Kind: captions
Language: en
and so our goal was a natural looking
gate it was real it was surprisingly
hard to get that to work
um and we but we did build an early
machine
uh we called it pet man prototype it was
the Prototype before the Pac-Man robot
and it had a really nice looking
gate where you know it would stick the
leg out it would do heel strike first
before it rolled onto the toe so you
didn't land with a flat foot you
extended your leg a little bit
um but even then it was hard to get the
robot to walk where when you're walking
that it fully extended its leg
and getting that all to work well
took such a long time in fact I I
probably didn't really see the nice
natural walking that I expected out of
our humanoids until maybe last year
and the team was developing on our newer
generation of Atlas you know some new
techniques
for developing a walking control
algorithm and they got that natural
looking motion as sort of a byproduct of
a just a different process that we're
applying to developing the control
so that probably took 15 years 10 to 15
years to sort of get that from from you
know
the Petman prototype was probably in
2008 and what was it 2022 last year that
I think I saw a good walking on Atlas
the following is a conversation with
Robert plater CEO of Boston Dynamics a
legendary robotics company that over 30
years has created some of the most
elegant dexterous and simply amazing
robots ever built including the humanoid
robot Atlas and the robot dog spot
one or both of whom you've probably seen
on the internet either dancing doing
backflips opening doors or uh throwing
around heavy objects
Robert has led both the development of
Boston Dynamics humanoid robots and
their physics-based simulation software
he has been with the company from the
very beginning including its roots at
MIT where he received his PhD in
Aeronautical Engineering this was in
1994 at the legendary MIT leg lab he
wrote his PhD thesis on robot gymnastics
as part of which he programmed a bipedal
robot to do the world's first 3D robotic
somersault
Robert is a great engineer robot
assistant leader and Boston Dynamics to
me as a roboticist is a truly inspiring
company this conversation was a big
honor and pleasure and I hope to do a
lot of great work with these robots in
the years to come
this is the Lex Freedom podcast to
support it please check out our sponsors
in the description and now dear friends
here's Robert plater
when did you first fall in love with
robotics
let's start with love and robots well
love is is relevant because I think the
the fascination the Deep Fascination is
really about movement
and uh
I was visiting MIT looking for a place
to get a PhD and I wanted to do some
laboratory work and one of my professors
at in the Aero Department said go see
this guy mark raber down in the basement
of the AI lab
and so I walked down there and saw him
he showed me his robots
and he showed me this robot doing a
somersault
and I just immediately went whoa you
know yeah robots can do that and because
of my own interest in in gymnastics
there was like this immediate connection
and um you know I was interested in I
was in an arrow Astro degree because you
know flight and movement was all so
fascinating to me and then it turned out
that you know robotics had this big
challenge how do you how do you balance
uh how do you how do you build a legged
robot that can really get around
and that just that was a Fascination and
it still exists today you're still
working on perfecting Motion in robots
what about the elegance and the beauty
of the movement itself is is there
something
maybe grounded in your appreciation of
uh movement from your gymnastics days
did you
was there something you just
fundamentally appreciate about the
elegance and beauty of movement you know
we had this concept in in gymnastics of
um
letting your body do what it wanted to
do when you get really good at
gymnastics
part of what you're doing is putting
your your body into a position where the
physics and the body's inertia and
momentum will kind of push you in the
right direction in a very natural and
organic way
and the thing that Mark was doing you
know in the basement of that laboratory
was trying to figure out how to build
machines to take advantage of those
ideas how do you build something so that
the physics of the machine just kind of
inherently wants to do what it wants to
do and he was building these springy
pogo stick type you know his first cut
at Lego Locomotion was a pogo stick
where it's bouncing and there's a spring
Mass
a system that's oscillating has its own
sort of natural frequency there and sort
of figuring out how to augment those
natural physics
with also intent how do you then control
that but not overpower it it's that
coordination that I think creates real
potential we could call it Beauty yeah
you could call it I don't know Synergy
that people have different words for it
but I think that that was inherent from
the beginning that was clear to me that
that's part of what Mark was trying to
do he asked me to do that in my research
work so
um you know that's where I got going so
part of the thing that I think I'm
calling elegance and Beauty in this case
which was there even with the pogo stick
is maybe the efficiency so letting the
body do what it wants to do
trying to discover the efficient
movement it's definitely more efficient
it also becomes easier to control in its
own way because the physics are solving
some of the problem itself it's not like
you have to do all this calculation and
overpower the physics the physics
naturally inherently want to do the
right thing
there can even be you know a feedback
mechanisms stabilizing mechanisms
that occur simply by virtue of the
physics of the body and it's you know
not all
not all in the computer or not even all
in your mind as a person and I there's
something interesting in that melding
you were with Mark for many many many
years but you were there in this kind of
legendary space
of leg lab and a my team in the basement
all great things happen in the basement
is there some memories uh is there some
money from that time
that you have because it's so it's such
Cutting Edge work
in in in robotics and artificial
intelligence
the memories the distinctive lessons I
would say I learned in that in that time
period
and um and that I think Mark was a great
teacher of
was it's okay to pursue your interest
your curiosity do something because you
love it
um you'll do it a lot better if you love
it
um
that that is a lasting lesson that I
think uh we apply at the company still
um and really is a core value so the
interesting thing is I got to um
uh with people like Ross Cedric and
um and others like the students that
work at those robotics labs are like
some of the happiest people I've ever
met
I don't know what that is I mean a lot
of PhD students a lot of them are kind
of broken by the wear and tear of the
process uh but roboticists are while
they work extremely hard and work a long
hours
there's a
there's a happiness there the only other
group of people I met like that are
people that Skydive a lot like for some
reason there's a deep fulfilling
happiness maybe from like a long period
of struggle to get a thing to work and
it works and there's a magic to it I
don't know exactly because it's so
fundamentally Hands-On and you're
bringing a thing to life I don't know
what it is but they're happy
we see you know our our attrition at the
company is really low people come and
they love the pursuit
and I think part of that is that there's
perhaps an external connection to it
it's a little bit easier to connect when
you have a robot that's moving around in
the world and part of your goal is to
make it move around in the world
you can identify with that and and this
is on a this is one of the unique things
about the kinds of robots we're building
is this physical interaction
lets you perhaps identify with it so I
think that is a source of happiness I
don't think it's Unique to robotics I
think anybody also who is just pursuing
something they love
it's easier to work hard at it and be
good at it and
um
not everybody gets to find that I I do
feel lucky in that way and I think we're
lucky as an organization that we've been
able to build a business around this and
that keeps people engaged
so if it's all right let's link on mark
for a little bit longer Mark raybert so
he's a legend
uh he's a legendary engineer Roboto says
what what have you learned about life
about Robotics and Mark through all the
many years you worked with him I think
the most important lesson which was you
know have the courage of your
convictions and and do what you think is
interesting
um
be willing to try to find big big
problems to go after and at the time you
know like at Locomotion
um especially in a dynamic machine
nobody had solved it and that felt like
a
multi-decade problem to go after
and so you know have the courage to go
after that because you're interested
don't worry if it's going to make money
you know that that's been a theme so
that that's really uh probably the most
uh important lesson I think that uh I
got from Mark how crazy is the effort of
doing legged
robotics at that time especially
you know Mark got some stuff to work uh
starting from the simple ideas oh so
maybe the other another important idea
that has really become a value of the
company is try to simplify a thing to
the Core Essence
and and while you know Mark was showing
videos of animals running across the
Savannah or uh uh climbing mountains
what he started with was a pogo stick
because he was trying to reduce the
problem to something that was manageable
and and getting the pogo stick to
balance had in it
the fundamental problems that if we
solve those you could eventually
extrapolate to something that galloped
like a horse
and so look for those simplifying
principles
um how tough is the job of simplifying a
robot so I I'd say in the early days the
the thing that made Boston
the researchers at Boston Dynamics
special
is that we we worked on under figuring
out what that that Central principle was
and then building software or machines
around that principle and that was not
easy in the early days and and it it
took
um real expertise in understanding the
Dynamics of motion and feedback control
principles and how to build and with
computers at the time how to build a
feedback control algorithm that was
simple enough that it could run in real
time at a thousand Hertz
and actually get that machine to work
um and that was not something everybody
was doing you know at that time
now the world's changing now and I I I
think the approach is to controlling
robots are going to change
um but uh and they're going to become
more broadly yet
um available
um but at the time there weren't many
groups who could really sort of work at
that principled level
with both the software and
and make the hardware work
and I'll and I'll say one other thing
about your sort of talking about what
are the special things the other thing
was it's okay it's good to break stuff
you know
um you know use the robots break them
repair them
um you know fix and repeat test fix and
repeat and that and that's also a core
principle that has become part of the
company
and it lets you be Fearless in your work
too often if you are working with a very
expensive robot maybe one that you
bought from somebody else or that you
don't know how to fix then you treat it
with kit gloves and you can't actually
make progress you have to be able to
break something and so I think that's a
been a a principle as well so just the
link on that psychologically how do you
deal with that because I remember I had
uh
uh I built a RC car
with that some
uh it had some custom stuff like compute
on it and all that kind of stuff cameras
and uh because I didn't sleep much the
code I wrote has an issue where it
didn't stop the car and then the car got
confused and at full speed at like 20 25
miles an hour slammed into a wall
and I just remember sitting there alone
in the deep sadness
um
sort of
full of regret I think almost anger
um
uh but also like sadness because you
think about well these robots especially
for autonomous vehicles like like you
should be taking safety very seriously
even in these kinds of things but just
no good feelings and made me more afraid
probably to do this kind of experiments
in the future perhaps the right way to
have seen that is positively
like it's it's too it depends if you
could have built that car or or just
gotten another one right that would have
been the approach
um I remember
um
when I got to grad school
um you know I got some training about uh
operating a lathe and a mill up in the
machine shop and I could start to make
my own parts and I remember breaking
some piece of equipment in the lab and
then realizing
because I maybe this was a unique part
and I couldn't go buy it and I realized
oh I can just go make it
that was an enabling feeling yeah then
you're not afraid yeah it might take
time it might take more work than you
thought it was going to be required to
get this thing done
but you can just go make it and that's
freeing in a way that nothing else is
you mentioned uh the the feedback
control the Dynamics sorry for the
Romantic question but is in the early
days and even now is the Dynamics
probably more appropriate for the early
days is it more art or science
there's a lot of science around it
and and trying to develop you know
scientific principles
that let you extrapolate from like one
legged machine to another
you know develop a core set of
principles like like a spring Mass
bouncing system and then figure out how
to apply that from a one-legged machine
to a two or a four-legged machine those
principles are really important and and
we're definitely a core part of our work
there's also
you know when we started to pursue
humanoid robots
um there was so much complexity in that
machine
that
you know one of the benefits of the
humanoid form is you have some intuition
about how it should look while it's
moving
and that's a little bit of an art I
think and now I'd say or maybe it's just
tapping into a knowledge that you have
deep in your body and then trying to
express that in the machine but that's
an intuition that's a little bit more on
the art side maybe it it predates your
knowledge you know before you have the
knowledge of how to control it you try
to work through the Art Channel and
humanoids sort of make that available to
you if it had been a different shape
maybe we wouldn't have had the same
intuition about it yeah so you're
knowledge about moving through the world
is not made explicit to you
so you just that's why it's art and it
might yeah it might be hard to actually
articulate exactly you know there's
something about
um and being a competitive uh athlete
there's something about
seeing a movement you know a coach one
of their greatest strengths a coach has
is being able to see you know some
little change in what the athlete is
doing and then being able to articulate
that to the athlete you know and then
maybe even trying to say and you should
try to feel this
um so there's something just in scene
and again you know sometimes it's hard
to articulate what it is you're seeing
but there's a
receiving the motion at a rate that is
again sometimes hard to put into words
yeah I
Wonder
how it is possible to achieve sort of
truly elegant movement you have a movie
like ex machina I'm not sure if you've
seen it but the main actress in that who
plays the AI robot I think is a
ballerina I mean just a natural
elegance and the I don't know eloquence
of movement
it's it's it looks efficient and easy
and just it looks right
it looks it looks right is sort of the
key yeah and then you you look at uh
especially early robots I mean they
they're so cautious in in the way they
move that
it's not it's not the caution that looks
wrong it's it's something about the
movement that looks wrong that feels
like it's very inefficient unnecessarily
so and it's hard to put that into words
exactly we think that and part of the
reason why people are attracted to the
machines we build
is because the inherent dynamics of
movement are are closer to right
um because we we try to use you know
walking Gates or we build a machine
around this gate where you're trying to
work with the Dynamics of the machine
instead of
to stop them you know some of the early
walking machines
you know you're essentially you're
really trying hard to not let them fall
over and so you're always stopping the
Tipping motion you know
and sort of the insight
of dynamic stability and a lighted
machine is to go with it you know let
the Tipping happen you know let yourself
fall but then catch her catch yourself
with that next foot and there's
something about getting those physics to
be expressed in the machine
that people interpret as
lifelike or or elegant or just natural
looking and so I think if you get the
physics right
it also ends up being more efficient
likely
there's a benefit that it probably ends
up being
more stable in the long run you know it
could it could walk stably over a wider
uh rain range of conditions
and it's uh and it's more beautiful and
attractive at the same time so how hard
is it to get the humanoid robot Atlas
to do some of the things that's recently
been doing let's forget the flips and
all of that let's just look at the
running
maybe you can correct me but there's
something about running I mean that's
not careful at all that's you're falling
forward
you're jumping forward and they're
falling so how hard is it to get that
right our first humanoid we needed to
deliver natural looking walking you know
we took a contract uh from the army they
wanted a robot that could walk naturally
they wanted to put a suit on the robot
and be able to test it in a gas
environment and so they wanted that the
motion to be natural
and so our goal was a natural looking
gate it was real it was surprisingly
hard to get that to work
and we but we did build an early machine
we called it pet man prototype it was
the Prototype before the Pac-Man robot
and it had a really nice looking
gate where you know it would stick the
leg out it would do heel strike first
before it rolled onto the toe so you
didn't land with a flat foot you
extended your leg a little bit
but even then it was hard to get the
robot to walk where when you're walking
that it fully extended its leg and
essentially landed on an extended leg
and if you watch closely how you walk
you probably land on an extended leg but
then you immediately flex your knee as
you start to make that contact
and getting that all to work well
took such a long time in fact I I
probably didn't really see the nice
natural walking that I expected out of
our humanoids until maybe last year
and the team was developing on our newer
generation of Atlas you know some new
techniques
um uh for developing a walking control
algorithm and they got that natural
looking motion as sort of a byproduct of
a just a different process they were
applying to developing the control
so that probably took 15 years 10 to 15
years to sort of get that from from you
know the Petman prototype was probably
in 2008 and what was it 2022 last year
that I think I saw a good walking on
Atlas if you could just like Linger on
it what are some challenges of getting
good walking so is it uh
is this is this partially like a
hardware like actuator problem is it the
control is it the artistic element of
just observing the whole system
operating in different conditions
together I mean is there some kind of
interesting
quirks or challenges you can speak to
like the heel strike yeah so one of the
things that makes the like this straight
leg uh a challenge is you're sort of up
against a singularity a mathematical
single Singularity where you know when
your leg is fully extended it can't go
further the other direction right
there's only you can only move in One
Direction and that makes all of the
calculations around how to produce
twerks at that joint or positions makes
it more complicated and so having all
the mathematics so it can deal with
these singular configurations is one of
many challenges uh that we face and I'd
say in in the
you know in those earlier days again we
were working with these really
simplified models
so we're trying to boil all the physics
of the complex human body into a simpler
subsystem that we can more easily
describe in mathematics and sometimes
those simpler subsystems don't have all
of that complexity of the straight leg
built into them and so what's happened
more recently is we're able to apply
techniques that let us take the full
physics of the robot into account and
and deal with some of those strange
situations like this like the straight
leg so is there a fundamental challenge
here that it's uh maybe you can correct
me but is it under actuated are you
falling under actuated is the right word
right you can't you can't uh push the
robot in any direction you want to right
and so that that is one of the hard
problems of of uh like at Locomotion and
you have to do that for natural movement
it's not necessarily required for
natural movement it's just required
you know we don't have you know a
gravity force that you can hook yourself
onto to apply uh an external force in
the direction you want at all times
right the only the only external forces
are being mediated through your feet and
how they get mediated depend on how you
place your feet and uh you know you
can't just uh you know God's hand can't
reach down and give and push in any
direction you want you know so is there
uh is there some extra challenge to the
fact that Alice is such a big robot
there is the humanoid form is um
attractive in many ways but it's also a
challenge in many ways
um
you have this big upper body that has a
lot of mass and inertia
um and throwing that inertia around
increases the complexity of maintaining
balance and as soon as you pick up
something heavy in your arms you've made
that problem even harder
and so uh in the early work in the leg
lab and in the early days at the company
and we were pursuing these quadruped
robots which had a a kind of built-in
simplification you had this big rigid
body and then really light legs so when
you swing the legs
the leg motion didn't impact the body
motion very much
all the mass and inertia was in the body
but when you have the humanoid that
doesn't work you have big heavy legs you
swing the legs it affects everything
else
and so
dealing with all of that interaction
does make the humanoid a much more
complicated platform
and I also saw that at least recently
you've been doing more explicit modeling
of the stuff you pick up yeah which is
very real
um really interesting so you have to
what model the shape
the weight distribution
[Music]
I don't know what like you have to under
like include that as part of the
modeling as part of the planning because
okay so for people who don't know
uh so Atlas at least in like a recent
video like throws a heavy bag throws a
bunch of stuff
so what what's involved in uh picking up
a thing a heavy thing
uh and when that thing is a bunch of
different non-standard things I think it
also picked up like a barbell
and to be able to throw in some cases
what are some interesting challenges
there
so we were definitely trying to show
that the robot and the techniques were
applying to the robe uh to Atlas let us
deal with heavy things in the world
because if the robot's going to be
useful it's actually got to move stuff
around yeah and that and that needs to
be significant stuff that's an
appreciable portion of the the body
weight of the robot
and we also think this differentiates us
from the other humanoid robot activities
that you're seeing out there mostly
they're not picking stuff up yet
and not heavy stuff anyway
um but just like you or me you know you
need to anticipate that moment you know
you're reaching out to pick something up
and as soon as you pick it up your
center of mass is going to shift
and if you're gonna you know turn in a
circle you have to take that inertia
into account and if you're gonna throw a
thing you know you've got all of that
has to be sort of included in in the
model of what you're trying to do so the
robot needs to have some idea or
expectation of what that weight is and
then and sort of predict you know think
a couple of seconds ahead how do I
manage my now my my body plus this big
heavy thing together to get and and
still maintain balance right and so um I
I uh that's a big change for us and I
think the tools we've built are really
allowing that to happen
um quickly now some of those motions
that you saw in that most recent video
we were able to create in a matter of
days it used to be it took six months to
do anything new you know on your robot
and and now we're starting to develop
the tools that let us do that in a
matter of days and so we think that's
really exciting it means that the
ability to create new behaviors for the
robot is going to be um
a quicker process so being able to
explicitly model
new things that it might need to pick up
new type of thing and you know to some
degree you don't you don't want to have
to pay too much attention to each
specific thing right
um there's sort of a generalization here
yeah
um obviously when you grab a thing you
have to conform your your hand your end
effector to the surface of that shape
but once it's in your hands it's
probably just the mass and inertia that
matter and the the shape may not be as
important yeah and so you know for some
in some ways you want to pay attention
to that detailed shape and in others you
want to generalize it and say uh well
all I really care about is the center of
mass of this thing especially if I'm
going to throw it up on that scaffolding
and it's easier if the body is rigid
what if it's there's some doesn't it
throw like a sandbag type thing that
tool bag you know you've had loose had
loose stuff in it yeah so it it managed
that there are harder things that we
haven't done yet you know we could have
had a big jointed thing or I don't know
a bunch of loose wire or rope what about
carrying another robot how about that
yeah we haven't we haven't done that yet
I guess we did a little bit of uh we did
a a little skit around Christmas where
we had two spots holding up another spot
that was trying to put you know a bow on
a tree so I guess we're doing that in a
small way
okay that's pretty good uh let me ask
the all-important question uh do you
know how much Atlas can curl goodbye
have you
I mean you know this for us humans
that's really one of the most
fundamental questions you can ask
another human being a bench it probably
can't curl as much as we can yet but a
metric that I think is interesting is um
you know another way of looking at that
strength
is you know the box jump so if how high
of a box can you jump onto question and
uh Alice I don't know the exact height
it was probably a meter high or
something like that it was a pretty
pretty tall jump that Atlas was able to
manage when we last tried to do this and
and I have video of my chief technical
officer
doing the same jump and he really
struggled you know the human but the
human getting all the way on top of this
box but then you know Atlas was able to
do it
um we're now thinking about the next
generation of Atlas and we're probably
going to be in the realm of a person
can't do it you know with this with the
Next Generation you know the robots the
actuators are going to get stronger
where there really is the case that at
least some of these joints some of these
motions will be stronger and to
understand how high it can jump you
probably had to do quite a bit of
testing oh yeah and there's lots of
videos of it trying and failing and
that's you know that's all you know we
don't always release those those videos
but they're a lot of fun to look at
uh so we'll talk a little bit about that
uh but if can you talk to the jumping
because you talked about the walking it
took a long time many many years to get
the walking to be natural but there's
also really natural looking
uh robust resilient jumping how hard is
it to do the jumping
well again this stuff has really evolved
rapidly in the last few years you know
the first time we did a somersault
um you know there's a lot of kind of
manual iteration
what is the trajectory you know how hard
do you throw it in fact in these early
days uh I actually would when I'd see
early experiments that the team was
doing I might make suggestions about how
to change the technique again kind of
borrowing from my own intuition about
how backflips work
um
but frankly they don't need that anymore
so in the early days you had to iterate
kind of in almost a manual way trying to
change these trajectories of the arms or
the legs to try to get you know a
successful backflip to happen
but more recently we're running
these model predictive uh control
techniques where we're able to the robot
essentially can think in advance for the
next second or two about how its motion
is going to transpire and you can you
know solve for optimal trajectories to
get from A to B so this is happening in
a much more natural way and we're really
seeing an acceleration happen in the
development of these behaviors again
partly due to these
optimization techniques uh sometimes
learning techniques
so it's there's it's hard in that
there's a lot of mathematics in behind
it but we're figuring that out so you
can do model predictive control for
uh I mean I don't even understand what
that looks like when the entire robot is
in the air flying and doing a back yeah
I mean but but that's the cool part
right so you know yeah you know the
physics we we can calculate physics
pretty well using you know Newton's laws
about how it's going to evolve over time
and the road you know this this the sick
trick which was a front somersault with
a half twist is a good example right
you saw the robot on various versions of
that trick I've seen it land in
different configurations and it still
manages to stabilize itself and so you
know what this model predictive control
means is again the in real time the
robot is projecting ahead you know a
second into the future and sort of
exploring options and if I if I move my
arm a little bit more this way how is
that going to affect the outcome and so
it can do these calculations many of
them you know uh and and basically solve
for you know given where I am now maybe
I took off a little bit screwy from how
I had planned I can adjust so you're
adjusting in there just on the fly so
the the model predictive control lets
you adjust on the Fly
and of course I think this is what you
know people adapt as well we when when
we do it even a gymnastics trick we try
to set it up so it's as close to the
same every time but we figured out how
to do some adjustment on the Fly and now
we're starting to figure out that the
robots can do this adjustment on the fly
as well using these techniques in the
air
and so I mean it just feels from a
robotics perspective just surreal well
that's sort of the you talked about
under actuated right so when you're when
you're in the air there's something
there's some things you can't change
right you can't change the momentum
while it's in the air because you can't
apply an external force or Torque and so
the momentum isn't going to change so
how do you work within the constraint of
that fixed momentum to still get from A
to B where you want to be that's really
unfortunate
you're in the air I mean you become a
drone for a brief moment of time no
you're not even a drone because you
can't can't ever you can't hover you're
gonna you're gonna impact soon be ready
yeah are you considered like a hover
type thing or no no it's too much weight
I mean it's just it's just incredible uh
and just even to have the guts to try
backflip with such a large body
that's wild but like uh we definitely
broke a few robots trying but that but
that's where the build it break it fix
it you know uh strategy comes in you
gotta be willing to break and what ends
up happening is you end up by breaking
the robot repeatedly you find the weak
points and then you end up redesigning
it so it doesn't break so easily next
time you know through the breaking
process you learn a lot like a lot of
lessons and you keep improving not just
how to make the backflip work but
everything and how to build the machine
better yeah yeah
I mean is there something about just the
guts that
come up with an idea of saying you know
what let's try to make it do a backflip
well I think the courage to do a
backflip in the first place and and to
not worry too much about the ridicule of
somebody saying why the heck are you
doing backflips with robots because a
lot of people have asked that you know
why why why are you doing this why go to
the moon in this decade and do the other
things JFK
[Laughter]
it's not because it's easy because it's
hard yeah exactly
don't ask questions okay so the uh the
jumping I mean it's just there's a lot
of incredible stuff if we can just
rewind a little bit to uh the DARPA
robotics challenge in 2015 I think which
was for people who are familiar with the
DARPA challenges
it uh was first with autonomous vehicles
and there's a lot of interesting
challenges around that and the DARPA
robotics challenge is when uh humanoid
robots were tasked to do all kinds of
uh you know manipulation walking driving
your car all these kinds of challenges
with if I remember correctly sort of
some slight capability to communicate
with humans but the communication was
very poor so it basically has to be
almost entirely autonomous you can have
periods where the communication was
entirely interrupted and the robot had
to be able to proceed yeah but you could
provide some high level guidance to the
robot basically load low bandwidth
Communications to steer it I watched
that challenge with kind of tears in my
eyes eating popcorn
but I wasn't personally losing uh you
know hundreds of thousands millions of
dollars and many years of incredible
hard work by some of the most brilliant
roboticists in the world so that was why
the tragic that's why tears came so
anyway what what have you uh just
looking back to that time what have you
learned from that experience
I mean maybe if you could describe what
it was uh sort of the setup for people
who haven't seen it well so there was a
contest where a bunch of different
um robots were asked to do a series of
tasks uh some of those that you
mentioned drive a vehicle get out open a
door go identify a vowel shutter valve
use a tool to maybe cut a hole in
um
a surface and then crawl over some
stairs and maybe some rough Terrain
so it was
the idea was have a general purpose
robot that could do lots of different
things
um it had to be mobility and
manipulation on board perception
and there was a contest which DARPA
likes at the time was running sort of
follow on to the
the Grand Challenge which was let's
let's try to push vehicle autonomy along
right they they encourage people to
build autonomous cars so they're trying
to basically push an industry forward
and
um
uh we were asked our role in this was to
build
um a humanoid at the time it was our
sort of first generation Atlas robot
and we built maybe 10 of them I don't
remember the exact number
and DARPA distributed those to various
teams
um that sort of won a contest showed
that they could you know program these
robots and then use them to compete
against each other and then other robots
were introduced as well some teams built
their own robots Carnegie
um melon for example built their own
robot and uh and all these robots
competed to see who could sort of get
through this this maze or the fastest
and again I think the purpose was to
kind of push the whole industry forward
we provided the robot and some baseline
software but we didn't we didn't
actually compete as a participant uh
where we were trying to uh you know
Drive the robot through this maze we
were just trying to support the other
teams
it was humbling because it was it was
really a hard task and honestly the
robots the tears were because mostly the
robots didn't do it you know they fell
down
you know repeatedly
um it was hard to get through this
contest uh some did and and you know
they were rewarded and won
but it was humbling because of just how
hard these tasks weren't all that hard a
person could have done it very easily
but it was really hard to get the robots
to do it you know the general nature of
it the variety of it the variety and
also that I don't know if the tasks were
sort of the task in themselves help us
understand what is difficult and what is
not I don't know if that was obvious
before the contest was designed so you
kind of tried to figure that out and I
think Atlas is really a general robot
platform and it's perhaps not best
suited for the specific tasks of that
contest like for just for example
probably the hardest task is not the
driving of the car but getting in and
out of the car and Atlas probably you
know if you were to design a robot that
can get into the car easily and get out
easily you probably would not make Atlas
that particular car yeah the robot was a
little bit big to get in and out of that
car right it doesn't fit yeah this is
the curse of a general purpose robot
that they're not perfect at any one
thing
but they might be able to do a wide
variety of things and and that is
that is the goal at the end of the day
you know I think we all want to build
general purpose robots that can be used
for lots of different activities
but it's hard and
um and the wisdom in in building
successful robots up until this point
have been go build a robot for a
specific task and it'll do it very well
and as long as you control that
environment it'll operate perfectly but
but robots need to be able to deal with
uncertainty if they're going to be
useful to us in the future
they need to be able to deal with
unexpected uh situations and that's sort
of the goal of a general purpose or
multi-purpose robot
and that's just darn hard and so some of
you know there's these curious little
failures like I remember one of the a
robot you know the first the first time
you start to try to push on the world
with a robot you you forget that the
world pushes back and and will push you
over if you're not ready for it and the
robot you know reached to grab the door
handle I think it missed the grasp of
the door handle was expecting that its
hand was on the door handle and so when
it tried to turn the knob it just threw
itself over it didn't realize oh I had
missed the door handle I didn't have I
didn't I was expecting a force back from
the door it wasn't there and then I lost
my balance so these little simple things
that you and I would take totally for
granted and deal with the robots don't
know how to deal with yet and so you
have to start to deal with all of those
uh circumstances well I think a lot of
us experience this in uh even when sober
but drunk too uh sort of you pick up a
thing and expect it to be
what is it heavy and it turns out to be
light yeah oh yeah and then so the same
and I'm sure if your depth perception
for whatever reason is screwed up if
you're if you're drunk or some other
reason and then you think you're putting
your hand on the table and you miss it I
mean it's the same kind of situation
yeah but
there's why you need to be able to
predict forward just a little bit and so
that's where this model predictive
control stuff comes in predict forward
what you think is going to happen
and then if and if that does happen
you're in good shape if something else
happens you better start predicting
again so if we like we re-uh regenerate
a plan yeah when you don't I mean that
um that also requires a very uh fast
feedback loop of updating
what your prediction how it matches to
the actual real world
yeah those things have to run pretty
quickly what's the challenge of running
things pretty quickly a thousand Hertz
of acting and sensing
quickly you know there's a few different
layers of that you you want at the
lowest level you like to run things
typically at around a thousand Hertz
which means that you know at each joint
of the robot you're measuring position
or force and then trying to control your
actuator whether it's a hydraulic or
electric motor trying to control the
force coming out of that actuator and
you want to do that really fast
something like a thousand Hertz and that
means you can't have too much
calculation going on at that joint
um but that's pretty manageable these
days and it's fairly common
and then there's another layer that
you're probably calculating you know
maybe at 100 Hertz maybe 10 times slower
which is now starting to look at the
overall body motion and thinking about
the the larger physics of of the uh of
the robot
um and then there's yet another loop
that's probably happening a little bit
slower which is where you start to bring
you know your perception and your vision
and things like that and so you need to
run all of these Loops sort of
simultaneously you do have to manage
your your computer time so that you can
squeeze in all the calculations you need
in real time in a very consistent way
and the amount of calculation we can do
is increasing as computers get better
which means we can start to do more
sophisticated calculations I can have a
more complex model
doing my forward prediction
and and that might allow me to do even
better predictions as I as I get better
and better and and it used to be again
we had
you know 10 years ago
we had to have pretty simple models
that we were running you know at those
fast rates because the computers weren't
as capable about calculating forward
with a sophisticated model but as as
computation gets better we can we can do
more of that what about the actual
pipeline of software engineering how
easy it is to keep updating Atlas like
Duke continuous development on it so how
many computers
are on there is there a nice pipeline
it's an important part of building a
team around it which means you know you
need to also have a software tools
simulation tools you know so
um we have always made strong use of
physics-based simulation tools to do uh
some of this calculation basically
tested in simulation before you put it
on the robot but you also want the same
code that you're running in simulation
to be the same code you're running on
the hardware and so even getting to the
point where it was the same code going
from one to the other
we probably didn't really get that
working until you know a few years
several years ago
um but that was a you know that was a
bit of a milestone and so you want to
work certainly work these pipelines so
that you can make it as easy as possible
and have a bunch of people working in
parallel especially when you know we
only have you know four of the atlas
robots the modern Atlas robots at the
company
and you know we probably have you know
40 developers there all trying to gain
access to it and so you need to share
resources and use some of these uh some
of the software pipeline well that's a
really exciting step to be able to run
the exact same code and simulation as on
the actual robot uh how hard is it to do
uh realistic simulation physics-based
simulation of of Atlas such that I mean
the dream is like if it works in
simulation works perfectly in reality
how hard is it to sort of keep workout
closing that Gap the root of some of our
physics-based simulation tools really
started at MIT and
um we built some some good physics-based
modeling tools there
the early days of the company we were
trying to develop those tools as a
commercial product so we continued to
develop them it wasn't a particularly
successful commercial product but we
ended up with some nice physics-based
simulation tools so that when we started
doing legged robotics again we had a
really nice tool to work with
and the things we paid attention to were
were things that weren't necessarily
handled very well in the commercial
tools you could buy off the shelf like
like interaction with the world like
foot ground contact so trying to model
those contact
um events well
in a way that captured the important
parts of the interaction
was a really important element uh to get
right and to also do in a way that was
computationally feasible
and could run fast because if you if
your simulation runs too slow you know
then your developers are sitting around
waiting for stuff to run and compile so
it's always about efficient uh a fast
operation as well so that's been a big
part of it you know I think developing
those tools in parallel to the
development of the the platform and
trying to scale them has has really been
essential I'd say to us being able to
assemble a team of people that could do
this yeah how to simulate contact
periods so flick ground contact but sort
of for manipulation
because don't you want to model all
kinds of surfaces yeah so it will be
even more complex with manipulation
because there's a lot more going on you
know and you need to capture I don't
know things slipping and moving you know
in in your in your hand
um it's a level of complexity that I
think goes above foot ground contact
when you really start doing dexterous
manipulation so there's challenges ahead
still so how far are we away from me
being able to walk with Atlas in the
sand along the beach and
us both drinking a beer
yeah maybe we can out of a kid maybe
Atlas could spill his beer because he's
got nowhere to put it
Alice could walk on the sand uh so can
it yeah uh yeah I mean you know have we
really had him out on the beach you know
we take them outside often you know
rocks Hills that sort of thing even just
around our lab in Waltham we probably
haven't been on the sand but I'm a salt
surface I don't doubt that we could deal
with it yeah we we might have to spend a
little bit of time to sort of make that
work but we did take uh we we had a had
to take
big dog to Thailand years ago and uh we
did this great video of the robot
walking in the sand walking into the
ocean up to I don't know its belly or
something like that and then turning
around and walking out all while playing
some cool beach music yeah great show
but then you know we didn't really clean
the robot off and the salt water was
really hard on it so you know we put it
in a box shipped it back by the time it
came back we had some problems with salt
it's the salt water it's not like old
stuff it's not like sand getting into
the components or something like this
but I'm sure if if this is a big
priority you can make it like waterproof
right that just wasn't our our goal at
the time well it's a personal goal of
mine to it walk along the beach but it's
a human problem too you get sand
everywhere it's it's just a jam mess
so soft surfaces are okay so I mean can
we just uh link on the the robotics
challenge there's a there's a pile of uh
like Rubble they had to walk over is
that's
um how difficult is that task
in the early days of developing big dog
the loose Rock was the epitome of the
hard walking surface because you step
down and then the Rock and you had these
little Point feet on the robot
and the rock and roll and and you have
to deal with that last minute you know
change in your foot placement yes so you
you step on the thing and that thing
responds to you stepping on it yeah and
and it moves where your point of support
is and so it's really that that became
kind of the essence of the test and so
that was the beginning of us starting to
build Rock piles in our parking lots and
and we would actually build boxes full
of rocks and bring them into the lab and
then we would have the robots walking
across these boxes of rocks because that
became the essential test
so you mentioned big dog can you can we
maybe take a stroll through the history
of Boston Dynamics uh so what and who's
Big Dog by the way is who
do you try not to anthropomorphize the
robots do you try not to
to try to remember that they're this is
like the division I have because I for
me it's impossible
for me there's a magic to the to the
being that is a robot it is not human
but it is
the same Magic
uh the living being has when it moves
about the world is there in the robot so
um I don't know what question I'm asking
but uh should I say what or who I guess
who is Big Dog what is big dog
well I'll say to address the medic
question
we don't try to draw hard lines around
it being an it or a him or a her
um it's okay right people I think part
of the magic of these kinds of machines
is by nature of their organic movement
of the of their Dynamics
we tend to want to identify with them we
tend to look at them and sort of
attribute
maybe feeling to that because we've only
seen things that move like this that
were alive
and so
um this is an opportunity it means that
you could have
feelings for a machine and you know
people have feelings for their cars you
know they get attracted to them attached
to them so that's inherently could be a
good thing as long as we manage what
that interaction is
so we don't put strong boundaries around
this and ultimately think it's a benefit
but it's also can be a bit of a curse
because I think people look at these
machines
and they attribute a level of
intelligence that the machines don't
have why because again they've seen
things move like this that we're living
beings
which are intelligent
and so they want to attribute
intelligence to the robots that isn't
appropriate yet even though they move
like an intelligent being but you try to
acknowledge that the
anthropomorphization is there and try to
first of all acknowledge that it's there
and have a little fun with it you know
our most recent video
it's just kind of fun you know to to
look at the robot we started off the the
video with Atlas
um kind of looking around for where the
bag of tools was because the guy up on
the scaffolding says send me some tools