
Speaker 2 (06:58):
We are a real company selling robots to people. Yes,
And that was part of the strategy from the get gos.
You know, I've been to robotics for twenty five years
now and seen a lot of companies over promise what's
possible with the technology, and then you know, their capital,
(07:19):
their investors getting patient, and they end up closing up shop.
And we see this time and time again.
Speaker 1 (07:24):
We're selling to Google and a happier version.
Speaker 2 (07:27):
Yeah, well you could do that too, which yeah, but
you know, really we wanted to have a viable business,
built a product that people liked and loved, and then
from there then you have permission to take the time
that it needs to really to do it correctly.
Speaker 1 (07:43):
How many have you sold?
Speaker 2 (07:45):
Yes, we have about three hundred robots in the world
across twenty three countries, and about forty five to different institutions,
from top tier research labs to small startups, to people
looking at vertical farming, to people looking at healthcare, you know,
a whole range of things.
Speaker 1 (08:04):
I saw a video about a man named Henry Vince
who uses one of your robots, and I found it
quite compelling. Yeah, tell me about Henry Evans, right.
Speaker 2 (08:18):
Henry and his care partner and wife, Jane. You know,
they're in California, not too far from us. They've collaborated
with my co founder, Charlie camp for at least I
think fifteen years now, so it's been a long relationship.
Henry is quadriplegic from a stroke. Jane is the primary caregiver.
(08:39):
But Henry is, you know, an early adopter. He's really
excited to find ways that technology can give him agency,
not just that, but make life easier on chain and
so you know, he's been looking at robotics as a
solution partial solution for that for quite some time. We've
collaborated with him, and really when we found it, Hello robot,
(09:01):
you know, he was one of the key well i'd
say inspirations and users end users that we had in mind.
So through some some work that was funded by the NIH,
we've spent it last three years working with him and
some other people in similar situations looking at how Stretch
can can benefit him. And so, you know, Henry isn't
(09:22):
able to communicate, to talk except through I gays, so
he uses I gaze on a letter board to spill
out words and then he can also use I gays
to control a computer mouse, and through that he's able
to control Stretch to do simple tasks, things that might
seem you know, trivial, are you not so exciting to
(09:43):
other people, but for him, it's it's it's it's really
life changing. And so a really really nice story with
with Henry is, you know, we came in thinking like, well,
you know, Stretch will do very practical tasks very quickly.
What we learned is there's a whole range of things
that are much more social, emotional about connection that we
(10:05):
just didn't anticipate with the device like this. And so
one example is his granddaughter at the time was probably
about three or four. She was a little bit scared
of him because you know, he's immobile, he's in a
powerchair in bed, can't talk, and so they didn't really
have a relationship. And what we found is through Stretch,
(10:27):
because he could control Stretch, he could start to play
with her, interact with her, and in the physical world.
She would decorate stretch, put stickers on it. They would
have a tablet on it, so you know, his face
would appear, and soon they started to play. So they'd
played basketball, pick up, a little nerve, basketball, play different games,
(10:49):
and you know, she started to really kind of see
him in a different light and former relationship with him,
and so now she calls him Papa Wheeli because he's
in his chair and and she, you know, can be
excited to see him. So it really bridge that gap
between the two of them, and and the type of
thing that with robot that there's so much potential and
(11:12):
I think we really don't quite know how far we
can go on the sort of the social side of it.
You know, really this can be this empathetic technology that
brings so much to people, and that's that's kind of
what we're excited to explore.
Speaker 1 (11:29):
As you were talking about the robot playing with a
you know, if you're a young child, the question of
safety crossed my mind, Like what do you have to
worry about in that setting? How do you deal with it?
Speaker 2 (11:41):
Like?
Speaker 1 (11:41):
How much is the robot weigh? Yeah, it's a great question.
Speaker 2 (11:44):
So this robot weighs about fifty pounds, a little over
fifty fifty pounds most that weighs down in the base, right,
And it's really it's an important question because there's so
much happening in robotics right now where we're seeing humanoids
in the home doing dishes right, and these are scary machines.
(12:05):
And really the fundamental thing you have to think about
is just that physics of it. Right, when you have
a humanoid that's balancing, particularly a human scale one, the
amount of weight that's up high, the potential energy that's
there is significant, and when things go.
Speaker 1 (12:23):
Wrong, if it falls over, you could smush things. That's fundamentally,
that's fundamentally it.
Speaker 2 (12:28):
And so just the pure physics physics of that are
fundamentally you know, we can't cheat it, right, you can't,
and things things will fail. So really, with from the
get go for for hello, robot and stretch. We thought
about safety from just the physical properties of it. How
do we make it so it's very hard for this
(12:49):
device to cause harm? So first off, it should be
light weight, that potential energy, the mass should be down
low so that if if it falls over, it's not
gonna you know, smash and break something. And then you
think about the robot joints themselves.
Speaker 1 (13:06):
And is there is there a pincer We haven't actually
talked about the sort of yes, gripper is it? Is
it like a two piece pincer.
Speaker 2 (13:12):
It's a it's basically like a grabber that you buy
an Amazon, you know, like to pick up trash. Yeah,
it's one of those that we've attached to motor to essentially.
But if you think about, you know, the robot kind
of motors. If if you have to hold your arm
out against gravity all day long, you know you're gonna
need a big motor there, right, because that's a lot
(13:33):
of energy to support that. If you don't, if you
can design the robot in a minimal way where you're
not spending all of your time fighting energy or first off,
you're more energy efficient. Second you have a smaller motor
which will be less dangerous. And so the arm moves
in a way, it is designed in a way that
it's n have to fight gravity, and that just makes
(13:53):
it inherently safer. Right, And so when you're working around
older adults, people with disabilities, kids, pets, these things really
really matter. And so that's that's just the starting point
from the design. The intrinsic safety has to be there.
The physics of it has has to be right.
Speaker 1 (14:12):
So people with quadriplegia, people with extremely limited mobility are
an obvious early use case. What else, like, what else
are people using this for? What else do you have
in mind for early use case?
Speaker 2 (14:26):
Yeah, one of the most interesting use cases is in
the home around caregiving. It's a very challenging one. It
is further out, and so you know, there's a lot
of use cases nearer term which I think will be
in service, you know, back of retail logistics. Uh, And
we're seeing you know, data centers, a lot of stuff
(14:47):
like that that we are seeing interest in and people exploring.
Speaker 1 (14:50):
Like picking and pecking. I mean there are specialized robots
for those things.
Speaker 2 (14:54):
Right. Well, yeah, there's some places where there's not yet, right, So,
for example, micro fulfillment little fulfillment, you know, like even
like a you know, stocking a seven to eleven. There's
lots of things where a robot like this could play
a role.
Speaker 1 (15:09):
Options And just to bit clear, my understanding is that
your robot is mainly not autonomous, that it's mainly remote
controlled tteley operated. Is that right? And what does that
mean for that use case?
Speaker 2 (15:21):
Right? It can be autonomous. There are autonomous it's completely
open source as a whole ecosystem of open source software
people have built for it, including autonomy, and the generation
of robot we're working on now will really be autonomy powered, right,
so in terms of the sensing of the software. But
(15:42):
you know, back to that sort of really minimal approach. Yeah,
we're trying to say what can we do today that
has value? And for someone like Henry Evans, he controls
the robot, it has autonomy mixed in there, so he's
not controlling every motion necessarily, and so what we see
is as a steady progression from direct teleoperation to mixed
(16:06):
autonomy to full autonomy, but doing that really deployed, right,
So this is not something in a research lab. It's
actually in the product. And you know, we also see
a path from people with really severe mobility impairments like Henry.
They're interesting because they have a high motivation to use
(16:27):
the device, and it's been the time to control it, right,
They've got time and motivation. And then you can layer
in the economy over time, and as you do that,
now you can broaden the set of people who can
who benefit from it. I think from people like Henry.
The next step in terms of in home care is
(16:49):
a remote operation by a family member a care provider.
You know, for example, I'm in New Orleans right now,
my mother is back in California, and you know, I'd
love to hop on and check in, and you know,
of course you can zoom, but sometimes you need to
go into the other room, open the cabinet door, and
you know, find the medication or something like that. And
(17:12):
so again there's a human in the loop, but they're
not going to be controlling every motion. They're going to
be clicking on a map and saying go to the bedroom, open,
open the drawer, that type of thing. Yeah, And then
from there, of course, you can bring in more autonomy,
and suddenly you have in home autonomous care and eventually
in home robots that are doing the laundry and dishes
(17:34):
and everything that you know, people are excited about. But
for us, again, we're taking this very pragmatic approach about
how do we stair step, how do we kind of
climb our way up to that future?
Speaker 1 (17:47):
Stairstep is an interesting choice of words for a robot
on wheels. But you know, one of the things people
always talk about in the context of robots in the
last few years is getting the data to train the
model for autonomy, right, And the obvious analog is large
(18:08):
language models got so good because the Internet is full
of text, right, and there is no Internet of the
physical world. There is no giant data set that you
can just train a model on so that robots can
understand how physics works in the real world. Right, how
are you dealing with that?
Speaker 2 (18:28):
You know, I would say, in terms of AI physical
AI as it's called around robotics, we're taking, you know,
an approach that's really about taking things that are mature
and integrating them into our autonomy, versus what's called end
to end learning, where you know, you might collect a
huge data set, right, you might train the robot tele
(18:51):
operated to open the dishwasher a thousand times, collect that
data and train a deep learned model to do that task,
and you see very exciting results from that, and I
do think that will continue to progress, but in terms
of being ready to be deployed, it'll be take some time.
And the amount of data that's required beyond just a dishwasher, Well,
(19:15):
now it's a different dishwasher. Now it's a dishwashering different
lighting conditions. Oh there's something in front of it. That
complexity gets so large so quickly.
Speaker 1 (19:23):
And it doesn't generalize, like there are not models where
you can teach it one dishwasher and it can make
inferences about other dishwashers.
Speaker 2 (19:31):
There's inklings of that happening, right, but we're not there yet,
and resourcers are still really exploring what's possible. You know,
some are saying it's going to learn from YouTube, some
are saying it's going to be all on simulation. Some
are our hiring teams of operators to collect that data. Right.
Our view is, you know that that world is progressing
(19:52):
quickly and it'll sort itself out soon. You know, we
will be able to train the robot to open that dishwasher.
I think what's more interesting is how do you get
it to do you know the five or six important
tasks you care about and actually do it in a
robot that people feel is trusted in the home, it's
safe in the home, is affordable, you know, all the
(20:15):
things that actually to have a business.
Speaker 1 (20:17):
Yeah, so you're basically it's again the like the physical
intelligence is the bright shiny object that you are not
chasing so that you can build a robot that works exactly.
Speaker 2 (20:27):
And and you know, think about uh, you know, you know,
is Apple building a big AI team? Well, no, it's
partnering with the people who are doing the AI. And
I think you know, for us, we're more on that
side of the fence.
Speaker 1 (20:42):
Are some of those AI gains that you are sort
of piggybacking on coming from autonomous vehicles? I mean when
you talk about saying go to the other room and
have it go there, I think of that, That's where
my mind goes, right that those are the robots out
in the world right now driving around San Francisco.
Speaker 2 (20:58):
Yeah, I mean it's slightly different in terms of the
requirements the sensors, but it's certainly certainly related. I mean,
we've all benefited from he's wrapping gains and sensing and
compute and and you know even ll MS is they're
being used to help their robots plan and think and replan. Right.
(21:19):
So you know, when I was started in robotics twenty
twenty years ago, I couldn't imagine being able to say
to the robot, hey, go to the other room and
find the blue mug that's next to the microwave. Right,
Like the context of that was just like, how are
we going to solve that? Are we going to write
every rule into the robot and that's.
Speaker 1 (21:38):
Like commodified like those are That's that's the easy part
now at some.
Speaker 2 (21:43):
Level, yes, I mean certainly there's more to be done.
I'd say, you know, the real challenges today are going
to be around desk to manipulation and being able to
do things that are really in dynamic, you know, kind
of home type environments.
Speaker 1 (22:00):
So when you when you talk about you know, dexter
sort of fine motor skill type things, what are you
thinking of? What do you want the robot to be
able to do in the relatively near term that it
can't do now?
Speaker 2 (22:11):
I think just broaden the range of tasks. Right. So
one thing that's really exciting to this user group is feeding, right,
I think feeding. Feeding is one of the things where
right now, so your partner care partner has to come
and feed you, and that there's a number of reasons
why that's just they would prefer to do it themselves.
Of course.
Speaker 1 (22:30):
The basic it's a basic function of adult life is
you feed yourself.
Speaker 2 (22:36):
You feed yourself, and depending on someone changes that relationship
with that person, right, And and so feeding is one
that we've looked at a fair amount. It is challenging.
There are safety considerations, so we're careful about what utensils
and what type of food. But I think if it
can feed, if you can feed yourself, if you can
(22:58):
get yourself a glass of water, you can do basic
you know, hygiene maintenance things for yourself. I think that's
kind of the base line of regaining a sense of agency, right,
And what do you have.
Speaker 1 (23:08):
To figure out to make feeding really work that you
haven't figured out yet?
Speaker 2 (23:12):
Yeah, it's a mixture of we have. Food is kind
of messy, mushy, right, it's not like a rigid object.
Speaker 1 (23:20):
Well, and it's it's super heterogeneous, right, Food is many
different kinds of.
Speaker 2 (23:25):
Ways exactly, and so but there is you know, one
thing we've seen is you know, the caregiver is willing
to prepare the food, maybe put it in a special tray,
certain types of food that we agree it can work with.
So but of course you'd love that to get broader.
But you know, food in itself is challenging. Then there's
the safety challenges. But also, you know, a really interesting
(23:46):
part of this, which kind of relates to you know,
where AI is in general, is understanding the person, right,
are they ready for another bite? You know, are they
you know, what side would they like to be fed from.
That's all these little things that are about that human
interaction that make the product useful, exciting, acceptable in the home.
(24:11):
And so much of what's happening in ai AI right
now doesn't really get at the human side of the story.
And I think that's one thing where we would love
to see a advancements is really having better understanding of
these sort of more subtle cues and interactions that the
robot will need in order to be accepted.
Speaker 1 (24:35):
We'll be back in just a minute. So I'm curious
because you have been working on robots for more than
twenty years, right, I mean, so much has happened in software, yeah, right,
(24:58):
and even in other kinds of hardware, Like you were
working on robots before there was an iPhone, right, and
so so much has changed, and it feels like, at
least as a as a layperson, certainly the home like
robots are not around the way iPhones are around. And
even now large language models are around. And so has
less happened with robots than you thought.
Speaker 2 (25:20):
I would say less in terms of deployed robots, right,
Like the rumba is still one of the most successful robots,
but there aren't many other examples like it, right, and
so why is that?
Speaker 1 (25:32):
Yes?
Speaker 2 (25:32):
What?
Speaker 1 (25:33):
Like that's a great way to ask it, and that
I mean, you worked with Rodney Brooks, right, one of
the rumbook guys at MIT, and that's like the only
robot anyone can name, and there are lots of failed
companies that were rumba.
Speaker 2 (25:45):
Like right, yeah, yeah, and so yeah, So I did
my PhD with rod Brooks at MIT and at the
time he was spinning up a robot, inventing the rumba.
And around the time I finished, he had just sold
his million rumba and he's he inspired me to go
off to my own company.
Speaker 1 (26:03):
I mean, it must have seemed like they're here home
robots are here, like you would think at that moment,
it's like, oh, hey, here we go.
Speaker 2 (26:10):
Yeah yeah, well but things seemed very hard, right, so
it's like, okay, a vacuum cleaner that bounces around off walls,
like we can do that, right, Yeah. And I think
the insight from Rumba that I still carry is, you know,
they thought really carefully about the price, the complexity of
the device. They weren't trying to There was a competitor
(26:32):
call the Electrial Lucks at the time that was about
three thousand dollars had all sorts of cool features and
computer and sensing. But with the Rumbo, what they did
is they went to the mall and they asked people,
how much can you spend on something without having to
ask your significant other for permission?
Speaker 1 (26:49):
Oh?
Speaker 2 (26:50):
And the answer was about two hundred dollars.
Speaker 1 (26:53):
I was just talking to a friend who basically does
a B to B business. That's business a business, and
apparently that's a sales strategy in business as well. Like
the person who you're selling to they have some spending
authority that they can spend on their own. Yes, of course,
it's about a thing times greater. Right, It's like they
can spend one hundred thousand dollars or whatever without asking
(27:13):
somebody else, So you really want to price it like that,
otherwise it's much harder.
Speaker 2 (27:17):
So anyway, yeah, exactly, But you know, I think the
room to success in part was they made a minimal
product within that spending, you know, authority and kind of
made it all work. And I think the reason we
haven't seen a lot beyond that is it's hard. It's
been hard traditionally to find the things that match all
(27:37):
those constraints. That technology can be there, the use case
can be there, right clearly in home care, there's so
much need for help in the home, but can we
do it at a price? And is it technology there?
You know, five thousand dollars would be a lot of
money for most families.
Speaker 1 (27:54):
Yes, although for a caregiver, right like, if you narrow
it not to like doing the dishes.
Speaker 2 (27:59):
But yeah, I.
Speaker 1 (28:01):
Mean, you know, many people have been bankrupted by just
needing to hire a simple caregiver, right Like, people could
definitely figure out how to spend five thousand dollars on
a robot that could significantly replace a caregiver. One would think,
I guess that's your bet, that's your.
Speaker 2 (28:20):
That's partial partially our bet. Yeah, and you know, I
think it's not replacing, it's it's augmenting, you know, I
think think what makes sense, yes, because we're not going
to do everything, and nor should we, you know, but
there's just aren't enough people ours caregiving hours in the
in the world today to take care of everyone who
needs care. That's just the truth of it.
Speaker 1 (28:39):
Yeah, zooming out from from Hello robot, from your company
to the field more generally, like what's the landscape look
to you? How does the landscape look to you right now?
Speaker 2 (28:48):
Yeah, it's fascinating there. There's a lot of first off,
a lot of capital flowing into it. Into two areas really,
one is robot foundation models. You know, this is sort
of deep learning the alms of robots. And then the
other would be humanoids, right, and sometimes there's both, right,
but there's a lot of excitement around that, and would
(29:09):
say generally they both have a little ways to go
before they get deployed and really applied.
Speaker 1 (29:15):
I mean the foundation model one, you know, an ll
M for robots sounds good to me, It seems interesting
to me. Tell me about what's happening there.
Speaker 2 (29:24):
Yeah, Well, it's first off, it's remarkable what you see
the robots doing with these these models, right, so peeling
an egg or tying shoelaces or folding laundry. And they're
getting better, right, They're they're getting they're having long I think.
Speaker 1 (29:41):
The peeling an egg was teleoperated.
Speaker 2 (29:43):
Okay, fair enough, but they are doing things like that
that you think, wow, I just could not imagine this.
And and they're getting better, and that they're having longer
memory so they can do stuff over a longer period
of time. They're getting better at recovering from from issues.
But also they haven't yet learned to generalize. Right, So
(30:03):
if you think about what they're learning, that's such as
we were saying, such richness to diversity in the environments
they meant to operate, and they're learning on one kind
of little island and the whole sea of data that
they need access to to really be generally useful. And
so that's kind of a real challenge that that field
(30:23):
is trying to work through, is how do you get
to that generality and that robustness to that diversity. And
it's very much an unsolved problem. I think they're making progress,
but it also means to me that this is not
something that we're going to be deploying into a robot
product anytime soon.
Speaker 1 (30:42):
You know, it's interesting because it seems like one of
the frankly big surprises about large language models is how
good they are at generalizing, right, how they often are
much better than specialist models.
Speaker 2 (30:57):
Right.
Speaker 1 (30:57):
Just take a frontier model and throw anything at it
and it will be really good. Not quite, but kind of.
And so it's interesting to hear that that's not at
all the case with robot foundation models. Why do you
pose that.
Speaker 2 (31:09):
Is, Well, you know, I think one thing is just
the amount of data that's available, Right, So if you
think of a ll M, one simple way to think
of it, it's going to predict the next word that
you say, and it has the whole corpus of the
Internet to make that prediction. Right, for a robot to
predict the next thing that's going to happen when it's
(31:30):
peeling the egg, there's not a lot of data out
there for it to understand that. And some people are saying, well,
I can do that in simulation, but very hard to
imagine simulating a lot of that kind of complexity that's
actually in the world, just.
Speaker 1 (31:47):
Because like the world is so diverse. What do you
even simulate or how.
Speaker 2 (31:50):
Do you even simulate an egg being peeled? And so,
you know, I think it's a challenge because robots are
in the physical world.
Speaker 1 (32:00):
Is there something as I understand it, the basic model,
the transformer model that powers all of the frontier elements
today was developed based on the nature of language itself, right, Like,
is there a thought that that's maybe just not the
best fundamental model for understanding the physical world for training robots.
(32:23):
Might it be the case that some very different approach
ends up working much better.
Speaker 2 (32:28):
Yeah, I know less about that. I do know there
is a large interest in world models today that can
kind of infer the physics of the world. And I
think one thing that that is beneficial is there is
so much structure in the world, both from physics.
Speaker 1 (32:44):
Like laws of physics, laws of physics, yep, but also.
Speaker 2 (32:47):
In the way that we design our homes, our appliances
are you know, there is a lot of structure to
be gained and learned, and so you can imagine these
models become very good at predicting what your home is
going to look like because it's seen a lot of
homes and they don't vary that much fundamentally, so so
that there is there is an enough structure to learn
(33:10):
from what that model and how that transpires. I don't know,
but people are exploring that.
Speaker 1 (33:16):
I mean, one of the one of the obvious challenges
is we're just making the robots come live in our world, right,
And I could imagine, for a case like someone with quadriplegia,
it might I could imagine a universe where it's like, well,
if you set up your room this way, not you
have to remodel your house, but you have to do
some set of things, then our robot can do a lot.
(33:36):
And I can imagine that being worthwhile in many settings.
Is that something you are doing or thinking about? Sort
of a compromised solution like that?
Speaker 2 (33:44):
Absolutely? And you know, as an example, a Stretch three
as it is today, has a hard time opening some refrigerators, right,
It's just the magnetic force is too high. But you
can attach a little you know, grabbing thing that breaks
the seal and makes it easy to open it. Right,
And so it's a three dollars apart, you stick it
onto your fridge. Now Stretch can open their fridge for
(34:07):
you and for for these users, you know, those accommodations
make a lot of sense, and uh, you know, I
do think you know, maybe over time you'll have special
plates at the robot nose it cannot you know, grab
and not break, and you know, there will be accommodations
that people make for it. But you know, for for now,
(34:29):
we are using and counting on those types of accommodations.
Speaker 1 (34:33):
So I want to do two kind of forward looking
things first, the sad one and the happy one. Basically okay,
first is how might it not work out? And then
the second is what does it look like if it
does work out for the low robots, So like what
might go wrong for for your company in a you know,
significant way, right.
Speaker 2 (34:57):
Well, there's there's always the risk of just the business
doesn't work out. I think, I think from from the
technology and product side, it's really about acceptance, right, So
I think implicit to bringing this technology and your life
into your home, there has to be trust, right, and
people have to trust and want this technology around them,
and so I think, you know, there there's always a
(35:19):
risk that people like you know, what we're going to
we're going to find different ways to addrust the caregiving problem,
you know, which you know, I should. There's lots of
going to be lots of solutions to that, but robots
may may not turn out to be part of that story.
We're bullish that it will, but there it's a real
kind of complex dynamic between the people and the technology.
(35:41):
So that's that's sort of the the bare case. The
bowl case would be, you know, ten years out. Call
it a five thousand dollars device that's in your home
that can provide caregiving support in the times when there's
not care present. So basically it's just kind of facilitating
quality of life at home, agency at home as you
(36:02):
want to kind of age independently without having so when
you're when your grandkids come over, you're not they're not
doing chores for you. You're interacting, right because of the
robots helping fill that gap.
Speaker 1 (36:17):
We'll be back in a minute with the lightning round.
Let's finish with the lightning round.
Speaker 2 (36:35):
Okay, do you.
Speaker 1 (36:37):
Still make art?
Speaker 2 (36:38):
Oh? I not for a while, But that is how
I got into robotics, was doing robotic sculpture, and then
that led one thing led to another. But yeah, really
fascinated by well, love art but also love you know,
kind of robotics and how it relates to art.
Speaker 1 (36:58):
What's the least practical robot you've ever built?
Speaker 2 (37:02):
Wow? Uh? So let's see in the in the nineteen
nineties in San Francisco, we had a robotic art performance
troupe and it was a theatrical multimedia performance of robots
on stage acting performing. You know, this was a kind
of robotic gallery of kind of characters from the from
(37:25):
the local neighborhood. So these were a street musician and
you know, someone who was got of causing trouble and
so they were kind of robotic denizens of the Tenderloin
of nineteen nineties, San Francisco put onto stage.
Speaker 1 (37:43):
What was the first robot you ever built?
Speaker 2 (37:46):
It was actually the Robotic Street Preacher. So it was
a little robot that would go around and big spinning
head of a speaker and it would be on stage
and perform and we would sit there and tell operate it.
And it had, you know, no real autonomy. This is
you know, everything with the hands salted, handbuilt and their
(38:06):
computers were super tiny, are super a small capability, And
we put that on stage and performed it.
Speaker 1 (38:15):
Did you ever try and build robots when you were
a kid?
Speaker 2 (38:19):
No, you know, Strangely, I never had a fascination with
robusts as a kid. I came out of grew up
in Washington State on a farm, and uh, you know,
I think my interest in robots came from having to
do farm work and learning about automation.
Speaker 1 (38:36):
Like Henry Ford. You know, Henry Ford was the same
story about a farm and thought that was too much
labor exactly, and I just thought, you know, there's got
to be a better way than this.
Speaker 2 (38:45):
My my mother's family, which is Mexican, had a history
and kind of uh kind of working in the fields
and kind of migrant labor, and so kind of the
relationship to labor and work kind of inspired my interest
in automation and then roproducts.
Speaker 1 (39:02):
Did you ever try and build a machine to do
your chores for you?
Speaker 2 (39:05):
I absolutely did. It never worked, but it was fun,
you know.
Speaker 1 (39:10):
I mean, I guess that's still what you're working on
at some level.
Speaker 2 (39:13):
Exactly.
Speaker 1 (39:15):
What's one thing you've learned working with Frank Gary?
Speaker 2 (39:17):
Oh? Right, So we had a project looking at this
conceptual but a large robotic kind of botanical garden and
this sole thing in Singapore, so that we were looking
at developing that with him and Uh, you know, I'd say, actually,
the thing that was most interesting is just he really
(39:41):
put a lot of trust into the young people in
the studio and really was generous in that way. And
you know, it's been an inspiration for me.
Speaker 1 (39:50):
That's a lovely surprise one. This is I guess some
bias that I have. I imagine a famous architect as
not being like that.
Speaker 2 (39:58):
So to hear that so lovely, very very generous in
a way that can be inspiring to to the to
people in their career. So I would try to carry
that forward in my own work.
Speaker 1 (40:09):
Uh, thank you for your time.
Speaker 2 (40:11):
Yeah, it's been a pleasure. Thank you.
Speaker 1 (40:20):
Aaron Edzinger is the co founder and CEO of Hello Robot.
Today's show was produced by Gabriel Hunter Chang and edited
by Lydia Jane Kott. Our engineer this week was Hansdale Sheet.
We're always looking for ideas for who to talk to
and what to cover on the show. You can email
us at problem at Pushkin dot fm. You can find
(40:42):
me on x at Jacob Goldstein. You can find me
on LinkedIn. I'm Jacob Goldstein. Thank you very much for
listening to the show, and we'll be back next week
with another episode
Popular Podcasts
Hey Jonas!
Hey Jonas! The official Jonas Brothers podcast. Hosted by Kevin, Joe, and Nick Jonas. It’s the Jonas Brothers you know... musicians, actors, and well, yes, brothers. Now, they’re sharing another side of themselves in the playful, intimate, and irreverent way only they can. Spend time with the Jonas Brothers here and stay a little bit longer for deep conversations like never before.
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by Audiochuck Media Company.
Betrayal Weekly
Betrayal Weekly is back for a new season. Every Thursday, Betrayal Weekly shares first-hand accounts of broken trust, shocking deceptions, and the trail of destruction they leave behind. Hosted by Andrea Gunning, this weekly ongoing series digs into real-life stories of betrayal and the aftermath. From stories of double lives to dark discoveries, these are cautionary tales and accounts of resilience against all odds. From the producers of the critically acclaimed Betrayal series, Betrayal Weekly drops new episodes every Thursday. If you would like to share your story, you can reach out to the Betrayal Team by emailing them at betrayalpod@gmail.com and follow us on Instagram at @betrayalpod and @glasspodcasts. Please join our Substack for additional exclusive content, curated book recommendations, and community discussions. Sign up FREE by clicking this link Beyond Betrayal Substack. Join our community dedicated to truth, resilience, and healing. Your voice matters! Be a part of our Betrayal journey on Substack.