back to indexRuss Tedrake: Underactuated Robotics, Control, Dynamics and Touch | Lex Fridman Podcast #114
link |
The following is a conversation with Russ Tedrick,
link |
a roboticist and professor at MIT
link |
and vice president of robotics research
link |
at Toyota Research Institute or TRI.
link |
He works on control of robots in interesting,
link |
complicated, underactuated, stochastic,
link |
difficult to model situations.
link |
He's a great teacher and a great person,
link |
one of my favorites at MIT.
link |
We'll get into a lot of topics in this conversation
link |
from his time leading MIT's Delta Robotics Challenge team
link |
to the awesome fact that he often runs
link |
close to a marathon a day to and from work barefoot.
link |
For a world class roboticist interested in elegant,
link |
efficient control of underactuated dynamical systems
link |
like the human body, this fact makes Russ
link |
one of the most fascinating people I know.
link |
Quick summary of the ads.
link |
Three sponsors, Magic Spoon Cereal, BetterHelp,
link |
Please consider supporting this podcast
link |
by going to magicspoon.com slash lex
link |
and using code lex at checkout,
link |
going to betterhelp.com slash lex
link |
and signing up at expressvpn.com slash lexpod.
link |
Click the links in the description,
link |
buy the stuff, get the discount.
link |
It really is the best way to support this podcast.
link |
If you enjoy this thing, subscribe on YouTube,
link |
review it with five stars on Apple Podcast,
link |
support it on Patreon, or connect with me
link |
on Twitter at lexfreedman.
link |
As usual, I'll do a few minutes of ads now
link |
and never any ads in the middle
link |
that can break the flow of the conversation.
link |
This episode is supported by Magic Spoon,
link |
low carb keto friendly cereal.
link |
I've been on a mix of keto or carnivore diet
link |
for a very long time now.
link |
That means eating very little carbs.
link |
I used to love cereal.
link |
Obviously, most have crazy amounts of sugar,
link |
which is terrible for you, so I quit years ago,
link |
but Magic Spoon is a totally new thing.
link |
Zero sugar, 11 grams of protein,
link |
and only three net grams of carbs.
link |
It tastes delicious.
link |
It has a bunch of flavors, they're all good,
link |
but if you know what's good for you,
link |
you'll go with cocoa, my favorite flavor
link |
and the flavor of champions.
link |
Click the magicspoon.com slash lex link in the description,
link |
use code lex at checkout to get the discount
link |
and to let them know I sent you.
link |
So buy all of their cereal.
link |
It's delicious and good for you.
link |
You won't regret it.
link |
This show is also sponsored by BetterHelp,
link |
spelled H E L P Help.
link |
Check it out at betterhelp.com slash lex.
link |
They figure out what you need
link |
and match you with a licensed professional therapist
link |
in under 48 hours.
link |
It's not a crisis line, it's not self help,
link |
it is professional counseling done securely online.
link |
As you may know, I'm a bit from the David Goggins line
link |
of creatures and still have some demons to contend with,
link |
usually on long runs or all nighters full of self doubt.
link |
I think suffering is essential for creation,
link |
but you can suffer beautifully
link |
in a way that doesn't destroy you.
link |
For most people, I think a good therapist can help in this.
link |
So it's at least worth a try.
link |
Check out the reviews, they're all good.
link |
It's easy, private, affordable, available worldwide.
link |
You can communicate by text anytime
link |
and schedule weekly audio and video sessions.
link |
Check it out at betterhelp.com slash lex.
link |
This show is also sponsored by ExpressVPN.
link |
Get it at expressvpn.com slash lex pod
link |
to get a discount and to support this podcast.
link |
Have you ever watched The Office?
link |
If you have, you probably know it's based
link |
on a UK series also called The Office.
link |
Not to stir up trouble, but I personally think
link |
the British version is actually more brilliant
link |
than the American one, but both are amazing.
link |
Anyway, there are actually nine other countries
link |
with their own version of The Office.
link |
You can get access to them with no geo restriction
link |
when you use ExpressVPN.
link |
It lets you control where you want sites
link |
to think you're located.
link |
You can choose from nearly 100 different countries,
link |
giving you access to content
link |
that isn't available in your region.
link |
So again, get it on any device at expressvpn.com slash lex pod
link |
to get an extra three months free
link |
and to support this podcast.
link |
And now here's my conversation with Russ Tedrick.
link |
What is the most beautiful motion
link |
of an animal or robot that you've ever seen?
link |
I think the most beautiful motion of a robot
link |
has to be the passive dynamic walkers.
link |
I think there's just something fundamentally beautiful.
link |
The ones in particular that Steve Collins built
link |
with Andy Ruina at Cornell, a 3D walking machine.
link |
So it was not confined to a boom or a plane
link |
that you put it on top of a small ramp,
link |
give it a little push, it's powered only by gravity.
link |
No controllers, no batteries whatsoever.
link |
It just falls down the ramp.
link |
And at the time it looked more natural, more graceful,
link |
more human like than any robot we'd seen to date
link |
powered only by gravity.
link |
Well, okay, the simplest model, it's kind of like a slinky.
link |
It's like an elaborate slinky.
link |
One of the simplest models we used to think about it
link |
is actually a rimless wheel.
link |
So imagine taking a bicycle wheel, but take the rim off.
link |
So it's now just got a bunch of spokes.
link |
If you give that a push,
link |
it still wants to roll down the ramp,
link |
but every time its foot, its spoke comes around
link |
and hits the ground, it loses a little energy.
link |
Every time it takes a step forward,
link |
it gains a little energy.
link |
Those things can come into perfect balance.
link |
And actually they want to, it's a stable phenomenon.
link |
If it's going too slow, it'll speed up.
link |
If it's going too fast, it'll slow down
link |
and it comes into a stable periodic motion.
link |
Now you can take that rimless wheel,
link |
which doesn't look very much like a human walking,
link |
take all the extra spokes away, put a hinge in the middle.
link |
Now it's two legs.
link |
That's called our compass gait walker.
link |
That can still, you give it a little push,
link |
it starts falling down a ramp.
link |
It looks a little bit more like walking.
link |
At least it's a biped.
link |
But what Steve and Andy,
link |
and Tad McGeer started the whole exercise,
link |
but what Steve and Andy did was they took it
link |
to this beautiful conclusion
link |
where they built something that had knees, arms, a torso.
link |
The arms swung naturally, give it a little push.
link |
And that looked like a stroll through the park.
link |
How do you design something like that?
link |
I mean, is that art or science?
link |
It's on the boundary.
link |
I think there's a science to getting close to the solution.
link |
I think there's certainly art in the way
link |
that they made a beautiful robot.
link |
But then the finesse, because they were working
link |
with a system that wasn't perfectly modeled,
link |
wasn't perfectly controlled,
link |
there's all these little tricks
link |
that you have to tune the suction cups at the knees,
link |
for instance, so that they stick,
link |
but then they release at just the right time.
link |
Or there's all these little tricks of the trade,
link |
which really are art, but it was a point.
link |
I mean, it made the point.
link |
We were, at that time, the walking robot,
link |
the best walking robot in the world was Honda's Asmo.
link |
Absolutely marvel of modern engineering.
link |
This was in 97 when they first released.
link |
It sort of announced P2, and then it went through.
link |
It was Asmo by then in 2004.
link |
And it looks like this very cautious walking,
link |
like you're walking on hot coals or something like that.
link |
I think it gets a bad rap.
link |
Asmo is a beautiful machine.
link |
It does walk with its knees bent.
link |
Our Atlas walking had its knees bent.
link |
But actually, Asmo was pretty fantastic.
link |
But it wasn't energy efficient.
link |
Neither was Atlas when we worked on Atlas.
link |
None of our robots that have been that complicated
link |
have been very energy efficient.
link |
But there's a thing that happens when you do control,
link |
when you try to control a system of that complexity.
link |
You try to use your motors to basically counteract gravity.
link |
Take whatever the world's doing to you and push back,
link |
erase the dynamics of the world,
link |
and impose the dynamics you want
link |
because you can make them simple and analyzable,
link |
mathematically simple.
link |
And this was a very sort of beautiful example
link |
that you don't have to do that.
link |
You can just let go.
link |
Let physics do most of the work, right?
link |
And you just have to give it a little bit of energy.
link |
This one only walked down a ramp.
link |
It would never walk on the flat.
link |
To walk on the flat,
link |
you have to give a little energy at some point.
link |
But maybe instead of trying to take the forces imparted
link |
to you by the world and replacing them,
link |
what we should be doing is letting the world push us around
link |
and we go with the flow.
link |
Very zen, very zen robot.
link |
Yeah, but okay, so that sounds very zen,
link |
but I can also imagine how many like failed versions
link |
they had to go through.
link |
Like how many, like, I would say it's probably,
link |
would you say it's in the thousands
link |
that they've had to have the system fall down
link |
before they figured out how to get it?
link |
I don't know if it's thousands, but it's a lot.
link |
It takes some patience.
link |
There's no question.
link |
So in that sense, control might help a little bit.
link |
Oh, I think everybody, even at the time,
link |
said that the answer is to do with that with control.
link |
But it was just pointing out
link |
that maybe the way we're doing control right now
link |
isn't the way we should.
link |
So what about on the animal side,
link |
the ones that figured out how to move efficiently?
link |
Is there anything you find inspiring or beautiful
link |
in the movement of any particular animal?
link |
I do have a favorite example.
link |
So it sort of goes with the passive walking idea.
link |
So is there, you know, how energy efficient are animals?
link |
Okay, there's a great series of experiments
link |
by George Lauder at Harvard and Mike Tranofilo at MIT.
link |
They were studying fish swimming in a water tunnel.
link |
And one of these, the type of fish they were studying
link |
were these rainbow trout,
link |
because there was a phenomenon well understood
link |
that rainbow trout, when they're swimming upstream
link |
in mating season, they kind of hang out behind the rocks.
link |
And it looks like, I mean,
link |
that's tiring work swimming upstream.
link |
They're hanging out behind the rocks.
link |
Maybe there's something energetically interesting there.
link |
So they tried to recreate that.
link |
They put in this water tunnel, a rock basically,
link |
a cylinder that had the same sort of vortex street,
link |
the eddies coming off the back of the rock
link |
that you would see in a stream.
link |
And they put a real fish behind this
link |
and watched how it swims.
link |
And the amazing thing is that if you watch from above
link |
what the fish swims when it's not behind a rock,
link |
it has a particular gate.
link |
You can identify the fish the same way you look
link |
at a human walking down the street.
link |
You sort of have a sense of how a human walks.
link |
The fish has a characteristic gate.
link |
You put that fish behind the rock, its gate changes.
link |
And what they saw was that it was actually resonating
link |
and kind of surfing between the vortices.
link |
Now, here was the experiment that really was the clincher.
link |
Because there was still, it wasn't clear how much of that
link |
was mechanics of the fish,
link |
how much of that is control, the brain.
link |
So the clincher experiment,
link |
and maybe one of my favorites to date,
link |
although there are many good experiments.
link |
They took, this was now a dead fish.
link |
They took a dead fish.
link |
They put a string that went,
link |
that tied the mouth of the fish to the rock
link |
so it couldn't go back and get caught in the grates.
link |
And then they asked what would that dead fish do
link |
when it was hanging out behind the rock?
link |
And so what you'd expect, it sort of flopped around
link |
like a dead fish in the vortex wake
link |
until something sort of amazing happens.
link |
And this video is worth putting in, right?
link |
The dead fish basically starts swimming upstream, right?
link |
It's completely dead, no brain, no motors, no control.
link |
But it's somehow the mechanics of the fish
link |
resonate with the vortex street
link |
and it starts swimming upstream.
link |
It's one of the best examples ever.
link |
Who do you give credit for that to?
link |
Is that just evolution constantly just figuring out
link |
by killing a lot of generations of animals,
link |
like the most efficient motion?
link |
Is that, or maybe the physics of our world completely like,
link |
is like if evolution applied not only to animals,
link |
but just the entirety of it somehow drives to efficiency,
link |
like nature likes efficiency?
link |
I don't know if that question even makes any sense.
link |
I understand the question.
link |
That's reasonable.
link |
I mean, do they co evolve?
link |
Yeah, somehow co, yeah.
link |
Like I don't know if an environment can evolve, but.
link |
I mean, there are experiments that people do,
link |
careful experiments that show that animals can adapt
link |
to unusual situations and recover efficiency.
link |
So there seems like at least in one direction,
link |
I think there is reason to believe
link |
that the animal's motor system and probably its mechanics
link |
adapt in order to be more efficient.
link |
But efficiency isn't the only goal, of course.
link |
Sometimes it's too easy to think about only efficiency,
link |
but we have to do a lot of other things first, not get eaten.
link |
And then all other things being equal, try to save energy.
link |
By the way, let's draw a distinction
link |
between control and mechanics.
link |
Like how would you define each?
link |
I mean, I think part of the point is that
link |
we shouldn't draw a line as clearly as we tend to.
link |
But on a robot, we have motors
link |
and we have the links of the robot, let's say.
link |
If the motors are turned off,
link |
the robot has some passive dynamics, okay?
link |
Gravity does the work.
link |
You can put springs, I would call that mechanics, right?
link |
If we have springs and dampers,
link |
which our muscles are springs and dampers and tendons.
link |
But then you have something that's doing active work,
link |
putting energy in, which are your motors on the robot.
link |
The controller's job is to send commands to the motor
link |
that add new energy into the system, right?
link |
So the mechanics and control interplay somewhere,
link |
the divide is around, you know,
link |
did you decide to send some commands to your motor
link |
or did you just leave the motors off,
link |
let them do their work?
link |
Would you say is most of nature
link |
on the dynamic side or the control side?
link |
So like, if you look at biological systems,
link |
we're living in a pandemic now,
link |
like, do you think a virus is a,
link |
do you think it's a dynamic system
link |
or is there a lot of control, intelligence?
link |
I think it's both, but I think we maybe have underestimated
link |
how important the dynamics are, right?
link |
I mean, even our bodies, the mechanics of our bodies,
link |
certainly with exercise, they evolve.
link |
But so I actually, I lost a finger in early 2000s
link |
and it's my fifth metacarpal.
link |
And it turns out you use that a lot
link |
in ways you don't expect when you're opening jars,
link |
even when I'm just walking around,
link |
if I bump it on something, there's a bone there
link |
that was used to taking contact.
link |
My fourth metacarpal wasn't used to taking contact,
link |
it used to hurt, it still does a little bit.
link |
But actually my bone has remodeled, right?
link |
Over a couple of years, the geometry,
link |
the mechanics of that bone changed
link |
to address the new circumstances.
link |
So the idea that somehow it's only our brain
link |
that's adapting or evolving is not right.
link |
Maybe sticking on evolution for a bit,
link |
because it's tended to create some interesting things.
link |
Bipedal walking, why the heck did evolution give us,
link |
I think we're, are we the only mammals that walk on two feet?
link |
No, I mean, there's a bunch of animals that do it a bit.
link |
I think we are the most successful bipeds.
link |
I think I read somewhere that the reason
link |
the evolution made us walk on two feet
link |
is because there's an advantage
link |
to being able to carry food back to the tribe
link |
or something like that.
link |
So like you can carry, it's kind of this communal,
link |
cooperative thing, so like to carry stuff back
link |
to a place of shelter and so on to share with others.
link |
Do you understand at all the value of walking on two feet
link |
from both a robotics and a human perspective?
link |
Yeah, there are some great books written
link |
about evolution of, walking evolution of the human body.
link |
I think it's easy though to make bad evolutionary arguments.
link |
Sure, most of them are probably bad,
link |
but what else can we do?
link |
I mean, I think a lot of what dominated our evolution
link |
probably was not the things that worked well
link |
sort of in the steady state, you know,
link |
when things are good, but for instance,
link |
people talk about what we should eat now
link |
because our ancestors were meat eaters or whatever.
link |
Oh yeah, I love that, yeah.
link |
But probably, you know, the reason
link |
that one pre Homo sapiens species versus another survived
link |
was not because of whether they ate well
link |
when there was lots of food.
link |
But when the ice age came, you know,
link |
probably one of them happened to be in the wrong place.
link |
One of them happened to forage a food that was okay
link |
even when the glaciers came or something like that, I mean.
link |
There's a million variables that contributed
link |
and we can't, and our, actually the amount of information
link |
we're working with and telling these stories,
link |
these evolutionary stories is very little.
link |
So yeah, just like you said, it seems like,
link |
if you study history, it seems like history turns
link |
on like these little events that otherwise
link |
would seem meaningless, but in a grant,
link |
like when you, in retrospect, were turning points.
link |
And that's probably how like somebody got hit in the head
link |
with a rock because somebody slept with the wrong person
link |
back in the cave days and somebody get angry
link |
and that turned, you know, warring tribes
link |
combined with the environment, all those millions of things
link |
and the meat eating, which I get a lot of criticism
link |
because I don't know what your dietary processes are like,
link |
but these days I've been eating only meat,
link |
which is, there's a large community of people who say,
link |
yeah, probably make evolutionary arguments
link |
and say you're doing a great job.
link |
There's probably an even larger community of people,
link |
including my mom, who says it's deeply unhealthy,
link |
it's wrong, but I just feel good doing it.
link |
But you're right, these evolutionary arguments
link |
can be flawed, but is there anything interesting
link |
There's a great book, by the way,
link |
well, a series of books by Nicholas Taleb
link |
about Fooled by Randomness and Black Swan.
link |
Highly recommend them, but yeah,
link |
they make the point nicely that probably
link |
it was a few random events that, yes,
link |
maybe it was someone getting hit by a rock, as you say.
link |
That said, do you think, I don't know how to ask this
link |
question or how to talk about this,
link |
but there's something elegant and beautiful
link |
about moving on two feet, obviously biased
link |
because I'm human, but from a robotics perspective, too,
link |
you work with robots on two feet,
link |
is it all useful to build robots that are on two feet
link |
as opposed to four?
link |
Is there something useful about it?
link |
I think the most, I mean, the reason I spent a long time
link |
working on bipedal walking was because it was hard
link |
and it challenged control theory in ways
link |
that I thought were important.
link |
I wouldn't have ever tried to convince you
link |
that you should start a company around bipeds
link |
or something like this.
link |
There are people that make pretty compelling arguments.
link |
I think the most compelling one is that the world
link |
is built for the human form, and if you want a robot
link |
to work in the world we have today,
link |
then having a human form is a pretty good way to go.
link |
There are places that a biped can go that would be hard
link |
for other form factors to go, even natural places,
link |
but at some point in the long run,
link |
we'll be building our environments for our robots, probably,
link |
and so maybe that argument falls aside.
link |
So you famously run barefoot.
link |
Do you still run barefoot?
link |
I still run barefoot.
link |
That's so awesome.
link |
Much to my wife's chagrin.
link |
Do you want to make an evolutionary argument
link |
for why running barefoot is advantageous?
link |
What have you learned about human and robot movement
link |
in general from running barefoot?
link |
Human or robot and or?
link |
Well, you know, it happened the other way, right?
link |
So I was studying walking robots,
link |
and there's a great conference called
link |
the Dynamic Walking Conference where it brings together
link |
both the biomechanics community
link |
and the walking robots community.
link |
And so I had been going to this for years
link |
and hearing talks by people who study barefoot running
link |
and other, the mechanics of running.
link |
So I did eventually read Born to Run.
link |
Most people read Born to Run in the first, right?
link |
The other thing I had going for me is actually
link |
that I wasn't a runner before,
link |
and I learned to run after I had learned
link |
about barefoot running, or I mean,
link |
started running longer distances.
link |
So I didn't have to unlearn.
link |
And I'm definitely, I'm a big fan of it for me,
link |
but I'm not going to,
link |
I tend to not try to convince other people.
link |
There's people who run beautifully with shoes on,
link |
But here's why it makes sense for me.
link |
It's all about the longterm game, right?
link |
So I think it's just too easy to run 10 miles,
link |
feel pretty good, and then you get home at night
link |
and you realize my knees hurt.
link |
I did something wrong, right?
link |
If you take your shoes off,
link |
then if you hit hard with your foot at all,
link |
You don't like run 10 miles
link |
and then realize you've done some damage.
link |
You have immediate feedback telling you
link |
that you've done something that's maybe suboptimal,
link |
and you change your gait.
link |
I mean, it's even subconscious.
link |
If I, right now, having run many miles barefoot,
link |
if I put a shoe on, my gait changes
link |
in a way that I think is not as good.
link |
So it makes me land softer.
link |
And I think my goals for running
link |
are to do it for as long as I can into old age,
link |
not to win any races.
link |
And so for me, this is a way to protect myself.
link |
Yeah, I think, first of all,
link |
I've tried running barefoot many years ago,
link |
probably the other way,
link |
just reading Born to Run.
link |
But just to understand,
link |
because I felt like I couldn't put in the miles
link |
And it feels like running for me,
link |
and I think for a lot of people,
link |
was one of those activities that we do often
link |
and we never really try to learn to do correctly.
link |
Like, it's funny, there's so many activities
link |
we do every day, like brushing our teeth, right?
link |
I think a lot of us, at least me,
link |
probably have never deeply studied
link |
how to properly brush my teeth, right?
link |
Or wash, as now with the pandemic,
link |
or how to properly wash our hands.
link |
We do it every day, but we haven't really studied,
link |
like, am I doing this correctly?
link |
But running felt like one of those things,
link |
it was absurd not to study how to do correctly,
link |
because it's the source of so much pain and suffering.
link |
Like, I hate running, but I do it.
link |
I do it because I hate it, but I feel good afterwards.
link |
But I think it feels like you need
link |
to learn how to do it properly.
link |
So that's where barefoot running came in,
link |
and then I quickly realized that my gait
link |
was completely wrong.
link |
I was taking huge steps,
link |
and landing hard on the heel, all those elements.
link |
And so, yeah, from that I actually learned
link |
to take really small steps, look.
link |
I already forgot the number,
link |
but I feel like it was 180 a minute or something like that.
link |
And I remember I actually just took songs
link |
that are 180 beats per minute,
link |
and then like tried to run at that beat,
link |
and just to teach myself.
link |
It took a long time, and I feel like after a while,
link |
you learn to run, you adjust properly,
link |
without going all the way to barefoot.
link |
But I feel like barefoot is the legit way to do it.
link |
I mean, I think a lot of people
link |
would be really curious about it.
link |
Can you, if they're interested in trying,
link |
what would you, how would you recommend
link |
they start, or try, or explore?
link |
That's the biggest thing people do,
link |
is they are excellent runners,
link |
and they're used to running long distances,
link |
or running fast, and they take their shoes off,
link |
and they hurt themselves instantly trying to do
link |
something that they were used to doing.
link |
I think I lucked out in the sense
link |
that I couldn't run very far when I first started trying.
link |
And I run with minimal shoes too.
link |
I mean, I will bring along a pair of,
link |
actually, like aqua socks or something like this,
link |
I can just slip on, or running sandals,
link |
I've tried all of them.
link |
What's the difference between a minimal shoe
link |
and nothing at all?
link |
What's, like, feeling wise, what does it feel like?
link |
There is a, I mean, I notice my gait changing, right?
link |
So, I mean, your foot has as many muscles
link |
and sensors as your hand does, right?
link |
Sensors, ooh, okay.
link |
And we do amazing things with our hands.
link |
And we stick our foot in a big, solid shoe, right?
link |
So there's, I think, you know, when you're barefoot,
link |
you're just giving yourself more proprioception.
link |
And that's why you're more aware of some of the gait flaws
link |
and stuff like this.
link |
Now, you have less protection too, so.
link |
I mean, yeah, so I think people who are afraid
link |
of barefoot running are worried about getting cuts
link |
or stepping on rocks.
link |
First of all, even if that was a concern,
link |
I think those are all, like, very short term.
link |
You know, if I get a scratch or something,
link |
it'll heal in a week.
link |
If I blow out my knees, I'm done running forever.
link |
So I will trade the short term for the long term anytime.
link |
But even then, you know, and this, again,
link |
to my wife's chagrin, your feet get tough, right?
link |
And, yeah, I can run over almost anything now.
link |
I mean, what, can you talk about,
link |
is there, like, is there tips or tricks
link |
that you have, suggestions about,
link |
like, if I wanted to try it?
link |
You know, there is a good book, actually.
link |
There's probably more good books since I read them.
link |
But Ken Bob, Barefoot Ken Bob Saxton.
link |
He's an interesting guy.
link |
But I think his book captures the right way
link |
to describe running, barefoot running,
link |
to somebody better than any other I've seen.
link |
So you run pretty good distances, and you bike,
link |
and is there, you know, if we talk about bucket list items,
link |
is there something crazy on your bucket list,
link |
athletically, that you hope to do one day?
link |
I mean, my commute is already a little crazy.
link |
What are we talking about here?
link |
What distance are we talking about?
link |
Well, I live about 12 miles from MIT,
link |
but you can find lots of different ways to get there.
link |
So, I mean, I've run there for many years, I've biked there.
link |
Yeah, but normally I would try to run in
link |
and then bike home, bike in, run home.
link |
But you have run there and back before?
link |
Yeah, or with minimal shoes or whatever that.
link |
It became kind of a game of how can I get to work?
link |
I've rollerbladed, I've done all kinds of weird stuff,
link |
but my favorite one these days,
link |
I've been taking the Charles River to work.
link |
So, I can put in the rowboat not so far from my house,
link |
but the Charles River takes a long way to get to MIT,
link |
so I can spend a long time getting there.
link |
And it's not about, I don't know, it's just about,
link |
I've had people ask me,
link |
how can you justify taking that time?
link |
But for me, it's just a magical time to think,
link |
to compress, decompress.
link |
Especially, I'll wake up, do a lot of work in the morning,
link |
and then I kind of have to just let that settle
link |
before I'm ready for all my meetings.
link |
And then on the way home, it's a great time to sort of
link |
You lead a large group of people.
link |
Is there days where you're like,
link |
oh shit, I gotta get to work in an hour?
link |
Like, I mean, is there a tension there?
link |
And like, if we look at the grand scheme of things,
link |
just like you said, long term,
link |
that meeting probably doesn't matter.
link |
Like, you can always say, I'll just, I'll run
link |
and let the meeting happen, how it happens.
link |
Like, what, how do you, that zen, how do you,
link |
what do you do with that tension
link |
between the real world saying urgently,
link |
you need to be there, this is important,
link |
everything is melting down,
link |
how are we gonna fix this robot?
link |
There's this critical meeting,
link |
and then there's this, the zen beauty of just running,
link |
the simplicity of it, you along with nature.
link |
What do you do with that?
link |
I would say I'm not a fast runner, particularly.
link |
Probably my fastest splits ever was when
link |
I had to get to daycare on time
link |
because they were gonna charge me, you know,
link |
some dollar per minute that I was late.
link |
I've run some fast splits to daycare.
link |
But those times are past now.
link |
I think work, you can find a work life balance in that way.
link |
I think you just have to.
link |
I think I am better at work
link |
because I take time to think on the way in.
link |
So I plan my day around it,
link |
and I rarely feel that those are really at odds.
link |
So what, the bucket list item.
link |
If we're talking 12 times two, or approaching a marathon,
link |
what, have you run an ultra marathon before?
link |
Is there, what's a...
link |
I'm not gonna like take a dinghy across the Atlantic
link |
or something if that's what you want.
link |
But if someone does and wants to write a book,
link |
I would totally read it
link |
because I'm a sucker for that kind of thing.
link |
No, I do have some fun things that I will try.
link |
You know, I like to, when I travel,
link |
I almost always bike to Logan Airport
link |
and fold up a little folding bike
link |
and then take it with me and bike to wherever I'm going.
link |
And it's taken me,
link |
or I'll take a stand up paddle board these days
link |
and then I'll try to paddle around where I'm going
link |
And I've done some crazy things, but...
link |
But not for the, you know, I now talk,
link |
I don't know if you know who David Goggins is by any chance.
link |
Not well, but yeah.
link |
But I talk to him now every day.
link |
So he's the person who made me do this stupid challenge.
link |
So he's insane and he does things for the purpose
link |
in the best kind of way.
link |
He does things like for the explicit purpose of suffering.
link |
Like he picks the thing that,
link |
like whatever he thinks he can do, he does more.
link |
So is that, do you have that thing in you or are you...
link |
I think it's become the opposite.
link |
So you're like that dynamical system
link |
that the walker, the efficient...
link |
Yeah, it's leave no pain, right?
link |
You should end feeling better than you started.
link |
But it's mostly, I think, and COVID has tested this
link |
because I've lost my commute.
link |
I think I'm perfectly happy walking around town
link |
with my wife and kids if they could get them to go.
link |
And it's more about just getting outside
link |
and getting away from the keyboard for some time
link |
just to let things compress.
link |
Let's go into robotics a little bit.
link |
What to use the most beautiful idea in robotics?
link |
Whether we're talking about control
link |
or whether we're talking about optimization
link |
and the math side of things or the engineering side of things
link |
or the philosophical side of things.
link |
I think I've been lucky to experience something
link |
that not so many roboticists have experienced,
link |
which is to hang out
link |
with some really amazing control theorists.
link |
And the clarity of thought
link |
that some of the more mathematical control theory
link |
can bring to even very complex, messy looking problems
link |
is really, it really had a big impact on me
link |
and I had a day even just a couple of weeks ago
link |
where I had spent the day on a Zoom robotics conference
link |
having great conversations with lots of people.
link |
Felt really good about the ideas
link |
that were flowing and the like.
link |
And then I had a late afternoon meeting
link |
with one of my favorite control theorists
link |
and we went from these abstract discussions
link |
about maybes and what ifs and what a great idea
link |
to these super precise statements
link |
about systems that aren't that much more simple
link |
or abstract than the ones I care about deeply.
link |
And the contrast of that is,
link |
I don't know, it really gets me.
link |
I think people underestimate
link |
maybe the power of clear thinking.
link |
And so for instance, deep learning is amazing.
link |
I use it heavily in our work.
link |
I think it's changed the world, unquestionable.
link |
It makes it easy to get things to work
link |
without thinking as critically about it.
link |
So I think one of the challenges as an educator
link |
is to think about how do we make sure people get a taste
link |
of the more rigorous thinking
link |
that I think goes along with some different approaches.
link |
Yeah, so that's really interesting.
link |
So understanding like the fundamentals,
link |
the first principles of the problem,
link |
where in this case it's mechanics,
link |
like how a thing moves, how a thing behaves,
link |
like all the forces involved,
link |
like really getting a deep understanding of that.
link |
I mean, from physics, the first principle thing
link |
come from physics, and here it's literally physics.
link |
Yeah, and this applies, in deep learning,
link |
this applies to not just, I mean,
link |
it applies so cleanly in robotics,
link |
but it also applies to just in any data set.
link |
I find this true, I mean, driving as well.
link |
There's a lot of folks in that work on autonomous vehicles
link |
that work on autonomous vehicles that don't study driving,
link |
I might be coming a little bit from the psychology side,
link |
but I remember I spent a ridiculous number of hours
link |
at lunch, at this like lawn chair,
link |
and I would sit somewhere in MIT's campus,
link |
there's a few interesting intersections,
link |
and we'd just watch people cross.
link |
So we were studying pedestrian behavior,
link |
and I felt like, as we record a lot of video,
link |
to try, and then there's the computer vision
link |
extracts their movement, how they move their head, and so on,
link |
but like every time, I felt like I didn't understand enough.
link |
I just, I felt like I wasn't understanding
link |
what, how are people signaling to each other,
link |
what are they thinking,
link |
how cognizant are they of their fear of death?
link |
Like, what's the underlying game theory here?
link |
What are the incentives?
link |
And then I finally found a live stream of an intersection
link |
that's like high def that I just, I would watch
link |
so I wouldn't have to sit out there.
link |
But it's interesting, so like, I feel.
link |
But that's tough, that's a tough example,
link |
because I mean, the learning.
link |
Humans are involved.
link |
Not just because human, but I think the learning mantra
link |
is that basically the statistics of the data
link |
will tell me things I need to know, right?
link |
And, you know, for the example you gave
link |
of all the nuances of, you know, eye contact,
link |
or hand gestures, or whatever that are happening
link |
for these subtle interactions
link |
between pedestrians and traffic, right?
link |
Maybe the data will tell that story.
link |
I maybe even, one level more meta than what you're saying.
link |
For a particular problem,
link |
I think it might be the case
link |
that data should tell us the story.
link |
But I think there's a rigorous thinking
link |
that is just an essential skill
link |
for a mathematician or an engineer
link |
that I just don't wanna lose it.
link |
There are certainly super rigorous control,
link |
or sorry, machine learning people.
link |
I just think deep learning makes it so easy
link |
to do some things that our next generation,
link |
are not immediately rewarded
link |
for going through some of the more rigorous approaches.
link |
And then I wonder where that takes us.
link |
Well, I'm actually optimistic about it.
link |
I just want to do my part
link |
to try to steer that rigorous thinking.
link |
So there's like two questions I wanna ask.
link |
Do you have sort of a good example of rigorous thinking
link |
where it's easy to get lazy and not do the rigorous thinking?
link |
And the other question I have is like,
link |
do you have advice of how to practice rigorous thinking
link |
in all the computer science disciplines that we've mentioned?
link |
Yeah, I mean, there are times where problems
link |
that can be solved with well known mature methods
link |
could also be solved with a deep learning approach.
link |
And there's an argument that you must use learning
link |
even for the parts we already think we know,
link |
because if the human has touched it,
link |
then you've biased the system
link |
and you've suddenly put a bottleneck in there
link |
that is your own mental model.
link |
But something like converting a matrix,
link |
I think we know how to do that pretty well,
link |
even if it's a pretty big matrix,
link |
and we understand that pretty well.
link |
And you could train a deep network to do it,
link |
but you shouldn't probably.
link |
So in that sense, rigorous thinking is understanding
link |
the scope and the limitations of the methods that we have,
link |
like how to use the tools of mathematics properly.
link |
Yeah, I think taking a class on analysis
link |
is all I'm sort of arguing is to take a chance to stop
link |
and force yourself to think rigorously
link |
about even the rational numbers or something.
link |
It doesn't have to be the end all problem.
link |
But that exercise of clear thinking,
link |
I think goes a long way,
link |
and I just wanna make sure we keep preaching it.
link |
But do you think when you're doing rigorous thinking
link |
or maybe trying to write down equations
link |
or sort of explicitly formally describe a system,
link |
do you think we naturally simplify things too much?
link |
Is that a danger you run into?
link |
Like in order to be able to understand something
link |
about the system mathematically,
link |
we make it too much of a toy example.
link |
But I think that's the good stuff, right?
link |
That's how you understand the fundamentals?
link |
I think maybe even that's a key to intelligence
link |
or something, but I mean, okay,
link |
what if Newton and Galileo had deep learning?
link |
And they had done a bunch of experiments
link |
and they told the world,
link |
here's your weights of your neural network.
link |
We've solved the problem.
link |
Where would we be today?
link |
I don't think we'd be as far as we are.
link |
There's something to be said
link |
about having the simplest explanation for a phenomenon.
link |
So I don't doubt that we can train neural networks
link |
to predict even physical F equals MA type equations.
link |
But I maybe, I want another Newton to come along
link |
because I think there's more to do
link |
in terms of coming up with the simple models
link |
for more complicated tasks.
link |
Yeah, let's not offend AI systems from 50 years
link |
from now that are listening to this
link |
that are probably better at,
link |
might be better coming up
link |
with F equals MA equations themselves.
link |
So sorry, I actually think learning is probably a route
link |
to achieving this, but the representation matters, right?
link |
And I think having a function that takes my inputs
link |
to outputs that is arbitrarily complex
link |
may not be the end goal.
link |
I think there's still the most simple
link |
or parsimonious explanation for the data.
link |
Simple doesn't mean low dimensional.
link |
That's one thing I think that we've,
link |
a lesson that we've learned.
link |
So a standard way to do model reduction
link |
or system identification and controls
link |
is the typical formulation is that you try to find
link |
the minimal state dimension realization of a system
link |
that hits some error bounds or something like that.
link |
And that's maybe not, I think we're learning
link |
that state dimension is not the right metric.
link |
But for me, I think a lot about contact,
link |
the mechanics of contact,
link |
if a robot hand is picking up an object or something.
link |
And when I write down the equations of motion for that,
link |
they look incredibly complex,
link |
not because, actually not so much
link |
because of the dynamics of the hand when it's moving,
link |
but it's just the interactions
link |
and when they turn on and off, right?
link |
So having a high dimensional,
link |
but simple description of what's happening out here is fine.
link |
But if when I actually start touching,
link |
if I write down a different dynamical system
link |
for every polygon on my robot hand
link |
and every polygon on the object,
link |
whether it's in contact or not,
link |
with all the combinatorics that explodes there,
link |
then that's too complex.
link |
So I need to somehow summarize that
link |
with a more intuitive physics way of thinking.
link |
And yeah, I'm very optimistic
link |
that machine learning will get us there.
link |
First of all, I mean, I'll probably do it
link |
in the introduction,
link |
but you're one of the great robotics people at MIT.
link |
You're a professor at MIT.
link |
You've teach him a lot of amazing courses.
link |
You run a large group
link |
and you have a important history for MIT, I think,
link |
as being a part of the DARPA Robotics Challenge.
link |
Can you maybe first say,
link |
what is the DARPA Robotics Challenge
link |
and then tell your story around it, your journey with it?
link |
So the DARPA Robotics Challenge,
link |
it came on the tails of the DARPA Grand Challenge
link |
and DARPA Urban Challenge,
link |
which were the challenges that brought us,
link |
put a spotlight on self driving cars.
link |
Gil Pratt was at DARPA and pitched a new challenge
link |
that involved disaster response.
link |
It didn't explicitly require humanoids,
link |
although humanoids came into the picture.
link |
This happened shortly after the Fukushima disaster in Japan
link |
and our challenge was motivated roughly by that
link |
because that was a case where if we had had robots
link |
that were ready to be sent in,
link |
there's a chance that we could have averted disaster.
link |
And certainly after the, in the disaster response,
link |
there were times we would have loved
link |
to have sent robots in.
link |
So in practice, what we ended up with was a grand challenge,
link |
a DARPA Robotics Challenge,
link |
where Boston Dynamics was to make humanoid robots.
link |
People like me and the amazing team at MIT
link |
were competing first in a simulation challenge
link |
to try to be one of the ones that wins the right
link |
to work on one of the Boston Dynamics humanoids
link |
in order to compete in the final challenge,
link |
which was a physical challenge.
link |
And at that point, it was already, so it was decided
link |
as humanoid robots early on.
link |
There were two tracks.
link |
You could enter as a hardware team
link |
where you brought your own robot,
link |
or you could enter through the virtual robotics challenge
link |
as a software team that would try to win the right
link |
to use one of the Boston Dynamics robots.
link |
Sure, called Atlas.
link |
Yeah, it was a 400 pound Marvel,
link |
but a pretty big, scary looking robot.
link |
Okay, so I mean, how did you feel
link |
at the prospect of this kind of challenge?
link |
I mean, it seems autonomous vehicles,
link |
yeah, I guess that sounds hard,
link |
but not really from a robotics perspective.
link |
It's like, didn't they do it in the 80s
link |
is the kind of feeling I would have,
link |
like when you first look at the problem,
link |
it's on wheels, but like humanoid robots,
link |
that sounds really hard.
link |
So what are your, psychologically speaking,
link |
what were you feeling, excited, scared?
link |
Why the heck did you get yourself involved
link |
in this kind of messy challenge?
link |
We didn't really know for sure what we were signing up for
link |
in the sense that you could have something that,
link |
as it was described in the call for participation,
link |
that could have put a huge emphasis on the dynamics
link |
of walking and not falling down
link |
and walking over rough terrain,
link |
or the same description,
link |
because the robot had to go into this disaster area
link |
and turn valves and pick up a drill,
link |
it cut the hole through a wall,
link |
it had to do some interesting things.
link |
The challenge could have really highlighted perception
link |
and autonomous planning,
link |
or it ended up that locomoting over complex terrain
link |
played a pretty big role in the competition.
link |
And the degree of autonomy wasn't clear.
link |
The degree of autonomy
link |
was always a central part of the discussion.
link |
So what wasn't clear was how we would be able,
link |
how far we'd be able to get with it.
link |
So the idea was always that you want semi autonomy,
link |
that you want the robot to have enough compute
link |
that you can have a degraded network link to a human.
link |
And so the same way we had degraded networks
link |
at many natural disasters,
link |
you'd send your robot in,
link |
you'd be able to get a few bits back and forth,
link |
but you don't get to have enough
link |
potentially to fully operate the robot
link |
in every joint of the robot.
link |
So, and then the question was,
link |
and the gamesmanship of the organizers
link |
was to figure out what we're capable of,
link |
push us as far as we could,
link |
so that it would differentiate the teams
link |
that put more autonomy on the robot
link |
and had a few clicks and just said,
link |
go there, do this, go there, do this,
link |
versus someone who's picking every footstep
link |
or something like that.
link |
So what were some memories,
link |
painful, triumphant from the experience?
link |
Like what was that journey?
link |
Maybe if you can dig in a little deeper,
link |
maybe even on the technical side, on the team side,
link |
that whole process of,
link |
from the early idea stages to actually competing.
link |
I mean, this was a defining experience for me.
link |
It came at the right time for me in my career.
link |
I had gotten tenure before I was due a sabbatical,
link |
and most people do something relaxing
link |
and restorative for a sabbatical.
link |
So you got tenure before this?
link |
It was a good time for me.
link |
We had a bunch of algorithms that we were very happy with.
link |
We wanted to see how far we could push them,
link |
and this was a chance to really test our mettle
link |
to do more proper software engineering.
link |
So the team, we all just worked our butts off.
link |
We were in that lab almost all the time.
link |
Okay, so there were some, of course,
link |
high highs and low lows throughout that.
link |
Anytime you're not sleeping
link |
and devoting your life to a 400 pound humanoid.
link |
I remember actually one funny moment
link |
where we're all super tired,
link |
and so Atlas had to walk across cinder blocks.
link |
That was one of the obstacles.
link |
And I remember Atlas was powered down
link |
and hanging limp on its harness,
link |
and the humans were there picking up
link |
and laying the brick down
link |
so that the robot could walk over it.
link |
And I thought, what is wrong with this?
link |
We've got a robot just watching us
link |
do all the manual labor
link |
so that it can take its little stroll across the train.
link |
But I mean, even the virtual robotics challenge
link |
was super nerve wracking and dramatic.
link |
I remember, so we were using Gazebo as a simulator
link |
and there was all these interesting challenges.
link |
I think the investment that OSR FC,
link |
whatever they were called at that time,
link |
Brian Gerkey's team at Open Source Robotics,
link |
they were pushing on the capabilities of Gazebo
link |
in order to scale it to the complexity of these challenges.
link |
So, you know, up to the virtual competition.
link |
So the virtual competition was,
link |
you will sign on at a certain time
link |
and we'll have a network connection
link |
to another machine on the cloud
link |
that is running the simulator of your robot.
link |
And your controller will run on this computer
link |
and the physics will run on the other
link |
and you have to connect.
link |
Now, the physics, they wanted it to run at real time rates
link |
because there was an element of human interaction.
link |
And humans, if you do want to teleop,
link |
it works way better if it's at frame rate.
link |
But it was very hard to simulate
link |
these complex scenes at real time rate.
link |
So right up to like days before the competition,
link |
the simulator wasn't quite at real time rate.
link |
And that was great for me because my controller
link |
was solving a pretty big optimization problem
link |
and it wasn't quite at real time rate.
link |
I was keeping up with the simulator.
link |
We were both running at about 0.7.
link |
And I remember getting this email.
link |
And by the way, the perception folks on our team hated
link |
that they knew that if my controller was too slow,
link |
the robot was gonna fall down.
link |
And no matter how good their perception system was,
link |
if I can't make my controller fast.
link |
Anyways, we get this email
link |
like three days before the virtual competition.
link |
It's for all the marbles.
link |
We're gonna either get a humanoid robot or we're not.
link |
And we get an email saying,
link |
good news, we made the robot, the simulator faster.
link |
It's now at one point.
link |
And I was just like, oh man, what are we gonna do here?
link |
So that came in late at night for me.
link |
I went over, it happened at Frank Permenter,
link |
who's a very, very sharp.
link |
He was a student at the time working on optimization.
link |
He was still in lab.
link |
Frank, we need to make the quadratic programming solver
link |
faster, not like a little faster.
link |
It's actually, you know, and we wrote a new solver
link |
for that QP together that night.
link |
It was terrifying.
link |
So there's a really hard optimization problem
link |
that you're constantly solving.
link |
You didn't make the optimization problem simpler?
link |
You wrote a new solver?
link |
So, I mean, your observation is almost spot on.
link |
What we did was what everybody,
link |
I mean, people know how to do this,
link |
but we had not yet done this idea of warm starting.
link |
So we are solving a big optimization problem
link |
at every time step.
link |
But if you're running fast enough,
link |
the optimization problem you're solving
link |
on the last time step is pretty similar
link |
to the optimization you're gonna solve with the next.
link |
We had course had told our commercial solver
link |
to use warm starting, but even the interface
link |
to that commercial solver was causing us these delays.
link |
So what we did was we basically wrote,
link |
we called it fast QP at the time.
link |
We wrote a very lightweight, very fast layer,
link |
which would basically check if nearby solutions
link |
to the quadratic program were,
link |
which were very easily checked,
link |
could stabilize the robot.
link |
And if they couldn't, we would fall back to the solver.
link |
You couldn't really test this well, right?
link |
I mean, so we always knew that if we fell back to,
link |
if we, it got to the point where if for some reason
link |
things slowed down and we fell back to the original solver,
link |
the robot would actually literally fall down.
link |
So it was a harrowing sort of edge we were,
link |
ledge we were sort of on.
link |
But I mean, it actually,
link |
like the 400 pound human could come crashing to the ground
link |
if your solver's not fast enough.
link |
But you know, we had lots of good experiences.
link |
So can I ask you a weird question I get
link |
about idea of hard work?
link |
So actually people, like students of yours
link |
that I've interacted with and just,
link |
and robotics people in general,
link |
but they have moments,
link |
at moments have worked harder than most people I know
link |
in terms of, if you look at different disciplines
link |
of how hard people work.
link |
But they're also like the happiest.
link |
Like, just like, I don't know.
link |
It's the same thing with like running.
link |
People that push themselves to like the limit,
link |
they also seem to be like the most like full of life
link |
And I get often criticized like,
link |
you're not getting enough sleep.
link |
What are you doing to your body?
link |
Blah, blah, blah, like this kind of stuff.
link |
And I usually just kind of respond like,
link |
I'm doing what I love.
link |
I'm passionate about it.
link |
I feel like it's, it's invigorating.
link |
I actually think, I don't think the lack of sleep
link |
is what hurts you.
link |
I think what hurts you is stress and lack of doing things
link |
that you're passionate about.
link |
But in this world, yeah, I mean,
link |
can you comment about why the heck robotics people
link |
are willing to push themselves to that degree?
link |
Is there value in that?
link |
And why are they so happy?
link |
I think, I think you got it right.
link |
I mean, I think the causality is not that we work hard.
link |
And I think other disciplines work very hard too,
link |
but it's, I don't think it's that we work hard
link |
and therefore we are happy.
link |
I think we found something
link |
that we're truly passionate about.
link |
It makes us very happy.
link |
And then we get a little involved with it
link |
and spend a lot of time on it.
link |
What a luxury to have something
link |
that you wanna spend all your time on, right?
link |
We could talk about this for many hours,
link |
but maybe if we could pick,
link |
is there something on the technical side
link |
on the approach that you took that's interesting
link |
that turned out to be a terrible failure
link |
or a success that you carry into your work today
link |
about all the different ideas that were involved
link |
in making, whether in the simulation or in the real world,
link |
making this semi autonomous system work?
link |
I mean, it really did teach me something fundamental
link |
about what it's gonna take to get robustness
link |
out of a system of this complexity.
link |
I would say the DARPA challenge
link |
really was foundational in my thinking.
link |
I think the autonomous driving community thinks about this.
link |
I think lots of people thinking
link |
about safety critical systems
link |
that might have machine learning in the loop
link |
are thinking about these questions.
link |
For me, the DARPA challenge was the moment
link |
where I realized we've spent every waking minute
link |
running this robot.
link |
And again, for the physical competition,
link |
days before the competition,
link |
we saw the robot fall down in a way
link |
it had never fallen down before.
link |
I thought, how could we have found that?
link |
We only have one robot, it's running almost all the time.
link |
We just didn't have enough hours in the day
link |
to test that robot.
link |
Something has to change, right?
link |
And then I think that, I mean,
link |
I would say that the team that won was,
link |
from KAIST, was the team that had two robots
link |
and was able to do not only incredible engineering,
link |
just absolutely top rate engineering,
link |
but also they were able to test at a rate
link |
and discipline that we didn't keep up with.
link |
What does testing look like?
link |
What are we talking about here?
link |
Like, what's a loop of tests?
link |
Like from start to finish, what is a loop of testing?
link |
Yeah, I mean, I think there's a whole philosophy to testing.
link |
There's the unit tests, and you can do that on a hardware,
link |
you can do that in a small piece of code.
link |
You write one function, you should write a test
link |
that checks that function's input and outputs.
link |
You should also write an integration test
link |
at the other extreme of running the whole system together,
link |
where they try to turn on all of the different functions
link |
that you think are correct.
link |
It's much harder to write the specifications
link |
for a system level test,
link |
especially if that system is as complicated
link |
as a humanoid robot.
link |
But the philosophy is sort of the same.
link |
On the real robot, it's no different,
link |
but on a real robot,
link |
it's impossible to run the same experiment twice.
link |
So if you see a failure,
link |
you hope you caught something in the logs
link |
that tell you what happened,
link |
but you'd probably never be able to run
link |
exactly that experiment again.
link |
And right now, I think our philosophy is just,
link |
basically Monte Carlo estimation,
link |
is just run as many experiments as we can,
link |
maybe try to set up the environment
link |
to make the things we are worried about happen
link |
as often as possible.
link |
But really we're relying on somewhat random search
link |
Maybe that's all we'll ever be able to,
link |
but I think, you know,
link |
cause there's an argument that the things that'll get you
link |
are the things that are really nuanced in the world.
link |
And there'd be very hard to, for instance,
link |
put back in a simulation.
link |
Yeah, I guess the edge cases.
link |
What was the hardest thing?
link |
Like, so you said walking over rough terrain,
link |
like just taking footsteps.
link |
I mean, people, it's so dramatic and painful
link |
in a certain kind of way to watch these videos
link |
from the DRC of robots falling.
link |
It's just so heartbreaking.
link |
Maybe it's because for me at least,
link |
we anthropomorphize the robot.
link |
Of course, it's also funny for some reason,
link |
like humans falling is funny for, I don't,
link |
it's some dark reason.
link |
I'm not sure why it is so,
link |
but it's also like tragic and painful.
link |
And so speaking of which, I mean,
link |
what made the robots fall and fail in your view?
link |
So I can tell you exactly what happened on our,
link |
we, I contributed one of those.
link |
Our team contributed one of those spectacular falls.
link |
Every one of those falls has a complicated story.
link |
I mean, at one time,
link |
the power effectively went out on the robot
link |
because it had been sitting at the door
link |
waiting for a green light to be able to proceed
link |
and its batteries, you know,
link |
and therefore it just fell backwards
link |
and smashed its head against the ground.
link |
And it was hilarious,
link |
but it wasn't because of bad software, right?
link |
But for ours, so the hardest part of the challenge,
link |
the hardest task in my view was getting out of the Polaris.
link |
It was actually relatively easy to drive the Polaris.
link |
Can you tell the story?
link |
Sorry to interrupt.
link |
The story of the car.
link |
People should watch this video.
link |
I mean, the thing you've come up with is just brilliant,
link |
but anyway, sorry, what's...
link |
Yeah, we kind of joke.
link |
We call it the big robot, little car problem
link |
because somehow the race organizers decided
link |
to give us a 400 pound humanoid.
link |
And then they also provided the vehicle,
link |
which was a little Polaris.
link |
And the robot didn't really fit in the car.
link |
So you couldn't drive the car with your feet
link |
under the steering column.
link |
We actually had to straddle the main column of the,
link |
and have basically one foot in the passenger seat,
link |
one foot in the driver's seat,
link |
and then drive with our left hand.
link |
But the hard part was we had to then park the car,
link |
get out of the car.
link |
It didn't have a door, that was okay.
link |
But it's just getting up from crouched, from sitting,
link |
when you're in this very constrained environment.
link |
First of all, I remember after watching those videos,
link |
I was much more cognizant of how hard it is for me
link |
to get in and out of the car,
link |
and out of the car, especially.
link |
It's actually a really difficult control problem.
link |
I'm very cognizant of it when I'm like injured
link |
for whatever reason.
link |
Oh, that's really hard.
link |
So how did you approach this problem?
link |
So we had, you think of NASA's operations,
link |
and they have these checklists,
link |
prelaunched checklists and the like.
link |
We weren't far off from that.
link |
We had this big checklist.
link |
And on the first day of the competition,
link |
we were running down our checklist.
link |
And one of the things we had to do,
link |
we had to turn off the controller,
link |
the piece of software that was running
link |
that would drive the left foot of the robot
link |
in order to accelerate on the gas.
link |
And then we turned on our balancing controller.
link |
And the nerves, jitters of the first day of the competition,
link |
someone forgot to check that box
link |
and turn that controller off.
link |
So we used a lot of motion planning
link |
to figure out a sort of configuration of the robot
link |
that we could get up and over.
link |
We relied heavily on our balancing controller.
link |
And basically, when the robot was in one
link |
of its most precarious sort of configurations,
link |
trying to sneak its big leg out of the side,
link |
the other controller that thought it was still driving
link |
told its left foot to go like this.
link |
And that wasn't good.
link |
But it turned disastrous for us
link |
because what happened was a little bit of push here.
link |
Actually, we have videos of us running into the robot
link |
with a 10 foot pole and it kind of will recover.
link |
But this is a case where there's no space to recover.
link |
So a lot of our secondary balancing mechanisms
link |
about like take a step to recover,
link |
they were all disabled because we were in the car
link |
and there was no place to step.
link |
So we were relying on our just lowest level reflexes.
link |
And even then, I think just hitting the foot on the seat,
link |
on the floor, we probably could have recovered from it.
link |
But the thing that was bad that happened
link |
is when we did that and we jostled a little bit,
link |
the tailbone of our robot was only a little off the seat,
link |
And the other foot came off the ground just a little bit.
link |
And nothing in our plans had ever told us what to do
link |
if your butt's on the seat and your feet are in the air.
link |
And then the thing is once you get off the script,
link |
things can go very wrong
link |
because even our state estimation,
link |
our system that was trying to collect all the data
link |
from the sensors and understand
link |
what's happening with the robot,
link |
it didn't know about this situation.
link |
So it was predicting things that were just wrong.
link |
And then we did a violent shake and fell off
link |
in our face first out of the robot.
link |
But like into the destination.
link |
That's true, we fell in, we got our point for egress.
link |
But so is there any hope for, that's interesting,
link |
is there any hope for Atlas to be able to do something
link |
when it's just on its butt and feet in the air?
link |
So you can, what do you?
link |
No, so that is one of the big challenges.
link |
And I think it's still true, you know,
link |
Boston Dynamics and Antimal and there's this incredible work
link |
on legged robots happening around the world.
link |
Most of them still are very good at the case
link |
where you're making contact with the world at your feet.
link |
And they have typically point feet relatively,
link |
they have balls on their feet, for instance.
link |
If those robots get in a situation
link |
where the elbow hits the wall or something like this,
link |
that's a pretty different situation.
link |
Now they have layers of mechanisms that will make,
link |
I think the more mature solutions have ways
link |
in which the controller won't do stupid things.
link |
But a human, for instance, is able to leverage
link |
incidental contact in order to accomplish a goal.
link |
In fact, I might, if you push me,
link |
I might actually put my hand out
link |
and make a new brand new contact.
link |
The feet of the robot are doing this on quadrupeds,
link |
but we mostly in robotics are afraid of contact
link |
on the rest of our body, which is crazy.
link |
There's this whole field of motion planning,
link |
collision free motion planning.
link |
And we write very complex algorithms
link |
so that the robot can dance around
link |
and make sure it doesn't touch the world.
link |
So people are just afraid of contact
link |
because contact the scene is a difficult.
link |
It's still a difficult control problem and sensing problem.
link |
Now you're a serious person, I'm a little bit of an idiot
link |
and I'm going to ask you some dumb questions.
link |
So I do martial arts.
link |
So like jiu jitsu, I wrestled my whole life.
link |
So let me ask the question, like whenever people learn
link |
that I do any kind of AI or like I mentioned robots
link |
and things like that, they say,
link |
when are we going to have robots that can win
link |
in a wrestling match or in a fight against a human?
link |
So we just mentioned sitting on your butt,
link |
if you're in the air, that's a common position.
link |
Jiu jitsu, when you're on the ground,
link |
you're a down opponent.
link |
Like how difficult do you think is the problem?
link |
And when will we have a robot that can defeat a human
link |
in a wrestling match?
link |
And we're talking about a lot, like, I don't know
link |
if you're familiar with wrestling, but essentially.
link |
It's basically the art of contact.
link |
It's like, it's because you're picking contact points
link |
and then using like leverage like to off balance
link |
to trick people, like you make them feel
link |
like you're doing one thing
link |
and then they change their balance
link |
and then you switch what you're doing
link |
and then results in a throw or whatever.
link |
So like, it's basically the art of multiple contacts.
link |
Awesome, that's a nice description of it.
link |
So there's also an opponent in there, right?
link |
Right, if you are wrestling a human
link |
and are in a game theoretic situation with a human,
link |
that's still hard, but just to speak to the, you know,
link |
quickly reasoning about contact part of it, for instance.
link |
Yeah, maybe even throwing the game theory out of it,
link |
almost like, yeah, almost like a non dynamic opponent.
link |
Right, there's reasons to be optimistic,
link |
but I think our best understanding of those problems
link |
are still pretty hard.
link |
I have been increasingly focused on manipulation,
link |
partly where that's a case where the contact
link |
has to be much more rich.
link |
And there are some really impressive examples
link |
of deep learning policies, controllers
link |
that can appear to do good things through contact.
link |
We've even got new examples of, you know,
link |
deep learning models of predicting what's gonna happen
link |
to objects as they go through contact.
link |
But I think the challenge you just offered there
link |
still eludes us, right?
link |
The ability to make a decision
link |
based on those models quickly.
link |
You know, I have to think though, it's hard for humans too,
link |
when you get that complicated.
link |
I think probably you had maybe a slow motion version
link |
of where you learned the basic skills
link |
and you've probably gotten better at it
link |
and there's much more subtle to you.
link |
But it might still be hard to actually, you know,
link |
really on the fly take a, you know, model of your humanoid
link |
and figure out how to plan the optimal sequence.
link |
That might be a problem we never solve.
link |
Well, the, I mean, one of the most amazing things to me
link |
about the, we can talk about martial arts.
link |
We could also talk about dancing.
link |
Doesn't really matter.
link |
Too human, I think it's the most interesting study
link |
It's not even the dynamic element of it.
link |
It's the, like when you get good at it, it's so effortless.
link |
Like I can just, I'm very cognizant
link |
of the entirety of the learning process
link |
being essentially like learning how to move my body
link |
in a way that I could throw very large weights
link |
around effortlessly, like, and I can feel the learning.
link |
Like I'm a huge believer in drilling of techniques
link |
and you can just like feel your, I don't,
link |
you're not feeling, you're feeling, sorry,
link |
you're learning it intellectually a little bit,
link |
but a lot of it is the body learning it somehow,
link |
like instinctually and whatever that learning is,
link |
that's really, I'm not even sure if that's equivalent
link |
to like a deep learning, learning a controller.
link |
I think it's something more,
link |
it feels like there's a lot of distributed learning
link |
Yeah, I think there's hierarchy and composition
link |
probably in the systems that we don't capture very well yet.
link |
You have layers of control systems.
link |
You have reflexes at the bottom layer
link |
and you have a system that's capable
link |
of planning a vacation to some distant country,
link |
which is probably, you probably don't have a controller,
link |
a policy for every possible destination you'll ever pick.
link |
But there's something magical in the in between
link |
and how do you go from these low level feedback loops
link |
to something that feels like a pretty complex
link |
You know, my guess is, I think there's evidence
link |
that you can plan at some of these levels, right?
link |
So Josh Tenenbaum just showed it in his talk the other day.
link |
He's got a game he likes to talk about.
link |
I think he calls it the pick three game or something,
link |
where he puts a bunch of clutter down in front of a person
link |
and he says, okay, pick three objects.
link |
And it might be a telephone or a shoe
link |
or a Kleenex box or whatever.
link |
And apparently you pick three items and then you pick,
link |
he says, okay, pick the first one up with your right hand,
link |
the second one up with your left hand.
link |
Now using those objects, now as tools,
link |
pick up the third object.
link |
Right, so that's down at the level of physics
link |
and mechanics and contact mechanics
link |
that I think we do learning or we do have policies for,
link |
we do control for, almost feedback,
link |
but somehow we're able to still,
link |
I mean, I've never picked up a telephone
link |
with a shoe and a water bottle before.
link |
And somehow, and it takes me a little longer to do that
link |
the first time, but most of the time
link |
we can sort of figure that out.
link |
So yeah, I think the amazing thing is this ability
link |
to be flexible with our models,
link |
plan when we need to use our well oiled controllers
link |
when we don't, when we're in familiar territory.
link |
Having models, I think the other thing you just said
link |
was something about, I think your awareness
link |
of what's happening is even changing
link |
as you improve your expertise, right?
link |
So maybe you have a very approximate model
link |
of the mechanics to begin with.
link |
And as you gain expertise,
link |
you get a more refined version of that model.
link |
You're aware of muscles or balance components
link |
that you just weren't even aware of before.
link |
So how do you scaffold that?
link |
Yeah, plus the fear of injury,
link |
the ambition of goals, of excelling,
link |
and fear of mortality.
link |
Let's see, what else is in there?
link |
As the motivations, overinflated ego in the beginning,
link |
and then a crash of confidence in the middle.
link |
All of those seem to be essential for the learning process.
link |
And if all that's good,
link |
then you're probably optimizing energy efficiency.
link |
Yeah, right, so we have to get that right.
link |
So there was this idea that you would have robots
link |
play soccer better than human players by 2050.
link |
That was the goal.
link |
Basically, it was the goal to beat world champion team,
link |
to become a world cup, beat like a world cup level team.
link |
So are we gonna see that first?
link |
Or a robot, if you're familiar,
link |
there's an organization called UFC for mixed martial arts.
link |
Are we gonna see a world cup championship soccer team
link |
that have robots, or a UFC champion mixed martial artist
link |
I mean, it's very hard to say one thing is harder,
link |
some problem is harder than the other.
link |
What probably matters is who started the organization that,
link |
I mean, I think RoboCup has a pretty serious following,
link |
and there is a history now of people playing that game,
link |
learning about that game, building robots to play that game,
link |
building increasingly more human robots.
link |
It's got momentum.
link |
So if you want to have mixed martial arts compete,
link |
you better start your organization now, right?
link |
I think almost independent of which problem
link |
is technically harder,
link |
because they're both hard and they're both different.
link |
That's a good point.
link |
I mean, those videos are just hilarious,
link |
like especially the humanoid robots
link |
trying to play soccer.
link |
I mean, they're kind of terrible right now.
link |
I mean, I guess there is robo sumo wrestling.
link |
There's like the robo one competitions,
link |
where they do have these robots that go on the table
link |
and basically fight.
link |
So maybe I'm wrong, maybe.
link |
First of all, do you have a year in mind for RoboCup,
link |
just from a robotics perspective?
link |
Seems like a super exciting possibility
link |
that like in the physical space,
link |
this is what's interesting.
link |
I think the world is captivated.
link |
I think it's really exciting.
link |
It inspires just a huge number of people
link |
when a machine beats a human at a game
link |
that humans are really damn good at.
link |
So you're talking about chess and go,
link |
but that's in the world of digital.
link |
I don't think machines have beat humans
link |
at a game in the physical space yet,
link |
but that would be just.
link |
You have to make the rules very carefully, right?
link |
I mean, if Atlas kicked me in the shins, I'm down
link |
So it's very subtle on what's fair.
link |
I think the fighting one is a weird one.
link |
Yeah, because you're talking about a machine
link |
that's much stronger than you.
link |
But yeah, in terms of soccer, basketball, all those kinds.
link |
Even soccer, right?
link |
I mean, as soon as there's contact or whatever,
link |
and there are some things that the robot will do better.
link |
I think if you really set yourself up to try to see
link |
could robots win the game of soccer
link |
as the rules were written, the right thing
link |
for the robot to do is to play very differently
link |
than a human would play.
link |
You're not gonna get the perfect soccer player robot.
link |
You're gonna get something that exploits the rules,
link |
exploits its super actuators, its super low bandwidth
link |
feedback loops or whatever, and it's gonna play the game
link |
differently than you want it to play.
link |
And I bet there's ways, I bet there's loopholes, right?
link |
We saw that in the DARPA challenge that it's very hard
link |
to write a set of rules that someone can't find
link |
Let me ask another ridiculous question.
link |
I think this might be the last ridiculous question,
link |
I aspire to ask as many ridiculous questions
link |
of a brilliant MIT professor.
link |
Okay, I don't know if you've seen the black mirror.
link |
It's funny, I never watched the episode.
link |
I know when it happened though, because I gave a talk
link |
to some MIT faculty one day on a unassuming Monday
link |
or whatever I was telling him about the state of robotics.
link |
And I showed some video from Boston Dynamics
link |
of the quadruped spot at the time.
link |
It was the early version of spot.
link |
And there was a look of horror that went across the room.
link |
And I said, I've shown videos like this a lot of times,
link |
And it turns out that this video had gone,
link |
this black mirror episode had changed
link |
the way people watched the videos I was putting out.
link |
The way they see these kinds of robots.
link |
So I talked to so many people who are just terrified
link |
because of that episode probably of these kinds of robots.
link |
I almost wanna say that they almost enjoy being terrified.
link |
I don't even know what it is about human psychology
link |
that kind of imagine doomsday,
link |
the destruction of the universe or our society
link |
and kind of like enjoy being afraid.
link |
I don't wanna simplify it, but it feels like
link |
they talk about it so often.
link |
It almost, there does seem to be an addictive quality to it.
link |
I talked to a guy, a guy named Joe Rogan,
link |
who's kind of the flag bearer
link |
for being terrified at these robots.
link |
Do you have two questions?
link |
One, do you have an understanding
link |
of why people are afraid of robots?
link |
And the second question is in black mirror,
link |
just to tell you the episode,
link |
I don't even remember it that much anymore,
link |
but these robots, I think they can shoot
link |
like a pellet or something.
link |
They basically have, it's basically a spot with a gun.
link |
And how far are we away from having robots
link |
that go rogue like that?
link |
Basically spot that goes rogue for some reason
link |
and somehow finds a gun.
link |
Right, so, I mean, I'm not a psychologist.
link |
I think, I don't know exactly why
link |
people react the way they do.
link |
I think we have to be careful about the way robots influence
link |
our society and the like.
link |
I think that's something, that's a responsibility
link |
that roboticists need to embrace.
link |
I don't think robots are gonna come after me
link |
with a kitchen knife or a pellet gun right away.
link |
And I mean, if they were programmed in such a way,
link |
but I used to joke with Atlas that all I had to do
link |
was run for five minutes and its battery would run out.
link |
But actually they've got to be careful
link |
and actually they've got a very big battery
link |
in there by the end.
link |
So it was over an hour.
link |
I think the fear is a bit cultural though.
link |
Cause I mean, you notice that, like, I think in my age,
link |
in the US, we grew up watching Terminator, right?
link |
If I had grown up at the same time in Japan,
link |
I probably would have been watching Astro Boy.
link |
And there's a very different reaction to robots
link |
in different countries, right?
link |
So I don't know if it's a human innate fear of metal marvels
link |
or if it's something that we've done to ourselves
link |
Yeah, the stories we tell ourselves through movies,
link |
through just through popular media.
link |
But if I were to tell, you know, if you were my therapist
link |
and I said, I'm really terrified that we're going
link |
to have these robots very soon that will hurt us.
link |
Like, how do you approach making me feel better?
link |
Like, why shouldn't people be afraid?
link |
There's a, I think there's a video
link |
that went viral recently.
link |
Everything, everything was spot in Boston,
link |
which goes viral in general.
link |
But usually it's like really cool stuff.
link |
Like they're doing flips and stuff
link |
or like sad stuff, the Atlas being hit with a broomstick
link |
or something like that.
link |
But there's a video where I think one of the new productions
link |
bought robots, which are awesome.
link |
It was like patrolling somewhere in like in some country.
link |
And like people immediately were like saying like,
link |
this is like the dystopian future,
link |
like the surveillance state.
link |
For some reason, like you can just have a camera,
link |
like something about spot being able to walk on four feet
link |
with like really terrified people.
link |
So like, what do you say to those people?
link |
I think there is a legitimate fear there
link |
because so much of our future is uncertain.
link |
But at the same time, technically speaking,
link |
it seems like we're not there yet.
link |
So what do you say?
link |
I mean, I think technology is complicated.
link |
It can be used in many ways.
link |
I think there are purely software attacks
link |
that somebody could use to do great damage.
link |
Maybe they have already, you know,
link |
I think wheeled robots could be used in bad ways too.
link |
Drones, right, I don't think that, let's see.
link |
I don't want to be building technology
link |
just because I'm compelled to build technology
link |
and I don't think about it.
link |
But I would consider myself a technological optimist,
link |
I guess, in the sense that I think we should continue
link |
to create and evolve and our world will change.
link |
And if we will introduce new challenges,
link |
we'll screw something up maybe,
link |
but I think also we'll invent ourselves
link |
out of those challenges and life will go on.
link |
So it's interesting because you didn't mention
link |
like this is technically too hard.
link |
I don't think robots are, I think people attribute
link |
a robot that looks like an animal
link |
as maybe having a level of self awareness
link |
or consciousness or something that they don't have yet.
link |
Right, so it's not, I think our ability
link |
to anthropomorphize those robots is probably,
link |
we're assuming that they have a level of intelligence
link |
that they don't yet have.
link |
And that might be part of the fear.
link |
So in that sense, it's too hard.
link |
But, you know, there are many scary things in the world.
link |
Right, so I think we're right to ask those questions.
link |
We're right to think about the implications of our work.
link |
Right, in the short term as we're working on it for sure,
link |
is there something long term that scares you
link |
about our future with AI and robots?
link |
A lot of folks from Elon Musk to Sam Harris
link |
to a lot of folks talk about the existential threats
link |
about artificial intelligence.
link |
Oftentimes, robots kind of inspire that the most
link |
because of the anthropomorphism.
link |
Do you have any fears?
link |
It's an important question.
link |
I actually, I think I like Rod Brooks answer
link |
maybe the best on this, I think.
link |
And it's not the only answer he's given over the years,
link |
but maybe one of my favorites is he says,
link |
it's not gonna be, he's got a book,
link |
Flesh and Machines, I believe, it's not gonna be
link |
the robots versus the people,
link |
we're all gonna be robot people.
link |
Because, you know, we already have smartphones,
link |
some of us have serious technology implanted
link |
in our bodies already, whether we have a hearing aid
link |
or a pacemaker or anything like this,
link |
people with amputations might have prosthetics.
link |
And that's a trend I think that is likely to continue.
link |
I mean, this is now wild speculation.
link |
But I mean, when do we get to cognitive implants
link |
and the like, and.
link |
Yeah, with neural link, brain computer interfaces,
link |
that's interesting.
link |
So there's a dance between humans and robots
link |
that's going to be, it's going to be impossible
link |
to be scared of the other out there, the robot,
link |
because the robot will be part of us, essentially.
link |
It'd be so intricately sort of part of our society that.
link |
Yeah, and it might not even be implanted part of us,
link |
but just, it's so much a part of our, yeah, our society.
link |
So in that sense, the smartphone is already the robot
link |
we should be afraid of, yeah.
link |
I mean, yeah, and all the usual fears arise
link |
of the misinformation, the manipulation,
link |
all those kinds of things that,
link |
the problems are all the same.
link |
They're human problems, essentially, it feels like.
link |
Yeah, I mean, I think the way we interact
link |
with each other online is changing the value we put on,
link |
you know, personal interaction.
link |
And that's a crazy big change that's going to happen
link |
and rip through our, has already been ripping
link |
through our society, right?
link |
And that has implications that are massive.
link |
I don't know if they should be scared of it
link |
or go with the flow, but I don't see, you know,
link |
some battle lines between humans and robots
link |
being the first thing to worry about.
link |
I mean, I do want to just, as a kind of comment,
link |
maybe you can comment about your just feelings
link |
about Boston Dynamics in general, but you know,
link |
I love science, I love engineering,
link |
I think there's so many beautiful ideas in it.
link |
And when I look at Boston Dynamics
link |
or legged robots in general,
link |
I think they inspire people, curiosity and feelings
link |
in general, excitement about engineering
link |
more than almost anything else in popular culture.
link |
And I think that's such an exciting,
link |
like responsibility and possibility for robotics.
link |
And Boston Dynamics is riding that wave pretty damn well.
link |
Like they found it, they've discovered that hunger
link |
and curiosity in the people and they're doing magic with it.
link |
I don't care if the, I mean, I guess is that their company,
link |
they have to make money, right?
link |
But they're already doing incredible work
link |
and inspiring the world about technology.
link |
I mean, do you have thoughts about Boston Dynamics
link |
and maybe others, your own work in robotics
link |
and inspiring the world in that way?
link |
I completely agree, I think Boston Dynamics
link |
is absolutely awesome.
link |
I think I show my kids those videos, you know,
link |
and the best thing that happens is sometimes
link |
they've already seen them, you know, right?
link |
I think, I just think it's a pinnacle of success
link |
in robotics that is just one of the best things
link |
that's happened, absolutely completely agree.
link |
One of the heartbreaking things to me is how many
link |
robotics companies fail, how hard it is to make money
link |
with a robotics company.
link |
Like iRobot like went through hell just to arrive
link |
at a Roomba to figure out one product.
link |
And then there's so many home robotics companies
link |
like Jibo and Anki, Anki, the cutest toy that's a great robot
link |
I thought went down, I'm forgetting a bunch of them,
link |
but a bunch of robotics companies fail,
link |
Rod's company, Rethink Robotics.
link |
Like, do you have anything hopeful to say
link |
about the possibility of making money with robots?
link |
Oh, I think you can't just look at the failures.
link |
I mean, Boston Dynamics is a success.
link |
There's lots of companies that are still doing amazingly
link |
good work in robotics.
link |
I mean, this is the capitalist ecology or something, right?
link |
I think you have many companies, you have many startups
link |
and they push each other forward and many of them fail
link |
and some of them get through and that's sort of
link |
the natural way of those things.
link |
I don't know that is robotics really that much worse.
link |
I feel the pain that you feel too.
link |
Every time I read one of these, sometimes it's friends
link |
and I definitely wish it went better or went differently.
link |
But I think it's healthy and good to have bursts of ideas,
link |
bursts of activities, ideas, if they are really aggressive,
link |
they should fail sometimes.
link |
Certainly that's the research mantra, right?
link |
If you're succeeding at every problem you attempt,
link |
then you're not choosing aggressively enough.
link |
Is it exciting to you, the new spot?
link |
When are you getting them as a pet or it?
link |
Yeah, I mean, I have to dig up 75K right now.
link |
I mean, it's so cool that there's a price tag,
link |
you can go and then actually buy it.
link |
I have a Skydio R1, love it.
link |
So no, I would absolutely be a customer.
link |
I wonder what your kids would think about it.
link |
I actually, Zach from Boston Dynamics would let my kid drive
link |
in one of their demos one time.
link |
And that was just so good, so good.
link |
And again, I'll forever be grateful for that.
link |
And there's something magical about the anthropomorphization
link |
of that arm, it adds another level of human connection.
link |
I'm not sure we understand from a control aspect,
link |
the value of anthropomorphization.
link |
I think that's an understudied
link |
and under understood engineering problem.
link |
There's been a, like psychologists have been studying it.
link |
I think it's part like manipulating our mind
link |
to believe things is a valuable engineering.
link |
Like this is another degree of freedom
link |
that can be controlled.
link |
I like that, yeah, I think that's right.
link |
I think there's something that humans seem to do
link |
or maybe my dangerous introspection is,
link |
I think we are able to make very simple models
link |
that assume a lot about the world very quickly.
link |
And then it takes us a lot more time, like you're wrestling.
link |
You probably thought you knew what you were doing
link |
with wrestling and you were fairly functional
link |
as a complete wrestler.
link |
And then you slowly got more expertise.
link |
So maybe it's natural that our first level of defense
link |
against seeing a new robot is to think of it
link |
in our existing models of how humans and animals behave.
link |
And it's just, as you spend more time with it,
link |
then you'll develop more sophisticated models
link |
that will appreciate the differences.
link |
Can you say what does it take to control a robot?
link |
Like what is the control problem of a robot?
link |
And in general, what is a robot in your view?
link |
Like how do you think of this system?
link |
I told you ridiculous questions.
link |
No, no, it's good.
link |
I mean, there's standard definitions
link |
of combining computation with some ability
link |
to do mechanical work.
link |
I think that gets us pretty close.
link |
But I think robotics has this problem
link |
that once things really work,
link |
we don't call them robots anymore.
link |
Like my dishwasher at home is pretty sophisticated,
link |
beautiful mechanisms.
link |
There's actually a pretty good computer,
link |
probably a couple of chips in there doing amazing things.
link |
We don't think of that as a robot anymore,
link |
Because then what roughly it means
link |
that robotics always has to solve the next problem
link |
and doesn't get to celebrate its past successes.
link |
I mean, even factory room floor robots
link |
are super successful.
link |
But that's not the ones,
link |
I mean, people think of them as robots,
link |
if you ask what are the successes of robotics,
link |
somehow it doesn't come to your mind immediately.
link |
So the definition of robot is a system
link |
with some level of automation that fails frequently.
link |
Something like, it's the computation plus mechanical work
link |
and an unsolved problem.
link |
It's an unsolved problem, yeah.
link |
So from a perspective of control and mechanics,
link |
dynamics, what is a robot?
link |
So there are many different types of robots.
link |
The control that you need for a Jibo robot,
link |
you know, some robot that's sitting on your countertop
link |
and interacting with you, but not touching you,
link |
for instance, is very different than what you need
link |
for an autonomous car or an autonomous drone.
link |
It's very different than what you need for a robot
link |
that's gonna walk or pick things up with its hands, right?
link |
My passion has always been for the places
link |
where you're interacting more,
link |
you're doing more dynamic interactions with the world.
link |
So walking, now manipulation.
link |
And the control problems there are beautiful.
link |
I think contact is one thing that differentiates them
link |
from many of the control problems we've solved classically,
link |
right, like modern control grew up stabilizing fighter jets
link |
that were passively unstable,
link |
and there's like amazing success stories from control
link |
all over the place.
link |
Power grid, I mean, there's all kinds of,
link |
it's everywhere that we don't even realize,
link |
just like AI is now.
link |
So you mentioned contact, like what's contact?
link |
So an airplane is an extremely complex system
link |
or a spacecraft landing or whatever,
link |
but at least it has the luxury
link |
of things change relatively continuously.
link |
That's an oversimplification.
link |
But if I make a small change
link |
in the command I send to my actuator,
link |
then the path that the robot will take
link |
tends to change only by a small amount.
link |
And there's a feedback mechanism here.
link |
That's what we're talking about.
link |
And there's a feedback mechanism.
link |
And thinking about this as locally,
link |
like a linear system, for instance,
link |
I can use more linear algebra tools
link |
to study systems like that,
link |
generalizations of linear algebra to these smooth systems.
link |
The robot has something very discontinuous
link |
that happens when it makes or breaks,
link |
when it starts touching the world.
link |
And even the way it touches or the order of contacts
link |
can change the outcome in potentially unpredictable ways.
link |
Not unpredictable, but complex ways.
link |
I do think there's a little bit of,
link |
a lot of people will say that contact is hard in robotics,
link |
And I think there's a little bit of a,
link |
there's truth to that,
link |
but maybe a misunderstanding around that.
link |
So what is limiting is that when we think about our robots
link |
and we write our simulators,
link |
we often make an assumption that objects are rigid.
link |
And when it comes down, that their mass moves all,
link |
stays in a constant position relative to each other itself.
link |
And that leads to some paradoxes
link |
when you go to try to talk about
link |
rigid body mechanics and contact.
link |
And so for instance, if I have a three legged stool
link |
with just imagine it comes to a point at the leg.
link |
So it's only touching the world at a point.
link |
If I draw my physics,
link |
my high school physics diagram of the system,
link |
then there's a couple of things
link |
that I'm given by elementary physics.
link |
I know if the system, if the table is at rest,
link |
if it's not moving, zero velocities,
link |
that means that the normal force,
link |
all the forces are in balance.
link |
So the force of gravity is being countered
link |
by the forces that the ground is pushing on my table legs.
link |
I also know since it's not rotating
link |
that the moments have to balance.
link |
And since it's a three dimensional table,
link |
it could fall in any direction.
link |
It actually tells me uniquely
link |
what those three normal forces have to be.
link |
If I have four legs on my table,
link |
four legged table and they were perfectly machined
link |
to be exactly the right same height
link |
and they're set down and the table's not moving,
link |
then the basic conservation laws don't tell me,
link |
there are many solutions for the forces
link |
that the ground could be putting on my legs
link |
that would still result in the table not moving.
link |
Now, the reason that seems fine, I could just pick one.
link |
But it gets funny now because if you think about friction,
link |
what we think about with friction is our standard model
link |
says the amount of force that the table will push back
link |
if I were to now try to push my table sideways,
link |
I guess I have a table here,
link |
is proportional to the normal force.
link |
So if I'm barely touching and I push, I'll slide,
link |
but if I'm pushing more and I push, I'll slide less.
link |
It's called coulomb friction is our standard model.
link |
Now, if you don't know what the normal force is
link |
on the four legs and you push the table,
link |
then you don't know what the friction forces are gonna be.
link |
And so you can't actually tell,
link |
the laws just aren't explicit yet
link |
about which way the table's gonna go.
link |
It could veer off to the left,
link |
it could veer off to the right, it could go straight.
link |
So the rigid body assumption of contact
link |
leaves us with some paradoxes,
link |
which are annoying for writing simulators
link |
and for writing controllers.
link |
We still do that sometimes because soft contact
link |
is potentially harder numerically or whatever,
link |
and the best simulators do both
link |
or do some combination of the two.
link |
But anyways, because of these kinds of paradoxes,
link |
there's all kinds of paradoxes in contact,
link |
mostly due to these rigid body assumptions.
link |
It becomes very hard to write the same kind of control laws
link |
that we've been able to be successful with
link |
Like fighter jets, we haven't been as successful
link |
writing those controllers for manipulation.
link |
And so you don't know what's going to happen
link |
at the point of contact, at the moment of contact.
link |
There are situations absolutely
link |
where our laws don't tell us.
link |
So the standard approach, that's okay.
link |
I mean, instead of having a differential equation,
link |
you end up with a differential inclusion, it's called.
link |
It's a set valued equation.
link |
It says that I'm in this configuration,
link |
I have these forces applied on me.
link |
And there's a set of things that could happen, right?
link |
And those aren't continuous, I mean, what...
link |
So when you're saying like non smooth,
link |
they're not only not smooth, but this is discontinuous?
link |
The non smooth comes in
link |
when I make or break a new contact first,
link |
or when I transition from stick to slip.
link |
So you typically have static friction,
link |
and then you'll start sliding,
link |
and that'll be a discontinuous change in philosophy.
link |
In philosophy, for instance,
link |
especially if you come to rest or...
link |
That's so fascinating.
link |
Okay, so what do you do?
link |
Sorry, I interrupted you.
link |
What's the hope under so much uncertainty
link |
about what's going to happen?
link |
What are you supposed to do?
link |
I mean, control has an answer for this.
link |
Robust control is one approach,
link |
but roughly you can write controllers
link |
which try to still perform the right task
link |
despite all the things that could possibly happen.
link |
The world might want the table to go this way and this way,
link |
but if I write a controller that pushes a little bit more
link |
and pushes a little bit,
link |
I can certainly make the table go in the direction I want.
link |
It just puts a little bit more of a burden
link |
on the control system, right?
link |
And this discontinuities do change the control system
link |
because the way we write it down right now,
link |
every different control configuration,
link |
including sticking or sliding
link |
or parts of my body that are in contact or not,
link |
looks like a different system.
link |
And I think of them,
link |
I reason about them separately or differently
link |
and the combinatorics of that blow up, right?
link |
So I just don't have enough time to compute
link |
all the possible contact configurations of my humanoid.
link |
Interestingly, I mean, I'm a humanoid.
link |
I have lots of degrees of freedom, lots of joints.
link |
I've only been around for a handful of years.
link |
It's getting up there,
link |
but I haven't had time in my life
link |
to visit all of the states in my system,
link |
certainly all the contact configurations.
link |
So if step one is to consider
link |
every possible contact configuration that I'll ever be in,
link |
that's probably not a problem I need to solve, right?
link |
Just as a small tangent, what's a contact configuration?
link |
What like, just so we can enumerate
link |
what are we talking about?
link |
How many are there?
link |
The simplest example maybe would be,
link |
imagine a robot with a flat foot.
link |
And we think about the phases of gait
link |
where the heel strikes and then the front toe strikes,
link |
and then you can heel up, toe off.
link |
Those are each different contact configurations.
link |
I only had two different contacts,
link |
but I ended up with four different contact configurations.
link |
Now, of course, my robot might actually have bumps on it
link |
so it could be much more subtle than that, right?
link |
But it's just even with one sort of box
link |
interacting with the ground already in the plane
link |
has that many, right?
link |
And if I was just even a 3D foot,
link |
then it probably my left toe might touch
link |
just before my right toe and things get subtle.
link |
Now, if I'm a dexterous hand
link |
and I go to talk about just grabbing a water bottle,
link |
if I have to enumerate every possible order
link |
that my hand came into contact with the bottle,
link |
then I'm dead in the water.
link |
Any approach that we were able to get away with that
link |
in walking because we mostly touched the ground
link |
within a small number of points, for instance,
link |
and we haven't been able to get dexterous hands that way.
link |
So you've mentioned that people think
link |
that contact is really hard
link |
and that that's the reason that robotic manipulation
link |
is problem is really hard.
link |
Is there any flaws in that thinking?
link |
So I think simulating contact is one aspect.
link |
I know people often say that we don't,
link |
that one of the reasons that we have a limit in robotics
link |
is because we do not simulate contact accurately
link |
in our simulators.
link |
And I think that is the extent to which that's true
link |
is partly because our simulators,
link |
we haven't got mature enough simulators.
link |
There are some things that are still hard, difficult,
link |
that we should change,
link |
but we actually, we know what the governing equations are.
link |
They have some foibles like this indeterminacy,
link |
but we should be able to simulate them accurately.
link |
We have incredible open source community in robotics,
link |
but it actually just takes a professional engineering team
link |
a lot of work to write a very good simulator like that.
link |
Now, where does, I believe you've written, Drake.
link |
There's a team of people.
link |
I certainly spent a lot of hours on it myself.
link |
But what is Drake and what does it take to create
link |
a simulation environment for the kind of difficult control
link |
problems we're talking about?
link |
Right, so Drake is the simulator that I've been working on.
link |
There are other good simulators out there.
link |
I don't like to think of Drake as just a simulator
link |
because we write our controllers in Drake,
link |
we write our perception systems a little bit in Drake,
link |
but we write all of our low level control
link |
and even planning and optimization.
link |
So it has optimization capabilities as well?
link |
I mean, Drake is three things roughly.
link |
It's an optimization library, which is sits on,
link |
it provides a layer of abstraction in C++ and Python
link |
for commercial solvers.
link |
You can write linear programs, quadratic programs,
link |
semi definite programs, sums of squares programs,
link |
the ones we've used, mixed integer programs,
link |
and it will do the work to curate those
link |
and send them to whatever the right solver is for instance,
link |
and it provides a level of abstraction.
link |
The second thing is a system modeling language,
link |
a bit like LabVIEW or Simulink,
link |
where you can make block diagrams out of complex systems,
link |
or it's like ROS in that sense,
link |
where you might have lots of ROS nodes
link |
that are each doing some part of your system,
link |
but to contrast it with ROS, we try to write,
link |
if you write a Drake system, then you have to,
link |
it asks you to describe a little bit more about the system.
link |
If you have any state, for instance, in the system,
link |
any variables that are gonna persist,
link |
you have to declare them.
link |
Parameters can be declared and the like,
link |
but the advantage of doing that is that you can,
link |
if you like, run things all on one process,
link |
but you can also do control design against it.
link |
You can do, I mean, simple things like rewinding
link |
and playing back your simulations, for instance,
link |
these things, you get some rewards
link |
for spending a little bit more upfront cost
link |
in describing each system.
link |
And I was inspired to do that
link |
because I think the complexity of Atlas, for instance,
link |
And I think, although, I mean,
link |
ROS has been an incredible, absolutely huge fan
link |
of what it's done for the robotics community,
link |
but the ability to rapidly put different pieces together
link |
and have a functioning thing is very good.
link |
But I do think that it's hard to think clearly
link |
about a bag of disparate parts,
link |
Mr. Potato Head kind of software stack.
link |
And if you can ask a little bit more
link |
out of each of those parts,
link |
then you can understand the way they work better.
link |
You can try to verify them and the like,
link |
or you can do learning against them.
link |
And then one of those systems, the last thing,
link |
I said the first two things that Drake is,
link |
but the last thing is that there is a set
link |
of multi body equations, rigid body equations,
link |
that is trying to provide a system that simulates physics.
link |
And we also have renderers and other things,
link |
but I think the physics component of Drake is special
link |
in the sense that we have done excessive amount
link |
of engineering to make sure
link |
that we've written the equations correctly.
link |
Every possible tumbling satellite or spinning top
link |
or anything that we could possibly write as a test is tested.
link |
We are making some, I think, fundamental improvements
link |
on the way you simulate contact.
link |
Just what does it take to simulate contact?
link |
I mean, it just seems,
link |
I mean, there's something just beautiful
link |
to the way you were like explaining contact
link |
and you were like tapping your fingers
link |
on the table while you're doing it, just.
link |
Easily, just like, just not even like,
link |
it was like helping you think, I guess.
link |
So you have this like awesome demo
link |
of loading or unloading a dishwasher,
link |
just picking up a plate,
link |
or grasping it like for the first time.
link |
That's just seems like so difficult.
link |
What, how do you simulate any of that?
link |
So it was really interesting that what happened was
link |
that we started getting more professional
link |
about our software development
link |
during the DARPA Robotics Challenge.
link |
I learned the value of software engineering
link |
and how these, how to bridle complexity.
link |
I guess that's what I want to somehow fight against
link |
and bring some of the clear thinking of controls
link |
into these complex systems we're building for robots.
link |
Shortly after the DARPA Robotics Challenge,
link |
Toyota opened a research institute,
link |
TRI, Toyota Research Institute.
link |
They put one of their, there's three locations.
link |
One of them is just down the street from MIT.
link |
And I helped ramp that up right up
link |
as a part of my, the end of my sabbatical, I guess.
link |
So TRI has given me, the TRI robotics effort
link |
has made this investment in simulation in Drake.
link |
And Michael Sherman leads a team there
link |
of just absolutely top notch dynamics experts
link |
that are trying to write those simulators
link |
that can pick up the dishes.
link |
And there's also a team working on manipulation there
link |
that is taking problems like loading the dishwasher.
link |
And we're using that to study these really hard corner cases
link |
kind of problems in manipulation.
link |
So for me, this, you know, simulating the dishes,
link |
we could actually write a controller.
link |
If we just cared about picking up dishes in the sink once,
link |
we could write a controller
link |
without any simulation whatsoever,
link |
and we could call it done.
link |
But we want to understand like,
link |
what is the path you take to actually get to a robot
link |
that could perform that for any dish in anybody's kitchen
link |
with enough confidence
link |
that it could be a commercial product, right?
link |
And it has deep learning perception in the loop.
link |
It has complex dynamics in the loop.
link |
It has controller, it has a planner.
link |
And how do you take all of that complexity
link |
and put it through this engineering discipline
link |
and verification and validation process
link |
to actually get enough confidence to deploy?
link |
I mean, the DARPA challenge made me realize
link |
that that's not something you throw over the fence
link |
and hope that somebody will harden it for you,
link |
that there are really fundamental challenges
link |
in closing that last gap.
link |
They're doing the validation and the testing.
link |
I think it might even change the way we have to think about
link |
the way we write systems.
link |
What happens if you have the robot running lots of tests
link |
and it screws up, it breaks a dish, right?
link |
How do you capture that?
link |
I said, you can't run the same simulation
link |
or the same experiment twice on a real robot.
link |
Do we have to be able to bring that one off failure
link |
back into simulation
link |
in order to change our controllers, study it,
link |
make sure it won't happen again?
link |
Do we, is it enough to just try to add that
link |
to our distribution and understand that on average,
link |
we're gonna cover that situation again?
link |
There's like really subtle questions at the corner cases
link |
that I think we don't yet have satisfying answers for.
link |
Like how do you find the corner cases?
link |
That's one kind of, is there,
link |
do you think that's possible to create a systematized way
link |
of discovering corner cases efficiently?
link |
In whatever the problem is?
link |
Yes, I mean, I think we have to get better at that.
link |
I mean, control theory has for decades
link |
talked about active experiment design.
link |
So people call it curiosity these days.
link |
It's roughly this idea of trying to exploration
link |
or exploitation, but in the active experiment design
link |
is even, is more specific.
link |
You could try to understand the uncertainty in your system,
link |
design the experiment that will provide
link |
the maximum information to reduce that uncertainty.
link |
If there's a parameter you wanna learn about,
link |
what is the optimal trajectory I could execute
link |
to learn about that parameter, for instance.
link |
Scaling that up to something that has a deep network
link |
in the loop and a planning in the loop is tough.
link |
We've done some work on, you know,
link |
with Matt Okely and Aman Sinha,
link |
we've worked on some falsification algorithms
link |
that are trying to do rare event simulation
link |
that try to just hammer on your simulator.
link |
And if your simulator is good enough,
link |
you can spend a lot of time,
link |
or you can write good algorithms
link |
that try to spend most of their time in the corner cases.
link |
So you basically imagine you're building an autonomous car
link |
and you wanna put it in, I don't know,
link |
downtown New Delhi all the time, right?
link |
And accelerated testing.
link |
If you can write sampling strategies,
link |
which figure out where your controller's
link |
performing badly in simulation
link |
and start generating lots of examples around that.
link |
You know, it's just the space of possible places
link |
where that can be, where things can go wrong is very big.
link |
So it's hard to write those algorithms.
link |
Yeah, rare event simulation
link |
is just a really compelling notion, if it's possible.
link |
We joked and we call it the black swan generator.
link |
It's a black swan.
link |
Because you don't just want the rare events,
link |
you want the ones that are highly impactful.
link |
I mean, that's the most,
link |
those are the most sort of profound questions
link |
we ask of our world.
link |
Like, what's the worst that can happen?
link |
But what we're really asking
link |
isn't some kind of like computer science,
link |
worst case analysis.
link |
We're asking like, what are the millions of ways
link |
this can go wrong?
link |
And that's like our curiosity.
link |
And we humans, I think are pretty bad at,
link |
we just like run into it.
link |
And I think there's a distributed sense
link |
because there's now like 7.5 billion of us.
link |
And so there's a lot of them.
link |
And then a lot of them write blog posts
link |
about the stupid thing they've done.
link |
So we learn in a distributed way.
link |
I think that's gonna be important for robots too.
link |
I mean, that's another massive theme
link |
at Toyota Research for Robotics
link |
is this fleet learning concept
link |
is the idea that I, as a human,
link |
I don't have enough time to visit all of my states, right?
link |
There's just a, it's very hard for one robot
link |
to experience all the things.
link |
But that's not actually the problem we have to solve, right?
link |
We're gonna have fleets of robots
link |
that can have very similar appendages.
link |
And at some point, maybe collectively,
link |
they have enough data
link |
that their computational processes
link |
should be set up differently than ours, right?
link |
It's this vision of just,
link |
I mean, all these dishwasher unloading robots.
link |
I mean, that robot dropping a plate
link |
and a human looking at the robot probably pissed off.
link |
But that's a special moment to record.
link |
I think one thing in terms of fleet learning,
link |
and I've seen that because I've talked to a lot of folks,
link |
just like Tesla users or Tesla drivers,
link |
they're another company
link |
that's using this kind of fleet learning idea.
link |
One hopeful thing I have about humans
link |
is they really enjoy when a system improves, learns.
link |
So they enjoy fleet learning.
link |
And the reason it's hopeful for me
link |
is they're willing to put up with something
link |
that's kind of dumb right now.
link |
And they're like, if it's improving,
link |
they almost like enjoy being part of the, like teaching it.
link |
Almost like if you have kids,
link |
like you're teaching them something, right?
link |
I think that's a beautiful thing
link |
because that gives me hope
link |
that we can put dumb robots out there.
link |
I mean, the problem on the Tesla side with cars,
link |
cars can kill you.
link |
That makes the problem so much harder.
link |
Dishwasher unloading is a little safe.
link |
That's why home robotics is really exciting.
link |
And just to clarify, I mean, for people who might not know,
link |
I mean, TRI, Toyota Research Institute.
link |
So they're, I mean, they're pretty well known
link |
for like autonomous vehicle research,
link |
but they're also interested in home robotics.
link |
Yep, there's a big group working on,
link |
multiple groups working on home robotics.
link |
It's a major part of the portfolio.
link |
There's also a couple other projects
link |
in advanced materials discovery,
link |
using AI and machine learning to discover new materials
link |
for car batteries and the like, for instance, yeah.
link |
And that's been actually an incredibly successful team.
link |
There's new projects starting up too, so.
link |
Do you see a future of where like robots are in our home
link |
and like robots that have like actuators
link |
that look like arms in our home
link |
or like, you know, more like humanoid type robots?
link |
Or is this, are we gonna do the same thing
link |
that you just mentioned that, you know,
link |
the dishwasher is no longer a robot.
link |
We're going to just not even see them as robots.
link |
But I mean, what's your vision of the home of the future
link |
10, 20 years from now, 50 years, if you get crazy?
link |
Yeah, I think we already have Roombas cruising around.
link |
We have, you know, Alexis or Google Homes
link |
on our kitchen counter.
link |
It's only a matter of time until they spring arms
link |
and start doing something useful like that.
link |
So I do think it's coming.
link |
I think lots of people have lots of motivations
link |
It's been super interesting actually learning
link |
about Toyota's vision for it,
link |
which is about helping people age in place.
link |
Cause I think that's not necessarily the first entry,
link |
the most lucrative entry point,
link |
but it's the problem maybe that we really need to solve
link |
And so I think there's a real opportunity.
link |
It's a delicate problem.
link |
How do you work with people, help people,
link |
keep them active, engaged, you know,
link |
but improve their quality of life
link |
and help them age in place, for instance.
link |
It's interesting because older folks are also,
link |
I mean, there's a contrast there
link |
because they're not always the folks
link |
who are the most comfortable with technology, for example.
link |
So there's a division that's interesting.
link |
You can do so much good with a robot for older folks,
link |
but there's a gap to fill of understanding.
link |
I mean, it's actually kind of beautiful.
link |
Robot is learning about the human
link |
and the human is kind of learning about this new robot thing.
link |
And it's also with, at least with,
link |
like when I talked to my parents about robots,
link |
there's a little bit of a blank slate there too.
link |
Like you can, I mean, they don't know anything
link |
about robotics, so it's completely like wide open.
link |
They don't have, they haven't,
link |
my parents haven't seen Black Mirror.
link |
So like they, it's a blank slate.
link |
Here's a cool thing, like what can it do for me?
link |
Yeah, so it's an exciting space.
link |
I think it's a really important space.
link |
I do feel like a few years ago,
link |
drones were successful enough in academia.
link |
They kind of broke out and started an industry
link |
and autonomous cars have been happening.
link |
It does feel like manipulation in logistics, of course,
link |
first, but in the home shortly after,
link |
seems like one of the next big things
link |
that's gonna really pop.
link |
So I don't think we talked about it,
link |
but what's soft robotics?
link |
So we talked about like rigid bodies.
link |
Like if we can just linger on this whole touch thing.
link |
Yeah, so what's soft robotics?
link |
So I told you that I really dislike the fact
link |
that robots are afraid of touching the world
link |
all over their body.
link |
So there's a couple reasons for that.
link |
If you look carefully at all the places
link |
that robots actually do touch the world,
link |
they're almost always soft.
link |
They have some sort of pad on their fingers
link |
or a rubber sole on their foot.
link |
But if you look up and down the arm,
link |
we're just pure aluminum or something.
link |
So that makes it hard actually.
link |
In fact, hitting the table with your rigid arm
link |
or nearly rigid arm has some of the problems
link |
that we talked about in terms of simulation.
link |
I think it fundamentally changes the mechanics of contact
link |
when you're soft, right?
link |
You turn point contacts into patch contacts,
link |
which can have torsional friction.
link |
You can have distributed load.
link |
If I wanna pick up an egg, right?
link |
If I pick it up with two points,
link |
then in order to put enough force
link |
to sustain the weight of the egg,
link |
I might have to put a lot of force to break the egg.
link |
If I envelop it with contact all around,
link |
then I can distribute my force across the shell of the egg
link |
and have a better chance of not breaking it.
link |
So soft robotics is for me a lot about changing
link |
the mechanics of contact.
link |
Does it make the problem a lot harder?
link |
Quite the opposite.
link |
It changes the computational problem.
link |
I think because of the, I think our world
link |
and our mathematics has biased us towards rigid.
link |
But it really should make things better in some ways, right?
link |
I think the future is unwritten there.
link |
But the other thing it can do.
link |
I think ultimately, sorry to interrupt,
link |
but I think ultimately it will make things simpler
link |
if we embrace the softness of the world.
link |
It makes things smoother, right?
link |
So the result of small actions is less discontinuous,
link |
but it also means potentially less instantaneously bad.
link |
For instance, I won't necessarily contact something
link |
and send it flying off.
link |
The other aspect of it
link |
that just happens to dovetail really well
link |
is that soft robotics tends to be a place
link |
where we can embed a lot of sensors too.
link |
So if you change your hardware and make it more soft,
link |
then you can potentially have a tactile sensor,
link |
which is measuring the deformation.
link |
So there's a team at TRI that's working on soft hands
link |
and you get so much more information.
link |
You can put a camera behind the skin roughly
link |
and get fantastic tactile information,
link |
which is, it's super important.
link |
Like in manipulation,
link |
one of the things that really is frustrating
link |
is if you work super hard on your head mounted,
link |
on your perception system for your head mounted cameras,
link |
and then you get a lot of information
link |
for your head mounted cameras,
link |
and then you've identified an object,
link |
you reach down to touch it,
link |
and the last thing that happens,
link |
right before the most important time,
link |
you stick your hand
link |
and you're occluding your head mounted sensors.
link |
So in all the part that really matters,
link |
all of your off board sensors are occluded.
link |
And really, if you don't have tactile information,
link |
then you're blind in an important way.
link |
So it happens that soft robotics and tactile sensing
link |
tend to go hand in hand.
link |
I think we've kind of talked about it,
link |
but you taught a course on underactuated robotics.
link |
I believe that was the name of it, actually.
link |
Can you talk about it in that context?
link |
What is underactuated robotics?
link |
Right, so underactuated robotics is my graduate course.
link |
It's online mostly now,
link |
in the sense that the lectures.
link |
Several versions of it, I think.
link |
Right, the YouTube.
link |
It's really great, I recommend it highly.
link |
Look on YouTube for the 2020 versions.
link |
Until March, and then you have to go back to 2019,
link |
No, I've poured my heart into that class.
link |
And lecture one is basically explaining
link |
what the word underactuated means.
link |
So people are very kind to show up
link |
and then maybe have to learn
link |
what the title of the course means
link |
over the course of the first lecture.
link |
That first lecture is really good.
link |
You should watch it.
link |
It's a strange name,
link |
but I thought it captured the essence
link |
of what control was good at doing
link |
and what control was bad at doing.
link |
So what do I mean by underactuated?
link |
So a mechanical system
link |
has many degrees of freedom, for instance.
link |
I think of a joint as a degree of freedom.
link |
And it has some number of actuators, motors.
link |
So if you have a robot that's bolted to the table
link |
that has five degrees of freedom and five motors,
link |
then you have a fully actuated robot.
link |
If you take away one of those motors,
link |
then you have an underactuated robot.
link |
Now, why on earth?
link |
I have a good friend who likes to tease me.
link |
He said, Ross, if you had more research funding,
link |
would you work on fully actuated robots?
link |
And the answer is no.
link |
The world gives us underactuated robots,
link |
whether we like it or not.
link |
I'm an underactuated robot,
link |
even though I have more muscles
link |
than my big degrees of freedom,
link |
because I have in some places
link |
multiple muscles attached to the same joint.
link |
But still, there's a really important degree of freedom
link |
that I have, which is the location of my center of mass
link |
in space, for instance.
link |
All right, I can jump into the air,
link |
and there's no motor that connects my center of mass
link |
to the ground in that case.
link |
So I have to think about the implications
link |
of not having control over everything.
link |
The passive dynamic walkers are the extreme view of that,
link |
where you've taken away all the motors,
link |
and you have to let physics do the work.
link |
But it shows up in all of the walking robots,
link |
where you have to use some of the actuators
link |
to push and pull even the degrees of freedom
link |
that you don't have an actuator on.
link |
That's referring to walking if you're falling forward.
link |
Is there a way to walk that's fully actuated?
link |
So it's a subtle point.
link |
When you're in contact and you have your feet on the ground,
link |
there are still limits to what you can do, right?
link |
Unless I have suction cups on my feet,
link |
I cannot accelerate my center of mass towards the ground
link |
faster than gravity,
link |
because I can't get a force pushing me down, right?
link |
But I can still do most of the things that I want to.
link |
So you can get away with basically thinking of the system
link |
as fully actuated,
link |
unless you suddenly needed to accelerate down super fast.
link |
But as soon as I take a step,
link |
I get into the more nuanced territory,
link |
and to get to really dynamic robots,
link |
or airplanes or other things,
link |
I think you have to embrace the underactuated dynamics.
link |
Manipulation, people think, is manipulation underactuated?
link |
Even if my arm is fully actuated, I have a motor,
link |
if my goal is to control the position and orientation
link |
of this cup, then I don't have an actuator
link |
for that directly.
link |
So I have to use my actuators over here
link |
to control this thing.
link |
Now it gets even worse,
link |
like what if I have to button my shirt, okay?
link |
What are the degrees of freedom of my shirt, right?
link |
I suddenly, that's a hard question to think about.
link |
It kind of makes me queasy
link |
thinking about my state space control ideas.
link |
But actually those are the problems
link |
that make me so excited about manipulation right now,
link |
is that it breaks some of the,
link |
it breaks a lot of the foundational control stuff
link |
that I've been thinking about.
link |
Is there, what are some interesting insights
link |
you could say about trying to solve an underactuated,
link |
a control in an underactuated system?
link |
So I think the philosophy there
link |
is let physics do more of the work.
link |
The technical approach has been optimization.
link |
So you typically formulate your decision making
link |
for control as an optimization problem.
link |
And you use the language of optimal control
link |
and sometimes often numerical optimal control
link |
in order to make those decisions and balance,
link |
these complicated equations of,
link |
and in order to control,
link |
you don't have to use optimal control
link |
to do underactuated systems,
link |
but that has been the technical approach
link |
that has borne the most fruit in our,
link |
at least in our line of work.
link |
And there's some, so in underactuated systems,
link |
when you say let physics do some of the work,
link |
so there's a kind of feedback loop
link |
that observes the state that the physics brought you to.
link |
So like you've, there's a perception there,
link |
there's a feedback somehow.
link |
Do you ever loop in like complicated perception systems
link |
into this whole picture?
link |
Right, right around the time of the DARPA challenge,
link |
we had a complicated perception system
link |
in the DARPA challenge.
link |
We also started to embrace perception
link |
for our flying vehicles at the time.
link |
We had a really good project
link |
on trying to make airplanes fly
link |
at high speeds through forests.
link |
Sirtash Karaman was on that project
link |
and we had, it was a really fun team to work on.
link |
He's carried it farther, much farther forward since then.
link |
And that's using cameras for perception?
link |
So that was using cameras.
link |
That was, at the time we felt like LIDAR
link |
was too heavy and too power heavy
link |
to be carried on a light UAV,
link |
and we were using cameras.
link |
And that was a big part of it was just
link |
how do you do even stereo matching
link |
at a fast enough rate with a small camera,
link |
small onboard compute.
link |
Since then we have now,
link |
so the deep learning revolution
link |
unquestionably changed what we can do
link |
with perception for robotics and control.
link |
So in manipulation, we can address,
link |
we can use perception in I think a much deeper way.
link |
And we get into not only,
link |
I think the first use of it naturally
link |
would be to ask your deep learning system
link |
to look at the cameras and produce the state,
link |
which is like the pose of my thing, for instance.
link |
But I think we've quickly found out
link |
that that's not always the right thing to do.
link |
Because what's the state of my shirt?
link |
Imagine, I've always,
link |
Very noisy, you mean, or?
link |
It's, if the first step of me trying to button my shirt
link |
is estimate the full state of my shirt,
link |
including like what's happening in the back here,
link |
whatever, whatever.
link |
That's just not the right specification.
link |
There are aspects of the state
link |
that are very important to the task.
link |
There are many that are unobservable
link |
and not important to the task.
link |
So you really need,
link |
it begs new questions about state representation.
link |
Another example that we've been playing with in lab
link |
has been just the idea of chopping onions, okay?
link |
Or carrots, turns out to be better.
link |
So onions stink up the lab.
link |
And they're hard to see in a camera.
link |
Details matter, yeah.
link |
Details matter, you know?
link |
So if I'm moving around a particular object, right?
link |
Then I think about,
link |
oh, it's got a position or an orientation in space.
link |
That's the description I want.
link |
Now, when I'm chopping an onion, okay?
link |
Like the first chop comes down.
link |
I have now a hundred pieces of onion.
link |
Does my control system really need to understand
link |
the position and orientation and even the shape
link |
of the hundred pieces of onion in order to make a decision?
link |
Probably not, you know?
link |
And if I keep going, I'm just getting,
link |
more and more is my state space getting bigger as I cut?
link |
So somehow there's a,
link |
I think there's a richer idea of state.
link |
It's not the state that is given to us
link |
by Lagrangian mechanics.
link |
There is a proper Lagrangian state of the system,
link |
but the relevant state for this is some latent state
link |
is what we call it in machine learning.
link |
But, you know, there's some different state representation.
link |
Some compressed representation, some.
link |
And that's what I worry about saying compressed
link |
because it doesn't,
link |
I don't mind that it's low dimensional or not,
link |
but it has to be something that's easier to think about.
link |
Or the algorithms being like control, optimal.
link |
So for instance, if the contact mechanics
link |
of all of those onion pieces and all the permutations
link |
of possible touches between those onion pieces,
link |
you know, you can give me
link |
a high dimensional state representation,
link |
I'm okay if it's linear.
link |
But if I have to think about all the possible
link |
shattering combinatorics of that,
link |
then my robot's gonna sit there thinking
link |
and the soup's gonna get cold or something.
link |
So since you taught the course,
link |
it kind of entered my mind,
link |
the idea of underactuated as really compelling
link |
to see the world in this kind of way.
link |
Do you ever, you know, if we talk about onions
link |
or you talk about the world with people in it in general,
link |
do you see the world as basically an underactuated system?
link |
Do you like often look at the world in this way?
link |
Or is this overreach?
link |
Underactuated is a way of life, man.
link |
Exactly, I guess that's what I'm asking.
link |
I do think it's everywhere.
link |
I think in some places,
link |
we already have natural tools to deal with it.
link |
You know, it rears its head.
link |
I mean, in linear systems, it's not a problem.
link |
We just, like an underactuated linear system
link |
is really not sufficiently distinct
link |
from a fully actuated linear system.
link |
It's a subtle point about when that becomes a bottleneck
link |
in what we know how to do with control.
link |
It happens to be a bottleneck,
link |
although we've gotten incredibly good solutions now,
link |
but for a long time that I felt
link |
that that was the key bottleneck in legged robots.
link |
And roughly now the underactuated course
link |
is me trying to tell people everything I can
link |
about how to make Atlas do a backflip, right?
link |
I have a second course now
link |
that I teach in the other semesters,
link |
which is on manipulation.
link |
And that's where we get into now more of the,
link |
that's a newer class.
link |
I'm hoping to put it online this fall completely.
link |
And that's gonna have much more aspects
link |
about these perception problems
link |
and the state representation questions,
link |
and then how do you do control.
link |
And the thing that's a little bit sad is that,
link |
for me at least, is there's a lot of manipulation tasks
link |
that people wanna do and should wanna do.
link |
They could start a company with it and be very successful
link |
that don't actually require you to think that much
link |
about underact, or dynamics at all even,
link |
but certainly underactuated dynamics.
link |
Once I have, if I reach out and grab something,
link |
if I can sort of assume it's rigidly attached to my hand,
link |
then I can do a lot of interesting,
link |
meaningful things with it
link |
without really ever thinking about the dynamics
link |
So we've built systems that kind of reduce the need for that.
link |
Enveloping grasps and the like.
link |
But I think the really good problems in manipulation.
link |
So manipulation, by the way, is more than just pick and place.
link |
That's like a lot of people think of that, just grasping.
link |
I don't mean that.
link |
I mean buttoning my shirt, I mean tying shoelaces.
link |
How do you program a robot to tie shoelaces?
link |
And not just one shoe, but every shoe, right?
link |
That's a really good problem.
link |
It's tempting to write down like the infinite dimensional
link |
state of the laces, that's probably not needed
link |
to write a good controller.
link |
I know we could hand design a controller that would do it,
link |
but I don't want that.
link |
I want to understand the principles that would allow me
link |
to solve another problem that's kind of like that.
link |
But I think if we can stay pure in our approach,
link |
then the challenge of tying anybody's shoes
link |
is a great challenge.
link |
That's a great challenge.
link |
I mean, and the soft touch comes into play there.
link |
That's really interesting.
link |
Let me ask another ridiculous question on this topic.
link |
How important is touch?
link |
We haven't talked much about humans,
link |
but I have this argument with my dad
link |
where like I think you can fall in love with a robot
link |
based on language alone.
link |
And he believes that touch is essential.
link |
Touch and smell, he says.
link |
But so in terms of robots, connecting with humans,
link |
we can go philosophical in terms of like a deep,
link |
meaningful connection, like love,
link |
but even just like collaborating in an interesting way,
link |
how important is touch like from an engineering perspective
link |
and a philosophical one?
link |
I think it's super important.
link |
Even just in a practical sense,
link |
if we forget about the emotional part of it.
link |
But for robots to interact safely
link |
while they're doing meaningful mechanical work
link |
in the close contact with or vicinity of people
link |
that need help, I think we have to have them,
link |
we have to build them differently.
link |
They have to be afraid, not afraid of touching the world.
link |
So I think Baymax is just awesome.
link |
That's just like the movie of Big Hero 6
link |
and the concept of Baymax, that's just awesome.
link |
I think we should, and we have some folks at Toyota
link |
that are trying to, Toyota Research
link |
that are trying to build Baymax roughly.
link |
And I think it's just a fantastically good project.
link |
I think it will change the way people physically interact.
link |
The same way, I mean, you gave a couple examples earlier,
link |
but if the robot that was walking around my home
link |
looked more like a teddy bear
link |
and a little less like the Terminator,
link |
that could change completely the way people perceive it
link |
and interact with it.
link |
And maybe they'll even wanna teach it, like you said, right?
link |
You could not quite gamify it,
link |
but somehow instead of people judging it
link |
and looking at it as if it's not doing as well as a human,
link |
they're gonna try to help out the cute teddy bear, right?
link |
Who knows, but I think we're building robots wrong
link |
and being more soft and more contact is important, right?
link |
Yeah, I mean, like all the magical moments
link |
I can remember with robots,
link |
well, first of all, just visiting your lab and seeing Atlas,
link |
but also Spotmini, when I first saw Spotmini in person
link |
and hung out with him, her, it,
link |
I don't have trouble engendering robots.
link |
I feel the robotics people really say, oh, is it it?
link |
I kinda like the idea that it's a her or a him.
link |
There's a magical moment, but there's no touching.
link |
I guess the question I have, have you ever been,
link |
like, have you had a human robot experience
link |
where a robot touched you?
link |
And like, it was like, wait,
link |
like, was there a moment that you've forgotten
link |
that a robot is a robot and like,
link |
the anthropomorphization stepped in
link |
and for a second you forgot that it's not human?
link |
I mean, I think when you're in on the details,
link |
then we, of course, anthropomorphized our work with Atlas,
link |
but in verbal communication and the like,
link |
I think we were pretty aware of it
link |
as a machine that needed to be respected.
link |
And I actually, I worry more about the smaller robots
link |
that could still move quickly if programmed wrong
link |
and we have to be careful actually
link |
about safety and the like right now.
link |
And that, if we build our robots correctly,
link |
I think then those, a lot of those concerns could go away.
link |
And we're seeing that trend.
link |
We're seeing the lower cost, lighter weight arms now
link |
that could be fundamentally safe.
link |
I mean, I do think touch is so fundamental.
link |
Ted Adelson is great.
link |
He's a perceptual scientist at MIT
link |
and he studied vision most of his life.
link |
And he said, when I had kids,
link |
I expected to be fascinated by their perceptual development.
link |
But what really, what he noticed was,
link |
felt more impressive, more dominant
link |
was the way that they would touch everything
link |
and lick everything.
link |
And pick things up, stick it on their tongue and whatever.
link |
And he said, watching his daughter convinced him
link |
that actually he needed to study tactile sensing more.
link |
So there's something very important.
link |
I think it's a little bit also of the passive
link |
versus active part of the world, right?
link |
You can passively perceive the world.
link |
But it's fundamentally different if you can do an experiment
link |
and if you can change the world
link |
and you can learn a lot more than a passive observer.
link |
So you can in dialogue, that was your initial example,
link |
you could have an active experiment exchange.
link |
But I think if you're just a camera watching YouTube,
link |
I think that's a very different problem
link |
than if you're a robot that can apply force.
link |
And I think that's a very different problem
link |
than if you're a robot that can apply force and touch.
link |
I think it's important.
link |
Yeah, I think it's just an exciting area of research.
link |
I think you're probably right
link |
that this hasn't been under researched.
link |
To me as a person who's captivated
link |
by the idea of human robot interaction,
link |
it feels like such a rich opportunity to explore touch.
link |
Not even from a safety perspective,
link |
but like you said, the emotional too.
link |
I mean, safety comes first,
link |
but the next step is like a real human connection.
link |
Even in the industrial setting,
link |
it just feels like it's nice for the robot.
link |
I don't know, you might disagree with this,
link |
but because I think it's important
link |
to see robots as tools often,
link |
I think they're just always going to be more effective
link |
once you humanize them.
link |
Like it's convenient now to think of them as tools
link |
because we want to focus on the safety,
link |
but I think ultimately to create like a good experience
link |
for the worker, for the person,
link |
there has to be a human element.
link |
I don't know, for me,
link |
it feels like an industrial robotic arm
link |
would be better if it has a human element.
link |
I think like Rethink Robotics had that idea
link |
with the Baxter and having eyes and so on,
link |
having, I don't know, I'm a big believer in that.
link |
It's not my area, but I am also a big believer.
link |
Do you have an emotional connection to Atlas?
link |
Like do you miss him?
link |
I mean, yes, I don't know if I more so
link |
than if I had a different science project
link |
that I'd worked on super hard, right?
link |
But yeah, I mean, the robot,
link |
we basically had to do heart surgery on the robot
link |
in the final competition because we melted the core.
link |
Yeah, there was something about watching that robot
link |
We know we had to compete with it in an hour
link |
and it was getting its guts ripped out.
link |
Those are all historic moments.
link |
I think if you look back like a hundred years from now,
link |
yeah, I think those are important moments in robotics.
link |
I mean, these are the early days.
link |
You look at like the early days
link |
of a lot of scientific disciplines.
link |
They look ridiculous, they're full of failure,
link |
but it feels like robotics will be important
link |
in the coming a hundred years.
link |
And these are the early days.
link |
So I think a lot of people are,
link |
look at a brilliant person such as yourself
link |
and are curious about the intellectual journey they've took.
link |
Is there maybe three books, technical, fiction,
link |
philosophical that had a big impact on your life
link |
that you would recommend perhaps others reading?
link |
Yeah, so I actually didn't read that much as a kid,
link |
but I read fairly voraciously now.