Russ Tedrake: Underactuated Robotics, Control, Dynamics and Touch
嘉宾：Russ Tedrake
全文文稿 · 按对话顺序保留完整内容，深入理解上下文。
本篇目录18 章节
§01Transcript · 00:00–08:45§02Transcript · 08:46–18:15§03Transcript · 18:20–27:21§04Transcript · 27:24–37:11§05Transcript · 37:14–46:51§06Transcript · 46:53–55:36§07Transcript · 55:39–1:04:06§08Transcript · 1:04:11–1:13:46§09Transcript · 1:13:49–1:23:30§10Transcript · 1:23:33–1:34:01§11Transcript · 1:34:04–1:42:50§12Transcript · 1:42:52–1:52:53§13Transcript · 1:52:55–2:01:55§14Transcript · 2:01:59–2:10:59§15Transcript · 2:11:02–2:19:44§16Transcript · 2:19:46–2:30:15§17Transcript · 2:30:17–2:39:59§18Transcript · 2:40:02–2:48:39§01Transcript · 00:00–08:45
The following is a conversation with Russ Tedrick,
a roboticist and professor at MIT
and vice president of robotics research
at Toyota Research Institute or TRI.
He works on control of robots in interesting,
complicated, underactuated, stochastic,
difficult to model situations.
He's a great teacher and a great person.
One of my favorites at MIT.
We'll get into a lot of topics in this conversation
from his time leading MIT's DARPA Robotics Challenge Team
to the awesome fact that he often runs
close to a marathon a day to and from work barefoot.
For a world-class roboticist interested
in elegant, efficient control
of underactuated dynamical systems like the human body,
this fact makes Russ one of the most
fascinating people I know.
Quick summary of the ads.
Three sponsors, Magic Spoon Cereal,
BetterHelp and ExpressVPN.
Please consider supporting this podcast
by going to magicspoon.com/lex
and using code LEX at checkout,
going to betterhelp.com/lex
and signing up at expressvpn.com/lexpod.
Click the links in the description,
buy the stuff, get the discount.
It really is the best way to support this podcast.
If you enjoy this thing, subscribe on YouTube,
review it with five stars on Apple Podcast,
support it on Patreon or connect with me
on Twitter @lexfriedman.
As usual, I'll do a few minutes of ads now
and never any ads in the middle
that can break the flow of the conversation.
This episode is supported by Magic Spoon,
low-carb, keto-friendly cereal.
I've been on a mix of keto or carnivore diet
for a very long time now.
That means eating very little carbs.
I used to love cereal.
Obviously, most have crazy amounts of sugar,
which is terrible for you, so I quit years ago.
But Magic Spoon is a totally new thing.
Zero sugar, 11 grams of protein,
and only three net grams of carbs.
It tastes delicious.
It has a bunch of flavors, they're all good,
but if you know what's good for you, you'll go with cocoa,
my favorite flavor and the flavor of champions.
Click the magicspoon.com/lex link in the description,
use code LEX at checkout to get the discount
and to let them know I sent you.
So buy all of their cereal.
It's delicious and good for you.
You won't regret it.
The show is also sponsored by BetterHelp,
spelled H-E-L-P, help.
Check it out at betterhelp.com/lex.
They figure out what you need
and match you with a licensed professional therapist
in under 48 hours.
It's not a crisis line, it's not self-help,
it is professional counseling done securely online.
As you may know, I'm a bit from the David Goggins line
of creatures and so have some demons to contend with,
usually on long runs or all-nighters full of self-doubt.
I think suffering is essential for creation,
but you can suffer beautifully
in a way that doesn't destroy you.
For most people, I think a good therapist can help in this,
so it's at least worth a try.
Check out the reviews, they're all good.
It's easy, private, affordable, available worldwide.
You can communicate by text anytime
and schedule weekly audio and video sessions.
Check it out at betterhelp.com/lex.
This show is also sponsored by ExpressVPN.
Get it at expressvpn.com/lexpod to get a discount
and to support this podcast.
Have you ever watched "The Office"?
If you have, you probably know it's based
on a UK series also called "The Office."
Not to stir up trouble, but I personally think
the British version is actually more brilliant
than the American one, but both are amazing.
Anyway, there are actually nine other countries
with their own version of "The Office."
You can get access to them with no geo-restriction
when you use ExpressVPN.
It lets you control where you want sites
to think you're located.
You can choose from nearly 100 different countries,
giving you access to content
that isn't available in your region.
So again, get it on any device at expressvpn.com/lexpod
to get an extra three months free
and to support this podcast.
And now here's my conversation with Russ Tedrick.
What is the most beautiful motion of a animal or robot
that you've ever seen?
- I think the most beautiful motion of a robot
has to be the passive dynamic walkers.
I think there's just something fundamentally beautiful.
The ones in particular that Steve Collins built
with Andy Ruina at Cornell, a 3D walking machine.
So it was not confined to a boom or a plane
that you put it on top of a small ramp,
give it a little push.
It's powered only by gravity,
no controllers, no batteries whatsoever.
It just falls down the ramp.
And at the time it looked more natural, more graceful,
more human-like than any robot we'd seen to date.
Powered only by gravity.
- How does it work?
- Well, okay, the simplest model,
it's kind of like a slinky.
It's like an elaborate slinky.
One of the simplest models we use to think about it
is actually a rimless wheel.
So imagine taking a bicycle wheel, but take the rim off.
So it's now just got a bunch of spokes.
If you give that a push,
it still wants to roll down the ramp.
But every time its foot, its spoke comes around
and hits the ground, it loses a little energy.
Every time it takes a step forward,
it gains a little energy.
Those things can come into perfect balance.
And actually they want to, it's a stable phenomenon.
If it's going too slow, it'll speed up.
If it's going too fast, it'll slow down.
And it comes into a stable periodic motion.
Now you can take that rimless wheel,
which doesn't look very much like a human walking,
take all the extra spokes away, put a hinge in the middle.
Now it's two legs.
That's called our compass gate walker.
That can still, you give it a little push,
starts falling down a ramp.
Looks a little bit more like walking.
At least it's a biped.
But what Steve and Andy and Ted McGeer
started the whole exercise,
but what Steve and Andy did was they took it
to this beautiful conclusion
where they built something that had knees, arms, a torso,
the arms swung naturally, give it a little push.
And that looked like a stroll through the park.
- How do you design something like that?
I mean, is that art or science?
- It's on the boundary.
I think there's a science to getting close to the solution.
I think there's certainly art in the way that they,
they made a beautiful robot.
But then the finesse, because this was,
they were working with a system
that wasn't perfectly modeled, wasn't perfectly controlled.
There's all these little tricks
that you have to tune the suction cups at the knees,
for instance, so that they stick,
but then they release at just the right time.
Or there's all these little tricks of the trade,
which really are art, but it was a point.
I mean, it made the point.
We were at that time, the walking robot,
the best walking robot in the world was Honda's ASIMO.
Absolutely marvel of modern engineering.
- This is 90s?
- This was in 97 when they first released,
it sort of announced P2 and then it went through,
it was ASIMO by then in 2004.
- It looks like this very cautious walking,
like you're walking on hot coals or something like that.
- I think it gets a bad rap.
ASIMO is a beautiful machine.
It does walk with its knees bent.
Our Atlas walking had its knees bent,
but actually ASIMO was pretty fantastic,
but it wasn't energy efficient.
Neither was Atlas when we worked on Atlas.
None of our robots that have been that complicated
have been very energy efficient.
But there's a thing that happens when you do control,
when you try to control a system of that complexity.
You try to use your motors to basically counteract gravity.
Take whatever the world's doing to you and push back,
erase the dynamics of the world
and impose the dynamics you want
because you can make them simple and analyzable,
mathematically simple.
And this was a very sort of beautiful example
that you don't have to do that.
You can just let go, let physics do most of the work.
And you just have to give it a little bit of energy.
This one only walked down a ramp.
It would never walk on the flat.
To walk on the flat,
§02Transcript · 08:46–18:15
you have to give a little energy at some point.
But maybe instead of trying to take the forces
imparted to you by the world and replacing them,
what we should be doing is letting the world push us around
and we go with the flow.
Very zen, very zen robot.
- Yeah, but okay, so that sounds very zen.
But I can also imagine how many
failed versions they had to go through.
I would say it's probably,
would you say it's in the thousands
that they've had to have the system fall down
before they figured out how to get--
- I don't know if it's thousands, but it's a lot.
It takes some patience, there's no question.
- So in that sense, control might help a little bit.
- Oh, I think everybody, even at the time,
said that the answer is to do that with control.
But it was just pointing out
that maybe the way we're doing control right now
isn't the way we should.
- Got it, so what about on the animal side?
The ones that figured out how to move efficiently?
Is there anything you find inspiring or beautiful
in the movement of any particular animal?
- I do have a favorite example.
- Okay.
(laughing)
- So it sort of goes with the passive walking idea.
So is there, how energy efficient are animals?
Okay, there's a great series of experiments
by George Lauder at Harvard and Mike Tranefilo at MIT.
They were studying fish swimming in a water tunnel, okay?
And one of these, the type of fish they were studying
were these rainbow trout,
because there was a phenomenon well understood
that rainbow trout,
when they're swimming upstream at mating season,
they kind of hang out behind the rocks.
And it looks like, I mean,
that's tiring work swimming upstream.
They're hanging out behind the rocks.
Maybe there's something energetically interesting there.
So they tried to recreate that.
They put in this water tunnel, a rock basically,
a cylinder that had the same sort of vortex street,
the eddies coming off the back of the rock
that you would see in a stream.
And they put a real fish behind this
and watched how it swims.
And the amazing thing is that if you watch from above
what the fish swims when it's not behind a rock,
it has a particular gate.
You can identify the fish the same way you look at a human
looking at walking down the street.
You sort of have a sense of how a human walks.
The fish has a characteristic gate.
You put that fish behind the rock, its gate changes.
And what they saw was that it was actually resonating
and kind of surfing between the vortices.
Now, here was the experiment that really was the clincher,
because there was still,
it wasn't clear how much of that was mechanics of the fish,
how much of that is control, the brain.
So the clincher experiment,
and maybe one of my favorites to date,
although there are many good experiments.
They took, this was now a dead fish.
They took a dead fish.
They put a string that went,
that tied the mouse of the fish to the rock.
So it couldn't go back and get caught in the grates.
And then they asked, what would that dead fish do
when it was hanging out behind the rock?
And so what you'd expect,
it sort of flopped around like a dead fish
in the vortex wake,
until something sort of amazing happens.
And this video is worth putting in.
- What happens?
- The dead fish basically starts swimming upstream.
It's completely dead, no brain, no motors, no control,
but it's somehow the mechanics of the fish
resonate with the vortex street,
and it starts swimming upstream.
It's one of the best examples ever.
- Who do you give credit for that to?
Is that just evolution constantly just figuring out
by killing a lot of generations of animals,
like the most efficient motion?
Or maybe the physics of our world completely,
is evolution applied not only to animals,
but just the entirety of it somehow drives to efficiency?
Like nature likes efficiency?
I don't know if that question even makes any sense.
- I understand the question.
I mean, do they co-evolve?
- Yeah, somehow co, yeah.
I don't know if an environment can evolve, but.
- I mean, there are experiments that people do,
careful experiments that show that animals
can adapt to unusual situations and recover efficiency.
So there seems like, at least in one direction,
I think there is reason to believe
that the animal's motor system,
and probably its mechanics,
adapt in order to be more efficient.
But efficiency isn't the only goal, of course.
Sometimes it's too easy to think about only efficiency.
But we have to do a lot of other things first,
not get eaten, and then all other things being equal,
try to save energy.
- By the way, let's draw a distinction
between control and mechanics.
Like how would you define each?
- Yeah, I mean, I think part of the point is that
we shouldn't draw a line as clearly as we tend to.
But on a robot, we have motors,
and we have the links of the robot, let's say.
If the motors are turned off,
the robot has some passive dynamics.
Gravity does the work.
You can put springs, I would call that mechanics.
If we have springs and dampers,
which our muscles are springs and dampers and tendons.
But then you have something that's doing active work,
putting energy in, which are your motors on the robot.
The controller's job is to send commands to the motor
that add new energy into the system.
So the mechanics and control interplay
somewhere the divide is around,
did you decide to send some commands to your motor,
or did you just leave the motors off,
let them do their work?
- So would you say is most of nature
on the dynamic side or the control side?
So like if you look at biological systems,
we're living in a pandemic now,
do you think a virus is a,
do you think it's a dynamic system,
or is there a lot of control, intelligence?
- I think it's both, but I think we maybe
have underestimated how important the dynamics are.
I mean even our bodies, the mechanics of our bodies,
certainly with exercise they evolve.
So I actually, I lost a finger in early 2000s,
and it's my fifth metacarpal.
And it turns out you use that a lot,
in ways you don't expect when you're opening jars,
even when I'm just walking around,
if I bump it on something,
there's a bone there that was used to taking contact.
My fourth metacarpal wasn't used to taking contact,
it used to hurt, still does a little bit.
But actually my bone has remodeled, right?
Over a couple of years, the geometry,
the mechanics of that bone changed
to address the new circumstances.
So the idea that somehow it's only our brain
that's adapting or evolving is not right.
- Maybe sticking on evolution for a bit,
'cause it's tended to create some interesting things.
Bipedal walking, why the heck did evolution give us,
I think we're, are we the only mammals that walk on two feet?
- No, I mean there's a bunch of animals that do it, a bit.
I think we are the most successful bipeds.
- I think I read somewhere that the reason
the evolution made us walk on two feet
is because there's an advantage to being able
to carry food back to the tribe or something like that.
So you can carry, it's kind of this communal,
cooperative thing, so to carry stuff back
to a place of shelter and so on to share with others.
- Do you understand at all the value of walking on two feet
from both a robotics and a human perspective?
- Yeah, there are some great books written
about evolution of, walking evolution of the human body.
I think it's easy though to make bad evolutionary arguments.
- Sure.
Most of them are probably bad, but what else can we do?
- I mean I think a lot of what dominated our evolution
probably was not the things that worked well
sort of in the steady state, when things are good.
But for instance, people talk about what we should eat now
because our ancestors were meat eaters or whatever.
- Oh yeah, I love that, yeah.
- But probably the reason that one pre-Homo sapien species
versus another survived was not because of
whether they ate well when there was lots of food,
but when the ice age came, probably one of them
happened to be in the wrong place,
one of them happened to forage a food that was okay
even when the glaciers came or something like that.
- There's a million variables that contributed
and we can't, and are actually, the amount of information
we're working with in telling these stories,
these evolutionary stories is very little.
So yeah, just like you said, it seems like,
if you study history, it seems like history turns
on these little events that otherwise would seem meaningless,
§03Transcript · 18:20–27:21
but in the grant, in retrospect, were turning points.
- Absolutely.
- And that's probably how, like somebody got hit
in the head with a rock because somebody slept
with the wrong person back in the cave days
and somebody got angry and that turned,
warring tribes combined with the environment,
all those millions of things and the meat eating,
which I get a lot of criticism because I don't know,
I don't know what your dietary processes are like,
but these days I've been eating only meat,
which is, there's a large community of people who say,
yeah, probably make evolutionary arguments
and say you're doing a great job.
There's probably an even larger community of people,
including my mom, who says it's a deeply unhealthy,
it's wrong, but I just feel good doing it.
But you're right, these evolutionary arguments
can be flawed, but is there anything interesting
to pull out for--
- There's a great book, by the way,
well, a series of books by Nicholas Taleb
about fooled by randomness and black swan.
Highly recommend them, but yeah,
they make the point nicely that probably it was
a few random events that, yes, maybe it was someone
getting hit by a rock, as you say.
- That said, do you think, I don't know how to ask
this question or how to talk about this,
but there's something elegant and beautiful
about moving on two feet.
Obviously biased, because I'm human,
but from a robotics perspective, too,
you work with robots on two feet.
Is it all useful to build robots that are on two feet
as opposed to four?
Is there something useful about it?
- I think the most, I mean, the reason I spent a long time
working on bipedal walking was because it was hard,
and it challenged control theory in ways
that I thought were important.
I wouldn't have ever tried to convince you
that you should start a company around bipeds
or something like this.
There are people that make pretty compelling arguments.
I think the most compelling one is that the world
is built for the human form, and if you want a robot
to work in the world we have today, then having a human form
is a pretty good way to go.
There are places that a biped can go that would be hard
for other form factors to go, even natural places.
But at some point in the long run, we'll be building
our environments for our robots, probably,
and so maybe that argument falls aside.
- So you famously run barefoot.
Do you still run barefoot?
- I still run barefoot.
- That's so awesome.
- Much to my wife's chagrin.
- Do you wanna make an evolutionary argument
for why running barefoot is advantageous?
What have you learned about human and robot movement
in general from running barefoot?
Human or robot and/or?
- Well, you know, it happened the other way, right?
So I was studying walking robots, and there's a great
conference called the Dynamic Walking Conference,
where it brings together both the biomechanics community
and the walking robots community.
And so I had been going to this for years and hearing
talks by people who study barefoot running
and other mechanics of running.
So I did eventually read Born to Run.
Most people read Born to Run and then--
- First then.
- Right?
The other thing I had going for me is actually that
I wasn't a runner before, and I learned to run
after I had learned about barefoot running,
or I mean, started running longer distances.
So I didn't have to unlearn.
And I'm definitely, I'm a big fan of it for me,
but I'm not gonna, I tend to not try to convince
other people.
There's people who run beautifully with shoes on,
and that's good.
But here's why it makes sense for me.
It's all about the long-term game, right?
So I think it's just too easy to run 10 miles,
feel pretty good, and then you get home at night
and you realize, my knees hurt.
I did something wrong, right?
If you take your shoes off, then if you hit hard
with your foot at all, then it hurts.
You don't like run 10 miles and then realize
you've done some damage.
You have immediate feedback telling you that you've done
something that's maybe suboptimal.
And you change your gait.
I mean, it's even subconscious.
If I, right now, having run many miles barefoot,
if I put a shoe on, my gait changes in a way
that I think is not as good.
So it makes me land softer.
And I think my goals for running are to do it
for as long as I can into old age, not to win any races.
And so for me, this is a way to protect myself.
- Yeah, I think, first of all, I've tried running barefoot
many years ago, probably the other way,
just reading Born to Run.
But just to understand, because I felt like I couldn't
put in the miles that I wanted to.
And it feels like running, for me, and I think
for a lot of people, was one of those activities
that we do often and we never really try to learn
to do correctly.
Like, it's funny, there's so many activities we do every day
like brushing our teeth, right?
I think a lot of us, at least me, probably have never
deeply studied how to properly brush my teeth, right?
Or wash, as now with the pandemic,
or how to properly wash our hands.
We do it every day, but we haven't really studied,
like, am I doing this correctly?
But running felt like one of those things
that it was absurd not to study how to do correctly
'cause it's the source of so much pain and suffering.
Like, I hate running, but I do it.
I do it because I hate it, but I feel good afterwards.
But I think it feels like you need to learn
how to do it properly, so that's where barefoot running
came in, and then I quickly realized that my gait
was completely wrong.
I was taking huge, like, steps and landing hard on the heel,
all those elements.
And so, yeah, from that, I actually learned
to take really small steps.
Look, I already forgot the number,
but I feel like it was 180 a minute or something like that.
And I remember I actually just took songs
that are 180 beats per minute,
and then, like, tried to run at that beat,
just to teach myself.
It took a long time, and I feel like after a while,
you learn to run, but you adjust it properly
without going all the way to barefoot.
I feel like barefoot is the legit way to do it.
I mean, I think a lot of people
would be really curious about it.
Can you, if they're interested in trying,
what would you, how would you recommend
they start or try or explore?
- Slowly. (laughs)
That's the biggest thing people do,
is they are excellent runners,
and they're used to running long distances or running fast,
and they take their shoes off, and they hurt themselves
instantly trying to do something
that they were used to doing.
I think I lucked out in the sense
that I couldn't run very far when I first started trying.
And I run with minimal shoes, too.
I mean, I will bring along a pair of,
actually, like, Aqua socks or something like this
I can just slip on, or running sandals.
I've tried all of them.
- What's the difference between a minimal shoe
and nothing at all?
What's, like, feeling-wise, what does it feel like?
- There is a, I mean, I notice my gait changing, right?
So, I mean, your foot has as many muscles and sensors
as your hand does, right?
- Sensors, ooh, okay.
- And we do amazing things with our hands.
And we stick our foot in a big, solid shoe, right?
So there's, I think, you know, when you're barefoot,
you're just giving yourself more proprioception.
And that's why you're more aware of some of the gait flaws
and stuff like this.
Now, you have less protection, too.
So--
- Rocks and stuff.
- I mean, yeah, so I think people who are afraid
of barefoot running, they're worried about getting cuts
or getting, stepping on rocks.
First of all, even if that was a concern,
I think those are all, like, very short-term.
You know, if I get a scratch or something,
it'll heal in a week.
If I blow out my knees, I'm done running forever.
So I will trade the short-term for the long-term, anytime.
But even then, you know, and this, again,
to my wife's chagrin, your feet get tough, right?
And--
- Your calves, okay.
- Yeah, I can run over almost anything now.
(laughing)
- I mean, what, maybe, can you talk about,
is there hint, like, is there tips or tricks
that you have, suggestions about,
§04Transcript · 27:24–37:11
like, if I wanted to try it?
- You know, there is a good book, actually.
There's probably more good books since I read them.
But Ken Bob, Barefoot Ken Bob Saxton.
He's an interesting guy.
But I think his book captures the right way
to describe running, barefoot running to somebody,
better than any other I've seen.
- So, you run pretty good distances, and you bike,
and is there, you know, if we talk about bucket list items,
is there something crazy on your bucket list,
athletically, that you hope to do one day?
- I mean, my commute is already a little crazy.
- What are we talking about here?
What distance are we talking about?
- Well, I live about 12 miles from MIT,
but you can find lots of different ways to get there.
So, I mean, I've run there for many years, I've biked there.
- Long ways?
- Yeah, but normally I would try to run in
and then bike home, bike in, run home.
- But you have run there and back before?
- Sure.
- Barefoot?
- Yeah, or with minimal shoes or whatever that--
- 12 times two?
- Yeah.
- Okay.
- It became kind of a game of how can I get to work.
I've rollerbladed, I've done all kinds of weird stuff,
but my favorite one these days,
I've been taking the Charles River to work.
So, I can put in a little rowboat, not so far from my house,
but the Charles River takes a long way to get to MIT.
So, I can spend a long time getting there.
And it's, you know, it's not about, I don't know,
it's just about, I've had people ask me,
how can you justify taking that time?
But for me, it's just a magical time to think,
to compress, decompress, you know,
especially I'll wake up, do a lot of work in the morning,
and then I kind of have to just let that settle
before I'm ready for all my meetings.
And then on the way home, it's a great time to load it,
sort of let that settle.
- You lead a large group of people.
I mean, is there days where you're like,
oh shit, I gotta get to work in an hour.
Like, I mean, is there a tension there?
And like, if we look at the grand scheme of things,
just like you said, long-term,
that meeting probably doesn't matter.
Like, you can always say, I'll run
and let the meeting happen how it happens.
Like, how do you, that zen,
what do you do with that tension
between the real world saying urgently,
you need to be there, this is important,
everything is melting down,
how are we gonna fix this robot?
There's this critical meeting,
and then there's this zen beauty of just running,
the simplicity of it, you along with nature.
What do you do with that?
- I would say I'm not a fast runner, particularly.
Probably my fastest splits ever
was when I had to get to daycare on time
because they were gonna charge me, you know,
some dollar per minute that I was late.
I've run some fast splits to daycare,
but those times are past now.
I think work, you can find a work-life balance in that way.
I think you just have to.
I think I am better at work
because I take time to think on the way in.
So I plan my day around it,
and I rarely feel that those are really at odds.
- So what, the bucket list item.
If we're talking 12 times two,
or approaching a marathon,
what, have you run an ultra marathon before?
Do you do races?
Is there, what's--
- Not to win.
- Not to-- (laughs)
- I'm not gonna take a dinghy across the Atlantic
or something if that's what you want,
but if someone does and wants to write a book,
I would totally read it
'cause I'm a sucker for that kind of thing.
No, I do have some fun things that I will try.
I like to, when I travel,
I almost always bike to the Logan Airport
and fold up a little folding bike
and then take it with me and bike to wherever I'm going.
And it's taken me,
or I'll take a stand-up paddleboard these days
on the airplane,
and then I'll try to paddle around
where I'm going or whatever.
And I've done some crazy things.
- But not for the,
I now talk, I don't know if you know
who David Goggins is by any chance.
- Not well, but yeah.
- But I talk to him now every day.
So he's the person who made me do this stupid challenge.
So he's insane, and he does things for the purpose,
in the best kind of way.
He does things like for the explicit purpose of suffering.
Like he picks the thing that,
like whatever he thinks he can do, he does more.
So is that, do you have that thing in you?
Or are you--
- I think it's become the opposite.
(Lex laughing)
- So you're like that dynamical system,
that the walker, the efficient--
- Yeah, it's leave no pain, right?
You should end feeling better than you started.
But it's mostly, I think,
and COVID has tested this 'cause I've lost my commute.
I think I'm perfectly happy walking around town
with my wife and kids if they could get them to go.
And it's more about just getting outside
and getting away from the keyboard for some time
just to let things compress.
- Let's go into robotics a little bit.
What to you is the most beautiful idea in robotics?
Whether we're talking about control,
or whether we're talking about optimization
and the math side of things,
or the engineering side of things,
or the philosophical side of things.
- I think I've been lucky to experience something
that not so many roboticists have experienced,
which is to hang out
with some really amazing control theorists.
And the clarity of thought
that some of the more mathematical control theory can bring
to even very complex, messy-looking problems
is really, it really had a big impact on me
and I had a day even just a couple of weeks ago
where I had spent the day on a Zoom robotics conference
having great conversations with lots of people.
Felt really good about the ideas
that were flowing and the like.
And then I had a late afternoon meeting
with one of my favorite control theorists
and we went from these abstract discussions
about maybes and what ifs and what a great idea,
to these super precise statements
about systems that aren't that much more simple
or abstract than the ones I care about deeply.
And the contrast of that is,
I don't know, it really gets me.
I think people underestimate
maybe the power of clear thinking.
And so for instance, deep learning is amazing.
I use it heavily in our work.
I think it's changed the world unquestionable.
It makes it easy to get things to work
without thinking as critically about it.
So I think one of the challenges as an educator
is to think about how do we make sure people get a taste
of the more rigorous thinking that I think goes along
with some different approaches.
- Yeah, so that's really interesting.
So understanding the fundamentals,
the first principles of the problem,
or in this case, it's mechanics,
like how a thing moves, how a thing behaves,
like all the forces involved,
like really getting a deep understanding of that.
I mean, from physics, the first principle thing
come from physics, and here it's literally physics.
Yeah, and this applies, in deep learning,
this applies to not just, I mean,
it applies so cleanly in robotics,
but it also applies to just in any data set.
I find this true, I mean, driving as well.
There's a lot of folks that work on autonomous vehicles
that don't study driving, like deeply.
I might be coming a little bit from the psychology side,
but I remember I spent a ridiculous number of hours
at lunch at this like lawn chair,
and I would sit somewhere in MIT's campus,
there's a few interesting intersections,
and we just watch people cross.
So we were studying pedestrian behavior,
and I felt like, I had to record a lot of video to try,
and then there's the computer vision
extracts their movement, how they move their head, and so on.
But like every time, I felt like I didn't understand enough.
I just, I felt like I wasn't understanding
what, how are people signaling to each other?
What are they thinking?
How cognizant are they of their fear of death?
Like, what's the underlying game theory here?
What are the incentives?
§05Transcript · 37:14–46:51
And then I finally found a live stream of an intersection
that's like high def that I just, I would watch,
so I wouldn't have to sit out there.
But that's interesting, so like, I feel--
- That's tough, that's a tough example,
because I mean, the learning--
- Humans are involved.
- Not just because human, but I think the learning mantra
is that basically the statistics of the data
will tell me things I need to know, right?
And for the example you gave of all the nuances
of eye contact or hand gestures or whatever
that are happening for these subtle interactions
between pedestrians and traffic, right?
Maybe the data will tell that story.
Maybe even one level more meta than what you're saying.
For a particular problem, I think it might be the case
that data should tell us the story.
But I think there's a rigorous thinking
that is just an essential skill for a mathematician
or an engineer that I just don't wanna lose it.
There are certainly super rigorous control,
or sorry, machine learning people.
I just think deep learning makes it so easy
to do some things that our next generation,
are not immediately rewarded for going through
some of the more rigorous approaches.
And then I wonder where that takes us.
Well, I'm actually optimistic about it.
I just want to do my part to try to steer
that rigorous thinking.
- So there's like two questions I wanna ask.
Do you have sort of a good example of rigorous thinking
where it's easy to get lazy and not do the rigorous thinking
and the other question I have is like,
do you have advice of how to practice rigorous thinking
in all the computer science disciplines
that we've mentioned?
- Yeah, I mean, there are times where problems
that can be solved with well-known mature methods
could also be solved with a deep learning approach
and there's an argument that you must use learning
even for the parts we already think we know
because if the human has touched it,
then you've biased the system
and you've suddenly put a bottleneck in there
that is your own mental model.
But something like inverting a matrix,
you know, I think we know how to do that pretty well,
even if it's a pretty big matrix
and we understand that pretty well
and you could train a deep network to do it,
but you shouldn't probably.
- So in that sense, rigorous thinking is understanding
the scope and the limitations of the methods that we have,
like how to use the tools of mathematics properly.
- Yeah, I think, you know, taking a class on analysis
is all I'm sort of arguing.
Take a chance to stop and force yourself
to think rigorously about even, you know,
the rational numbers or something,
you know, it doesn't have to be the end all problem,
but that exercise of clear thinking,
I think goes a long way
and I just want to make sure we keep preaching it.
- We don't lose it. - Yeah.
- But do you think when you're doing like rigorous thinking
or like maybe trying to write down equations
or sort of explicitly, like formally describe a system,
do you think we naturally simplify things too much?
Is that a danger you run into?
Like in order to be able to understand something
about the system mathematically,
we make it too much of a toy example.
- But I think that's the good stuff, right?
- That's how you understand the fundamentals?
- I think so.
I think maybe even that's a key to intelligence
or something, but I mean, okay,
what if Newton and Galileo had deep learning
and they had done a bunch of experiments
and they told the world,
"Here's your weights of your neural network.
"We've solved the problem."
- Yeah.
- Where would we be today?
I don't think we'd be as far as we are.
There's something to be said
about having the simplest explanation for a phenomenon.
So I don't doubt that we can train neural networks
to predict even physical,
F equals MA type equations,
but I maybe, I want another Newton to come along
'cause I think there's more to do
in terms of coming up with the simple models
for more complicated tasks.
- Yeah.
Let's not offend the AI systems from 50 years
from now that are listening to this
that are probably better at,
might be better at coming up
with F equals MA equations themselves.
- Oh, sorry, I actually think learning
is probably a route to achieving this,
but the representation matters, right?
And I think having a function
that takes my inputs to outputs
that is arbitrarily complex may not be the end goal.
I think there's still the most simple
or parsimonious explanation for the data.
Simple doesn't mean low dimensional.
That's one thing I think that we've,
a lesson that we've learned.
So a standard way to do model reduction
or system identification and controls
is to, the typical formulation is that you try
to find the minimal state dimension realization
of a system that hits some error bounds
or something like that.
And that's maybe not, I think we're learning
that that was, that state dimension
is not the right metric.
- Of complexity.
- Of complexity.
But for me, I think a lot about contact,
the mechanics of contact.
The robot hand is picking up an object or something.
And when I write down the equations of motion for that,
they look incredibly complex,
not because, actually not so much
because of the dynamics of the hand when it's moving,
but it's just the interactions
and when they turn on and off, right?
So having a high dimensional, you know,
but simple description of what's happening out here is fine.
But if, when I actually start touching,
if I write down a different dynamical system
for every polygon on my robot hand
and every polygon on the object,
whether it's in contact or not,
with all the combinatorics that explodes there,
then that's too complex.
So I need to somehow summarize that
with a more intuitive physics way of thinking.
And yeah, I'm very optimistic
that machine learning will get us there.
- First of all, I mean, I'll probably do it
in the introduction, but you're one
of the great robotics people at MIT.
You're a professor at MIT.
You've teach a lot of amazing courses.
You run a large group and you have a important history
for MIT, I think, as being a part
of the DARPA Robotics Challenge.
Can you maybe first say what is the DARPA Robotics Challenge
and then tell your story around it, your journey with it?
- Yeah, sure.
So the DARPA Robotics Challenge,
it came on the tails of the DARPA Grand Challenge
and DARPA Urban Challenge,
which were the challenges that brought us,
put a spotlight on self-driving cars.
Gil Pratt was at DARPA and pitched a new challenge
that involved disaster response.
It didn't explicitly require humanoids,
although humanoids came into the picture.
This happened shortly after the Fukushima disaster in Japan.
And our challenge was motivated roughly by that
because that was a case where if we had had robots
that were ready to be sent in,
there's a chance that we could have averted disaster.
And certainly after the, in the disaster response,
there were times where we would have loved
to have sent robots in.
So in practice, what we ended up with was a grand challenge,
a DARPA Robotics Challenge,
where Boston Dynamics was to make humanoid robots.
People like me and the amazing team at MIT
were competing first in a simulation challenge
to try to be one of the ones that wins the right
to work on one of the Boston Dynamics humanoids
in order to compete in the final challenge,
which was a physical challenge.
- And at that point, it was already,
so it was decided as humanoid robots early on.
- There were two tracks.
You could enter as a hardware team
where you brought your own robot,
or you could enter through the virtual robotics challenge
as a software team that would try to win the right
to use one of the Boston Dynamics robots.
- Which are called Atlas. - Atlas.
- Humanoid robots. - Yeah, it was a 400 pound
Marvel, but a pretty big, scary looking robot.
- Expensive too. - Expensive, yeah.
- Okay, so, I mean, how did you feel
the prospect of this kind of challenge?
I mean, it seems, you know, autonomous vehicles,
yeah, I guess that sounds hard,
but not really from a robotics perspective.
§06Transcript · 46:53–55:36
It's like, didn't they do it in the '80s
is the kind of feeling I would have
like when you first look at the problem.
It's on wheels, but like humanoid robots,
that sounds really hard.
So what, psychologically speaking,
what were you feeling, excited, scared?
Why the heck did you get yourself involved
in this kind of messy challenge?
- We didn't really know for sure
what we were signing up for,
in the sense that you could have had something that,
as it was described in the call for participation,
that could have put a huge emphasis
on the dynamics of walking and not falling down
and walking over rough terrain,
or the same description,
'cause the robot had to go into this disaster area
and turn valves and pick up a drill
and cut a hole through a wall.
It had to do some interesting things.
The challenge could have really highlighted perception
and autonomous planning,
or it ended up that, you know,
locomoting over a complex terrain
played a pretty big role in the competition.
So--
- And the degree of autonomy wasn't clear.
- The degree of autonomy was always
a central part of the discussion.
So what wasn't clear was how we would be able,
how far we'd be able to get with it.
So the idea was always that you want semi-autonomy,
that you want the robot to have enough compute
that you can have a degraded network link to a human.
And so the same way we had degraded networks
at many natural disasters,
you'd send your robot in,
you'd be able to get a few bits back and forth,
but you don't get to have enough,
potentially, to fully operate the robot,
every joint of the robot.
So, and then the question was,
and the gamesmanship of the organizers
was to figure out what we're capable of,
push us as far as we could,
so that it would differentiate the teams
that put more autonomy on the robot
and had a few clicks and just said,
"Go there, do this, go there, do this,"
versus someone who's picking every footstep
or something like that.
- So what were some memories,
painful, triumphant from the experience?
Like what was that journey?
Maybe if you can dig in a little deeper,
maybe even on the technical side, on the team side,
that whole process of,
from the early idea stages to actually competing.
- I mean, this was a defining experience for me.
It came at the right time for me in my career.
I had gotten tenure before I was due a sabbatical,
and most people do something relaxing
and restorative for a sabbatical.
- So you got tenure before this?
- Yeah, yeah, yeah.
It was a good time for me.
We had a bunch of algorithms that we were very happy with.
We wanted to see how far we could push them,
and this was a chance to really test our mettle,
to do more proper software engineering.
The team, we all just worked our butts off.
We're in that lab almost all the time.
Okay, so I mean, there were some, of course,
high highs and low lows throughout that,
anytime you're not sleeping and devoting your life
to a 400-pound humanoid.
I remember actually one funny moment
where we're all super tired,
and so Atlas had to walk across cinder blocks.
That was one of the obstacles.
And I remember Atlas was powered down,
hanging limp on its harness,
and the humans were there picking up
and laying the brick down
so that the robot could walk over it.
And I thought, "What is wrong with this?"
We've got a robot just watching us
do all the manual labor so that it can take its little
stroll across the terrain.
I mean, even the virtual robotics challenge
was super nerve-wracking and dramatic.
I remember, so we were using Gazebo as a simulator.
We were using Gazebo as a simulator on the cloud,
and there was all these interesting challenges.
I think the investment that OSRFC,
whatever they were called at that time,
Brian Gerke's team at Open Source Robotics,
they were pushing on the capabilities of Gazebo
in order to scale it to the complexity of these challenges.
So up to the virtual competition.
So the virtual competition was,
you will sign on at a certain time,
and we'll have a network connection
to another machine on the cloud
that is running the simulator of your robot.
And your controller will run on this computer,
and the physics will run on the other,
and you have to connect.
Now, the physics, they wanted it to run at real-time rates,
because there was an element of human interaction.
And humans, if you do want to tele-op,
it works way better if it's at frame rate.
- Oh, cool.
- But it was very hard to simulate
these complex scenes at real-time rate.
So right up to days before the competition,
the simulator wasn't quite at real-time rate.
And that was great for me,
because my controller was solving
a pretty big optimization problem,
and it wasn't quite at real-time rate.
So I was fine.
I was keeping up with the simulator.
We were both running at about 0.7.
And I remember getting this email,
and by the way, the perception folks on our team hated,
that they knew that if my controller was too slow,
the robot was gonna fall down.
And no matter how good their perception system was,
if I can't make my controller fast.
Anyways, we get this email
like three days before the virtual competition.
You know, it's for all the marbles.
We're gonna either get a humanoid robot or we're not.
And we get an email saying, "Good news.
"We made the robot, the simulator faster."
It's now 1.0.
And I was just like, "Oh man, what are we gonna do here?"
So that came in late at night for me.
- A few days ahead.
- A few days ahead.
I went over, it happened at Frank Permenter,
who's a very, very sharp,
he was a student at the time working on optimization.
He was still in lab.
Frank, we need to make
this quadratic programming solver faster.
Not like a little faster, it's actually, you know.
And we wrote a new solver for that QP,
together, that night.
- And you saw her. - It was terrifying.
- So there's a really hard optimization problem
that you're constantly solving.
You didn't make the optimization problem simpler?
You wrote a new solver?
- So, I mean, your observation is almost spot on.
What we did was what everybody,
I mean, people know how to do this,
but we had not yet done this idea of warm starting.
So we are solving a big optimization problem
at every time step.
But if you're running fast enough,
the optimization problem you're solving
on the last time step is pretty similar
to the optimization you're gonna solve with the next.
We, of course, had told our commercial solver
to use warm starting,
but even the interface to that commercial solver
was causing us these delays.
So what we did was we basically wrote,
we called it fast QP at the time.
We wrote a very lightweight, very fast layer,
which would basically check if nearby solutions
to the quadratic program,
which were very easily checked, could stabilize the robot.
And if they couldn't, we would fall back to the solver.
- You couldn't really test this well, right?
Or like?
- So we always knew that if we fell back,
if we, it got to the point where if for some reason
things slowed down and we fell back to the original solver,
the robot would actually literally fall down.
So it was a harrowing sort of ledge we were sort of on.
But I mean, actually, like the 400 pound humanoid
could come crashing to the ground
if your solver's not fast enough.
But we had lots of good experiences.
- Can I ask you a weird question I get
about the idea of hard work?
So actually people, like students of yours
that I've interacted with,
and just, and robotics people in general,
but they have moments, at moments have worked harder
than most people I know in terms of,
if you look at different disciplines
of how hard people work.
But they're also like the happiest.
Like, just like, I don't know.
It's the same thing with like running,
§07Transcript · 55:39–1:04:06
people that push themselves to like the limit,
they also seem to be like the most like full of life somehow.
And I get often criticized, like,
"You're not getting enough sleep.
"What are you doing to your body?"
Blah, blah, blah, like this kind of stuff.
And I usually just kind of respond like,
"I'm doing what I love, I'm passionate about it.
"I love it, I feel like it's invigorating."
I actually think, I don't think the lack of sleep
is what hurts you.
I think what hurts you is stress
and lack of doing things that you're passionate about.
But in this world, yeah, I mean, can you comment about
why the heck robotics people are
(laughs)
willing to push themselves to that degree?
Is there value in that?
And why are they so happy?
- I think you got it right.
I mean, I think the causality is not that we work hard.
And I think other disciplines work very hard too.
But I don't think it's that we work hard
and therefore we are happy.
I think we found something
that we're truly passionate about.
It makes us very happy.
And then we get a little involved with it
and spend a lot of time on it.
What a luxury to have something
that you wanna spend all your time on, right?
- We could talk about this for many hours,
but maybe if we could pick,
is there something on the technical side
on the approach that you took that's interesting
that turned out to be a terrible failure
or a success that you carry into your work today
about all the different ideas that were involved
in making, whether in the simulation or in the real world,
making the semi-autonomous system work?
- I mean, it really did teach me something fundamental
about what it's gonna take to get robustness
out of a system of this complexity.
I would say the DARPA challenge
really was foundational in my thinking.
I think the autonomous driving community thinks about this.
I think lots of people thinking about safety critical
systems that might have machine learning in the loop
are thinking about these questions.
For me, the DARPA challenge was the moment where I realized,
we've spent every waking minute running this robot.
And again, for the physical competition,
days before the competition,
we saw the robot fall down in a way
it had never fallen down before.
I thought, how could we have found that?
We only have one robot, it's running almost all the time.
We just didn't have enough hours in the day
to test that robot.
Something has to change, right?
And I think that, I mean, I would say that the team that won
from KAIST was the team that had two robots
and was able to do not only incredible engineering,
just absolutely top-rate engineering,
but also they were able to test at a rate
and discipline that we didn't keep up with.
- What does testing look like?
What are we talking about here?
Like what's a loop of tests?
Like from start to finish, what is a loop of testing?
- Yeah, I mean, I think there's a whole philosophy
to testing.
There's the unit tests, and you can do that on a hardware.
You can do that in a small piece of code.
You write one function, you should write a test
that checks that function's input and outputs.
You should also write an integration test
at the other extreme of running the whole system together,
where they try to turn on all of the different functions
that you think are correct.
It's much harder to write the specifications
for a system-level test,
especially if that system is as complicated
as a humanoid robot, but the philosophy is sort of the same.
On the real robot, it's no different,
but on a real robot,
it's impossible to run the same experiment twice.
So if you see a failure,
you hope you caught something in the logs
that tell you what happened,
but you'd probably never be able to run exactly
that experiment again.
And right now, I think our philosophy is just,
basically, Monte Carlo estimation,
is just run as many experiments as we can,
maybe try to set up the environment
to make the things we are worried about
happen as often as possible,
but really we're relying on somewhat random search
in order to test.
Maybe that's all we'll ever be able to,
but I think, you know, 'cause there's an argument
that the things that'll get you
are the things that are really nuanced in the world,
and it'd be very hard to, for instance,
put back in a simulation.
- Yeah, I guess the edge cases.
What was the hardest thing?
Like, so you said walking over rough terrain,
like just taking footsteps.
I mean, people, it's so dramatic and painful
in a certain kind of way to watch these videos
from the DRC of robots falling.
- Yep.
- It's just so heartbreaking.
I don't know.
Maybe it's because, for me at least,
we anthropomorphize the robot.
Of course, it's also funny for some reason.
Like, humans falling is funny.
It's some dark reason.
I'm not sure why it is so,
but it's also like tragic and painful.
And so speaking of which,
I mean, what made the robots fall and fail in your view?
- So I can tell you exactly what happened.
I contributed one of those.
Our team contributed one of those spectacular falls.
Every one of those falls has a complicated story.
I mean, at one time,
the power effectively went out on the robot
because it had been sitting at the door
waiting for a green light to be able to proceed
and its batteries, you know,
and therefore it just fell backwards
and smashed its head to the ground.
And it was hilarious,
but it wasn't because of bad software, right?
But for ours, so the hardest part of the challenge,
the hardest task in my view
was getting out of the Polaris.
It was actually relatively easy to drive the Polaris.
- Can you tell the story?
Sorry to interrupt. - No, of course.
- The story of the car.
People should watch this video.
I mean, the thing you've come up with is just brilliant,
but anyway, sorry.
- Yeah, we kind of joke,
we call it the big robot little car problem
because somehow the race organizers
decided to give us a 400 pound humanoid
and that they also provided the vehicle,
which was a little Polaris.
And the robot didn't really fit in the car.
So you couldn't drive the car with your feet
under the steering column.
We actually had to straddle the main column
of the, and have basically one foot in the passenger seat,
one foot in the driver's seat
and then drive with our left hand.
But the hard part was we had to then park the car,
get out of the car.
It didn't have a door, that was okay,
but it's just getting up from crouch, from sitting
when you're in this very constrained environment.
- First of all, I remember after watching those videos,
I was much more cognizant of how hard is it,
it is for me to get in and out of the car
and out of the car especially.
Like it's actually a really difficult control problem.
- Yeah.
- And I'm very cognizant of it when I'm like injured
for whatever reason. - No, it's really hard.
- Yeah.
So how did you approach this problem?
- So we had, you think of NASA's operations
and they have these checklists, pre-launch checklists
and they're like, we weren't far off from that.
We had this big checklist and on the first day
of the competition, we were running down our checklist.
And one of the things we had to do,
we had to turn off the controller,
the piece of software that was running
that would drive the left foot of the robot
in order to accelerate on the gas.
And then we turned on our balancing controller.
And the nerves, jitters of the first day of the competition,
someone forgot to check that box
and turn that controller off.
So we used a lot of motion planning to figure out
a sort of configuration of the robot
that we could get up and over.
We relied heavily on our balancing controller.
And basically there were, when the robot was in one
of its most precarious sort of configurations,
trying to sneak its big leg out of the side,
the other controller that thought it was still driving
told its left foot to go like this.
And that wasn't good, but it turned disastrous for us
§08Transcript · 1:04:11–1:13:46
because what happened was a little bit of push here.
Actually, we have videos of us running into the robot
with a 10 foot pole and it kind of will recover.
But this is a case where there's no space to recover.
So a lot of our secondary balancing mechanisms
about like take a step to recover,
they were all disabled because we were in the car
and there's no place to step.
So we were relying on our just lowest level reflexes.
And even then, I think just hitting the foot on the seat,
on the floor, we probably could have recovered from it.
But the thing that was bad that happened
is when we did that and we jostled a little bit,
the tailbone of our robot was only a little off the seat,
it hit the seat.
And the other foot came off the ground just a little bit.
And nothing in our plans had ever told us what to do
if your butt's on the seat and your feet are in the air.
- Feet in the air.
- And then the thing is, once you get off the script,
things can go very wrong because even our state estimation,
our system that was trying to collect all the data
from the sensors and understand
what's happening with the robot,
it didn't know about this situation.
So it was predicting things that were just wrong.
And then we did a violent shake and fell off
in our face first out of the robot.
- But like into the destination.
- That's true, we fell in and we got our point for egress.
- But so is there any hope for, that's interesting,
is there any hope for Atlas to be able to do something
when it's just on its butt and feet in the air?
- Absolutely.
No, so that is one of the big challenges.
And I think it's still true, you know,
Boston Dynamics and Andy Mow and there's this incredible
work on legged robots happening around the world.
Most of them still are very good at the case
where you're making contact with the world at your feet.
And they have typically point feet relatively,
their balls on their feet, for instance.
If those robots get in a situation where the elbow
hits the wall or something like this,
that's a pretty different situation.
Now they have layers of mechanisms that will make,
I think the more mature solutions have ways in which
the controller won't do stupid things.
But a human for instance, is able to leverage
incidental contact in order to accomplish a goal.
In fact, I might, if you push me, I might actually
put my hand out and make a new brand new contact.
The feet of the robot are doing this on quadrupeds,
but we mostly in robotics are afraid of contact
on the rest of our body, which is crazy.
There's this whole field of motion planning,
collision free motion planning.
And we write very complex algorithms so that the robot
can dance around and make sure it doesn't touch the world.
- So people are just afraid of contact
'cause contact is seen as a difficult.
- It's still a difficult control problem
and sensing problem.
- Now you're a serious person.
(laughs)
- I'm a little bit of an idiot
and I'm going to ask you some dumb questions.
So I do martial arts.
So like jujitsu, I wrestled my whole life.
So let me ask the question,
whenever people learn that I do any kind of AI
or like I mentioned robots and things like that,
they say when are we gonna have robots
that can win in a wrestling match
or in a fight against a human?
So we just mentioned sitting on your butt,
if you're in the air, that's a common position,
jujitsu when you're on the ground,
you're a down opponent.
Like how difficult do you think is the problem
and when will we have a robot that can defeat a human
in a wrestling match?
And we're talking about a lot,
I don't know if you're familiar with wrestling,
but essentially--
- Not very.
- It's basically the art of contact.
It's like, 'cause you're picking contact points
and then using like leverage,
like to off balance, to trick people.
It's like you make them feel like you're doing one thing
and then they change their balance
and then you switch what you're doing
and then results in a throw or whatever.
So it's basically the art of multiple contacts.
- Awesome, that's a nice description of it.
So there's also an opponent in there, right?
So if--
- Very dynamic.
- Right, if you are wrestling a human
and are in a game theoretic situation with a human,
that's still hard.
But just to speak to the quickly reasoning
about contact part of it, for instance.
- Yeah, maybe even throwing the game theory out of it,
almost like a, yeah, almost like a non-dynamic opponent.
- Right.
There's reasons to be optimistic,
but I think our best understanding of those problems
are still pretty hard.
I have been increasingly focused on manipulation,
partly where that's a case where the contact
has to be much more rich.
And there are some really impressive examples
of deep learning policies, controllers,
that can appear to do good things through contact.
We've even got new examples of deep learning models
of predicting what's gonna happen to objects
as they go through contact.
But I think the challenge you just offered there
still eludes us, right?
The ability to make a decision
based on those models quickly.
I have to think though, it's hard for humans too,
when you get that complicated.
I think probably you had maybe a slow motion version
of where you learn the basic skills,
and you've probably gotten better at it,
and there's much more subtlety.
But it might still be hard to actually,
really on the fly, take a model of your humanoid
and figure out how to plan the optimal sequence.
That might be a problem we never solve.
- Well, I mean, one of the most amazing things to me
about the, we could talk about martial arts,
we could also talk about dancing,
doesn't really matter, too human.
I think it's the most interesting study of contact.
It's not even the dynamic element of it.
Like when you get good at it, it's so effortless.
Like I can just, I'm very cognizant
of the entirety of the learning process
being essentially like learning how to move my body
in a way that I could throw very large weights around
effortlessly.
And I can feel the learning.
Like I'm a huge believer in drilling of techniques.
And you can just like feel your,
you're not feeling, you're feeling, sorry,
you're learning it intellectually a little bit,
but a lot of it is the body learning it somehow,
like instinctually.
And whatever that learning is, that's really,
I'm not even sure if that's equivalent
to like a deep learning, learning a controller.
I think it's something more,
it feels like there's a lot of distributed learning
going on.
- Yeah, I think there's hierarchy and composition
probably in the systems
that we don't capture very well yet.
You have layers of control systems,
you have reflexes at the bottom layer,
and you have a system that's capable of planning a vacation
to some distant country,
which is probably,
you probably don't have a controller or a policy
for every possible destination you'll ever pick, right?
But there's something magical in the in-between.
And how do you go from these low level feedback loops
to something that feels like a pretty complex
set of outcomes?
My guess is, I think there's evidence
that you can plan at some of these levels, right?
So Josh Tenenbaum just showed it in his talk the other day.
He's got a game he likes to talk about,
I think he calls it the pick three game or something,
where he puts a bunch of clutter down in front of a person,
and he says, "Okay, pick three objects."
And it might be a telephone or a shoe
or a Kleenex box or whatever.
And apparently you pick three items and then you pick,
he says, "Okay, pick the first one up with your right hand,
the second one up with your left hand.
Now using those objects,
now as tools, pick up the third object."
Right, so that's down at the level of physics and mechanics
and contact mechanics that I think we do learning,
or we do have policies for,
we do control for, almost feedback.
But somehow we're able to still,
I mean, I've never picked up a telephone
with a shoe and a water bottle before,
and somehow, and it takes me a little longer
to do that the first time,
but most of the time we can sort of figure that out.
So, yeah, I think the amazing thing
is this ability to be flexible with our models,
plan when we need to,
use our well-oiled controllers when we don't,
§09Transcript · 1:13:49–1:23:30
when we're in familiar territory.
Having models, I think the other thing you just said
was something about,
I think your awareness of what's happening
is even changing as you improve your expertise, right?
So maybe you have a very approximate model
of the mechanics to begin with,
and as you gain expertise,
you get a more refined version of that model.
You're aware of muscles or balanced components
that you just weren't even aware of before.
So how do you scaffold that?
- Yeah, plus the fear of injury,
the ambition of goals, of excelling,
and fear of mortality.
Let's see what else is in there as motivations.
Overinflated ego in the beginning,
and then a crash of confidence in the middle.
All of those seem to be essential for the learning process.
And if all that's good,
then you're probably optimizing energy efficiency.
- Yeah, right, so we have to get that right.
So there was this idea that you would have robots play soccer
better than human players by 2050.
That was the goal.
Basically, was the goal to beat world champion team,
to become a World Cup,
beat like a World Cup level team?
So are we gonna see that first,
or a robot, if you're familiar,
there's an organization called UFC for mixed martial arts.
Are we gonna see a World Cup championship soccer team
that have robots,
or a UFC champion mixed martial artist that's a robot?
- I mean, it's very hard to say one thing is harder,
some problem's harder than the other.
What probably matters is who started the organization.
I mean, I think RoboCup has a pretty serious following,
and there is a history now of people playing that game,
learning about that game,
building robots to play that game,
building increasingly more human robots.
It's got momentum.
So if you want to have mixed martial arts compete,
you better start your organization now, right?
I think almost independent of which problem
is technically harder,
'cause they're both hard and they're both different.
- That's a good point.
I mean, those videos are just hilarious.
Like, especially the humanoid robots trying to play soccer.
I mean, they're kind of terrible right now.
- I mean, I guess there is RoboSumo wrestling.
There's like the RoboOne competitions
where they do have these robots that go on a table
and basically fight.
So maybe I'm wrong.
- First of all, do you have a year in mind for RoboCup,
just from a robotics perspective?
Seems like a super exciting possibility.
- That, like in the physical space,
this is what's interesting.
I think the world is captivated.
I think it's really exciting.
It inspires just a huge number of people
when a machine beats a human
at a game that humans are really damn good at.
So you're talking about chess and Go,
but that's in the world of digital.
I don't think machines have beat humans
at a game in the physical space yet,
but that would be just--
- You have to make the rules very carefully, right?
I mean, if Atlas kicked me in the shins, I'm down,
and game over.
So it's very subtle on what's fair.
- I think the fighting one is a weird one, yeah,
'cause you're talking about a machine
that's much stronger than you.
But yeah, in terms of soccer, basketball,
all those kinds. - Even soccer, right?
I mean, as soon as there's contact or whatever,
and there are some things that the robot will do better.
I think if you really set yourself up to try to see
could robots win the game of soccer
as the rules were written,
the right thing for the robot to do
is to play very differently than a human would play.
You're not gonna get the perfect soccer player robot.
You're gonna get something that exploits the rules,
exploits its super actuators,
its super low bandwidth feedback loops or whatever,
and it's gonna play the game differently
than you want it to play.
And I bet there's ways, I bet there's loopholes, right?
We saw that in the DARPA challenge,
that it's very hard to write a set of rules
that someone can't find a way to exploit.
- Let me ask another ridiculous question.
I think this might be the last ridiculous question, but--
- I doubt it.
(laughing)
- I aspire to ask as many ridiculous questions
of a brilliant MIT professor.
Okay, I don't know if you've seen The Black Mirror.
- It's funny, I never watched the episode.
I know when it happened, though,
because I gave a talk to some MIT faculty one day,
on an unassuming Monday or whatever,
I was telling them about the state of robotics.
And I showed some video from Boston Dynamics
of the quadruped spot at the time.
It was their early version of spot.
And there was a look of horror that went across the room.
And I said, "I've shown videos like this a lot of times.
"What happened?"
And it turns out that this video had,
this Black Mirror episode had changed the way people watched
the videos I was putting out.
- The way they see these kinds of robots.
So I talked to so many people who are just terrified
because of that episode probably of these kinds of robots.
I almost want to say that they almost enjoy being terrified.
I don't even know what it is about human psychology
that kind of imagine doomsday,
the destruction of the universe or our society,
and kind of enjoy being afraid.
I don't want to simplify it,
but it feels like they talk about it so often,
it almost, there does seem to be an addictive quality to it.
I talked to a guy, a guy named Joe Rogan,
who's kind of the flag bearer
for being terrified of these robots.
Do you have a, two questions.
One, do you have an understanding
of why people are afraid of robots?
And the second question is, in Black Mirror,
just to tell you the episode,
I don't even remember it that much anymore,
but these robots, I think they can shoot
like a pellet or something.
They basically have, it's basically a spot with a gun.
And how far are we away from having robots
that go rogue like that?
You know, basically spot that goes rogue for some reason
and somehow finds a gun.
- Right, so, I mean, I'm not a psychologist.
I think, I don't know exactly why people react
the way they do.
I think we have to be careful
about the way robots influence our society and the like.
I think that's something, that's a responsibility
that roboticists need to embrace.
I don't think robots are gonna come after me
with a kitchen knife or a pellet gun right away.
And I mean, if they were programmed in such a way,
but I used to joke with Atlas
that all I had to do was run for five minutes
and its battery would run out.
But actually they've got a very big battery
in there by the end.
So it was over an hour.
I think the fear is a bit cultural though.
'Cause I mean, you notice that,
like I think in my age in the US,
we grew up watching Terminator, right?
If I had grown up at the same time in Japan,
I probably would have been watching Astro Boy.
And there's a very different reaction to robots
in different countries, right?
So I don't know if it's a human innate fear
of metal marvels,
or if it's something that we've done to ourselves
with our sci-fi.
- Yeah, the stories we tell ourselves through movies,
through just, through popular media.
But if I were to tell,
if you were my therapist and I said,
"I'm really terrified that we're going to have these robots
very soon that will hurt us."
Like, how do you approach making me feel better?
Like, why shouldn't people be afraid?
There's a, I think there's a video that went viral recently.
Everything was spot in Boston,
then it goes viral in general.
But usually it's like really cool stuff.
Like they're doing flips and stuff,
or like sad stuff,
Atlas being hit with a broomstick or something like that.
But there's a video where I think
one of the new productions bought robots,
which are awesome.
It was like patrolling somewhere in some country.
And people immediately were saying,
this is the dystopian future, the surveillance state.
For some reason, you can just have a camera.
Something about spot, being able to walk on four feet
with like really terrified people.
So what do you say to those people?
I think there is a legitimate fear there
§10Transcript · 1:23:33–1:34:01
because so much of our future is uncertain.
But at the same time, technically speaking,
it seems like we're not there yet.
So what do you say?
- I mean, I think technology is complicated.
It can be used in many ways.
I think there are purely software attacks
somebody could use to do great damage.
Maybe they have already.
I think wheeled robots could be used in bad ways too.
- Drones. - Drones, right?
I don't think that, let's see.
I don't want to be building technology
just because I'm compelled to build technology
and I don't think about it.
But I would consider myself a technological optimist,
I guess, in the sense that I think we should continue
to create and evolve and our world will change.
And if we will introduce new challenges,
we'll screw something up maybe.
But I think also we'll invent ourselves
out of those challenges and life will go on.
- So it's interesting 'cause you didn't mention
this is technically too hard.
- I don't think robots are,
I think people attribute a robot that looks like an animal
as maybe having a level of self-awareness
or consciousness or something that they don't have yet.
I think our ability to anthropomorphize those robots
is probably, we're assuming that they have
a level of intelligence that they don't yet have
and that might be part of the fear.
So in that sense, it's too hard.
But there are many scary things in the world, right?
So I think we're right to ask those questions,
we're right to think about the implications of our work.
- Right, in the short term as we're working on it, for sure.
Is there something long-term that scares you
about our future with AI and robots?
A lot of folks from Elon Musk to Sam Harris
to a lot of folks talk about the existential threats
about artificial intelligence.
Oftentimes robots kind of inspire that the most
because of the anthropomorphism.
Do you have any fears?
- It's an important question.
I actually, I think I like Rod Brooks answer
maybe the best on this.
I think, and it's not the only answer he's given
over the years, but maybe one of my favorites is,
he says, it's not gonna be,
he's got a book, "Flesh and Machines," I believe.
It's not gonna be the robots versus the people.
We're all gonna be robot people.
Because we already have smartphones,
some of us have serious technology
implanted in our bodies already,
whether we have a hearing aid or a pacemaker
or anything like this.
People with amputations might have prosthetics.
That's a trend I think that is likely to continue.
I mean, this is now wild speculation.
But I mean, when do we get to cognitive implants
and the like?
Yeah, with neural link, brain-computer interfaces.
That's interesting.
So there's a dance between humans and robots.
It's going to be,
it's going to be impossible to be scared
of the other out there, the robot,
because the robot will be part of us, essentially.
It'd be so intricately sort of part of our society.
- Yeah, and it might not even be implanted part of us,
but just it's so much a part of our society.
- So in that sense, the smartphone is already the robot
we should be afraid of, yeah.
I mean, yeah, and all the usual fears arise
of the misinformation, the manipulation,
all those kinds of things that,
the problems are all the same.
They're human problems, essentially, it feels like.
- Yeah, I mean, I think the way we interact
with each other online is changing the value we put on
personal interaction, and that's a crazy big change
that's going to happen and has already been ripping
through our society, right?
And that has implications that are massive.
I don't know if they should be scared of it
or go with the flow, but I don't see
some battle lines between humans and robots
being the first thing to worry about.
- I mean, I do want to just, as a kind of comment,
maybe you can comment about your just feelings
about Boston Dynamics in general,
but I love science, I love engineering.
I think there's so many beautiful ideas in it.
And when I look at Boston Dynamics
or legged robots in general, I think they inspire people
curiosity and feelings in general, excitement
about engineering more than almost anything else
in popular culture.
And I think that's such an exciting responsibility
and possibility for robotics.
And Boston Dynamics is riding that wave pretty damn well.
Like they found it, they've discovered that hunger
and curiosity in the people, and they're doing magic with it.
I don't care if they, I mean, I guess it's their company,
they have to make money, right?
But they're already doing incredible work
in inspiring the world about technology.
I mean, do you have thoughts about Boston Dynamics
and maybe others, your own work in robotics
and inspiring the world in that way?
- I completely agree.
I think Boston Dynamics is absolutely awesome.
I think I show my kids those videos,
and the best thing that happens is sometimes
they've already seen them, right?
I think, I just think it's a pinnacle of success
in robotics that is just one of the best things
that's happened.
I absolutely, completely agree.
- One of the heartbreaking things to me
is how many robotics companies fail.
How hard it is to make money with the robotics company.
Like iRobot like went through hell just to arrive at Arumba
to figure out one product.
And then there's so many home robotics companies
like Jibo and Anki, the cutest toy.
There's a great robot I thought went down.
I'm forgetting a bunch of them,
but a bunch of robotics companies fail.
Rod's company, Rethink Robotics.
Like, do you have anything hopeful to say
about the possibility of making money with robots?
- Oh, I think you can't just look at the failures.
You can, I mean, Boston Dynamics is a success.
There's lots of companies that are still
doing amazingly good work in robotics.
I mean, this is the capitalist ecology or something, right?
I think you have many companies, you have many startups,
and they push each other forward, and many of them fail,
and some of them get through,
and that's sort of the natural--
- Way of things.
- Way of those things.
I don't know that, is robotics really that much worse?
I feel the pain that you feel too.
Every time I read one of these, sometimes it's friends,
and I definitely wish it went better, went differently.
But I think it's healthy and good to have bursts of ideas,
bursts of activities, ideas.
If they are really aggressive, they should fail sometimes.
Certainly that's the research mantra, right?
If you're succeeding at every problem you attempt,
then you're not choosing aggressively enough.
- Is it exciting to you, the new Spot?
- Oh, it's so good.
- When are you getting him as a pet, or it?
- Yeah, I mean, I have to dig up 75K right now.
(laughing)
- I mean, it's so cool that there's a price tag.
You can go and then actually buy it.
- I have a Skydio R1, love it.
So, no, I would absolutely be a customer.
- I wonder what your kids would think about it.
- I actually, Zach from Boston Dynamics
would let my kid drive in one of their demos one time,
and that was just so good, so good.
So, I'll forever be grateful for that.
- And there's something magical
about the anthropomorphization of that arm.
It adds another level of human connection.
I'm not sure we understand from a control aspect
the value of anthropomorphization.
I think that's an understudied
and underunderstood engineering problem.
There's been a, psychologists have been studying it.
I think it's part, like, manipulating our mind
to believe things is a valuable engineering.
Like, this is another degree of freedom
that can be controlled.
- I like that, yeah, I think that's right.
I think there's something that humans seem to do,
or maybe my dangerous introspection is,
I think we are able to make very simple models
that assume a lot about the world very quickly,
and then it takes us a lot more time, like your wrestling.
You probably thought you knew
what you were doing with wrestling,
and you were fairly functional as a complete wrestler,
and then you slowly got more expertise.
So maybe it's natural that our first level of defense
against seeing a new robot is to think of it
in our existing models of how humans and animals behave.
And it's just, as you spend more time with it,
then you'll develop more sophisticated models
that will appreciate the differences.
- Exactly.
Can you say what does it take to control a robot?
§11Transcript · 1:34:04–1:42:50
Like, what is the control problem of a robot?
And in general, what is a robot in your view?
Like, how do you think of this system?
- What is a robot?
- What is a robot?
- I think robotics-- - I told you
ridiculous questions.
- No, no, it's good.
I mean, there's standard definitions
of combining computation with some ability
to do mechanical work.
I think that gets us pretty close.
But I think robotics has this problem
that once things really work,
we don't call them robots anymore.
Like, my dishwasher at home is pretty sophisticated,
beautiful mechanisms.
There's actually a pretty good computer,
probably a couple of chips in there doing amazing things.
We don't think of that as a robot anymore,
which isn't fair, 'cause then,
roughly it means that robotics always has to
solve the next problem
and doesn't get to celebrate its past successes.
I mean, even factory room floor robots
are super successful.
They're amazing.
But that's not the ones,
I mean, people think of them as robots,
but they don't, if you ask what are the successes
of robotics, somehow it doesn't come
to your mind immediately.
- So the definition of robot is a system
with some level of automation that fails frequently.
- Something like, it's the computation
plus mechanical work and an unsolved problem.
(laughing)
- Unsolved problem, yeah.
So from a perspective of control
and mechanics, dynamics, what is a robot?
- So there are many different types of robots.
The control that you need for a Jibo robot,
some robot that's sitting on your countertop
and interacting with you,
but not touching you, for instance,
is very different than what you need for an autonomous car
or an autonomous drone.
It's very different than what you need for a robot
that's gonna walk or pick things up with its hands.
My passion has always been for the places
where you're interacting more,
you're doing more dynamic interactions with the world.
So walking, now manipulation.
And the control problems there are beautiful.
I think contact is one thing that differentiates them
from many of the control problems we've solved classically.
Right, like modern control grew up stabilizing fighter jets
that were passively unstable
and there's like amazing success stories
from control all over the place.
Power grid, I mean, there's all kinds of,
it's everywhere that we don't even realize,
just like AI is now.
- So you mentioned contact, like what's contact?
- So an airplane is an extremely complex system
or a spacecraft landing or whatever,
but at least it has the luxury
of things change relatively continuously.
That's an oversimplification.
But if I make a small change
in the command I send to my actuator,
then the path that the robot will take
tends to change only by a small amount.
- And there's a feedback mechanism here.
That's what we're talking about.
- And there's a feedback mechanism.
And thinking about this as locally,
like a linear system, for instance,
I can use more linear algebra tools
to study systems like that,
generalizations of linear algebra
to these smooth systems.
What is contact?
A robot has something very discontinuous
that happens when it makes or breaks,
when it starts touching the world.
And even the way it touches or the order of contacts
can change the outcome in potentially unpredictable ways.
Not unpredictable, but complex ways.
I do think there's a little bit of a,
a lot of people will say that contact is hard in robotics,
even to simulate.
And I think there's a little bit of a,
there's truth to that,
but maybe a misunderstanding around that.
So what is limiting is that when we think about our robots
and we write our simulators,
we often make an assumption that objects are rigid.
And when it comes down,
you know, that their mass moves all,
you know, it stays in a constant position
relative to each other itself.
And that leads to some paradoxes
when you go to try to talk about
rigid body mechanics and contact.
And so for instance,
if I have a three-legged stool with just a,
imagine it comes to a point at the legs.
So it's only touching the world at a point.
If I draw my physics,
my high school physics diagram of this system,
then there's a couple of things that I'm given
by elementary physics.
I know if the system, if the table is at rest,
if it's not moving, it's zero velocities.
That means that the normal force,
all the forces are in balance.
So the force of gravity is being countered
by the forces that the ground is pushing on my table legs.
I also know since it's not rotating
that the moments have to balance.
And since it can, it's a three-dimensional table,
it could fall in any direction.
It actually tells me uniquely
what those three normal forces have to be.
If I have four legs on my table, four-legged table,
and they were perfectly machined
to be exactly the right, same height,
and they're set down and the table's not moving,
then the basic conservation laws don't tell me
there are many solutions for the forces
that the ground could be putting on my legs
that would still result in the table not moving.
Now, the reason, that seems fine.
I could just pick one.
But it gets funny now because if you think about friction,
what we think about with friction is our standard model says
the amount of force that the table will push back,
if I were to now try to push my table sideways,
I guess I have a table here,
is proportional to the normal force.
So if I'm barely touching and I push, I'll slide,
but if I'm pushing more and I push, I'll slide less.
It's called Coulomb friction, is our standard model.
Now, if you don't know what the normal force is
on the four legs and you push the table,
then you don't know what the friction forces are gonna be.
And so you can't actually tell,
the laws just aren't explicit yet
about which way the table's gonna go.
It could veer off to the left,
it could veer off to the right,
it could go straight.
So the rigid body assumption of contact
leaves us with some paradoxes,
which are annoying for writing simulators
and for writing controllers.
We still do that sometimes because soft contact
is potentially harder numerically or whatever,
and the best simulators do both
or do some combination of the two.
But anyways, because of these kinds of paradoxes,
there's all kinds of paradoxes in contact.
Mostly due to these rigid body assumptions.
It becomes very hard to write the same kind of control laws
that we've been able to be successful with
for like fighter jets.
We haven't been as successful writing those controllers
for manipulation.
- And so you don't know what's going to happen
at the point of contact, at the moment of contact.
- There are situations absolutely
where our laws don't tell us.
So the standard approach, that's okay.
I mean, instead of having a differential equation,
you end up with a differential inclusion, it's called.
It's a set valued equation.
It says that I'm in this configuration,
I have these forces applied on me,
and there's a set of things that could happen.
And you can-- - And those aren't continuous,
I mean, what, so when you say non-smooth,
they're not only not smooth, but this is discontinuous?
- The non-smooth comes in when I make
or break a new contact first,
or when I transition from stick to slip.
So you typically have static friction,
and then you'll start sliding,
and that'll be a discontinuous change
in velocity, for instance,
especially if you come to rest or--
- That's so fascinating.
Okay, so what do you do?
Sorry, I interrupted you. - That's fine.
- What's the hope under so much uncertainty
about what's going to happen?
What are you supposed to do?
- I mean, control has an answer for this.
Robust control is one approach,
but roughly, you can write controllers
§12Transcript · 1:42:52–1:52:53
which try to still perform the right task
despite all the things that could possibly happen.
The world might want the table to go this way and this way,
but if I write a controller
that pushes a little bit more and pushes a little bit,
I can certainly make the table go in the direction I want.
It just puts a little bit more of a burden
on the control system, right?
And these discontinuities do change the control system
because the way we write it down right now,
every different control configuration,
including sticking or sliding
or parts of my body that are in contact or not,
looks like a different system.
And I think of them,
I reason about them separately or differently,
and the combinatorics of that blow up, right?
So I just don't have enough time to compute
all the possible contact configurations of my humanoid.
Interestingly, I mean, I'm a humanoid.
I have lots of degrees of freedom, lots of joints.
I've only been around for a handful of years.
It's getting up there,
but I haven't had time in my life
to visit all of the states in my system,
certainly all the contact configurations.
So if step one is to consider
every possible contact configuration that I'll ever be in,
that's probably not a problem I need to solve, right?
- Just as a small tangent,
what's a contact configuration?
Just so we can enumerate, what are we talking about?
How many are there?
- The simplest example maybe would be,
imagine a robot with a flat foot.
And we think about the phases of gait
where the heel strikes and then the front toe strikes,
and then you can heel up, toe off.
Those are each different contact configurations.
I only had two different contacts,
but I ended up with four different contact configurations.
Now, of course, my robot might actually have bumps on it
or other things, so it could be much more subtle than that.
But it's just even with one sort of box
interacting with the ground already in the plane
has that many, right?
And if I was just even a 3D foot,
then it probably my left toe might touch
just before my right toe and things get subtle.
Now, if I'm a dexterous hand
and I go to talk about just grabbing a water bottle,
if I have to enumerate every possible order
that my hand came into contact with the bottle,
then I'm dead in the water.
Any approach that we were able to get away with that
in walking because we mostly touched the ground
more than a small number of points, for instance,
and we haven't been able to get dexterous hands that way.
- So you've mentioned that people think that contact
is really hard and that that's the reason
that robotic manipulation problem is really hard.
Is there any flaws in that thinking?
- So I think simulating contact is one aspect.
I know people often say that we don't,
that one of the reasons that we have a limit in robotics
is because we do not simulate contact accurately
in our simulators.
And I think that is,
the extent to which that's true is partly
because our simulators,
we haven't got mature enough simulators.
There are some things that are still hard, difficult,
that we should change.
But we actually, we know what the governing equations are.
They have some foibles, like this indeterminacy,
but we should be able to simulate them accurately.
We have incredible open source community in robotics,
but it actually just takes a professional engineering team
a lot of work to write a very good simulator like that.
- My word is, I believe you've written, Drake.
- There's a team of people.
I certainly spent a lot of hours on it myself.
- Well, what is Drake?
What does it take to create a simulation environment
for the kind of difficult control problems
we're talking about?
- Right, so Drake is the simulator that I've been working on.
There are other good simulators out there.
I don't like to think of Drake as just a simulator,
'cause we write our controllers in Drake,
we write our perception systems a little bit in Drake,
but we write all of our low-level control
and even planning and optimization.
- So it has optimization capabilities as well?
- Absolutely, yeah.
I mean, Drake is three things, roughly.
It's an optimization library, which is,
sits on, it provides a layer of abstraction
in C++ and Python for commercial solvers.
You can write linear programs, quadratic programs,
semi-definite programs, sums of squares programs,
the ones we've used, mixed integer programs,
and it will do the work to curate those
and send them to whatever the right solver is,
for instance, and it provides a level of abstraction.
The second thing is a system modeling language,
a bit like LabVIEW or Simulink,
where you can make block diagrams out of complex systems,
or it's like ROS in that sense,
where you might have lots of ROS nodes
that are each doing some part of your system,
but to contrast it with ROS,
we try to write, if you write a Drake system,
then you have to, it asks you to describe
a little bit more about the system.
If you have any state, for instance, in the system,
any variables that are gonna persist,
you have to declare them.
Parameters can be declared and the like,
but the advantage of doing that is that you can,
if you like, run things all on one process,
but you can also do control design against it.
You can do, I mean, simple things like rewinding
and playing back your simulations, for instance.
You know, these things, you get some rewards
for spending a little bit more upfront cost
in describing each system.
And I was inspired to do that
because I think the complexity of Atlas, for instance,
is just so great.
And I think, although, I mean,
ROS has been incredible, absolute,
huge fan of what it's done for the robotics community,
but the ability to rapidly put different pieces together
and have a functioning thing is very good.
But I do think that it's hard to think clearly
about a bag of disparate parts,
Mr. Potato Head kind of software stack.
And if you can, you know, ask a little bit more
out of each of those parts,
then you can understand the way they work better.
You can try to verify them and the like,
or you can do learning against them.
And then one of those systems, the last thing,
I said the first two things that Drake is,
but the last thing is that there is a set
of multi-body equations, rigid body equations,
that is trying to provide a system that simulates physics.
And that, we also have renderers and other things,
but I think the physics component of Drake is special
in the sense that we have done excessive amount
of engineering to make sure
that we've written the equations correctly.
Every possible tumbling satellite or spinning top
or anything that we could possibly write as a test
is tested.
We are making some, you know, I think,
fundamental improvements on the way you simulate contact.
- Just what does it take to simulate contact?
I mean, it just seems,
I mean, there's something just beautiful
with the way you were explaining contact
and you were tapping your fingers on the table
while you're doing it.
Just--
- Easily, right?
- Easily, just not even,
it was helping you think, I guess.
You have this awesome demo of loading
or unloading a dishwasher.
Just picking up a plate,
grasping it like for the first time.
That just seems like so difficult.
What, how do you simulate any of that?
- So it was really interesting that what happened was that
we started getting more professional
about our software development
during the DARPA Robotics Challenge.
I learned the value of software engineering
and how to bridle complexity.
I guess that's what I want to somehow fight against
and bring some of the clear thinking of controls
into these complex systems we're building for robots.
Shortly after the DARPA Robotics Challenge,
Toyota opened a research institute,
TRI, Toyota Research Institute.
They put one of their, there's three locations.
One of them is just down the street from MIT
and I helped ramp that up
as a part of my, the end of my sabbatical, I guess.
So TRI has given me, the TRI Robotics effort
has made this investment in simulation in Drake.
And Michael Sherman leads a team there
of just absolutely top-notch dynamics experts
that are trying to write those simulators
that can pick up the dishes.
And there's also a team working on manipulation there
that is taking problems like loading the dishwasher
and we're using that to study these really hard corner cases
kind of problems in manipulation.
§13Transcript · 1:52:55–2:01:55
So for me, this, you know, simulating the dishes,
we could actually write a controller.
If we just cared about picking up dishes in the sink once,
we could write a controller
without any simulation whatsoever
and we could call it done.
But we want to understand like,
what is the path you take to actually get to a robot
that could perform that for any dish in anybody's kitchen
with enough confidence
that it could be a commercial product, right?
And it has deep learning perception in the loop.
It has complex dynamics in the loop.
It has controller, it has a planner.
And how do you take all of that complexity
and put it through this engineering discipline
and verification and validation process
to actually get enough confidence to deploy?
I mean, the DARPA challenge made me realize
that that's not something you throw over the fence
and hope that somebody will harden it for you.
That there are really fundamental challenges
in closing that last gap.
- They're doing the validation and the testing.
- I think it might even change the way we have to think
about the way we write systems.
What happens if you have the robot running lots of tests
and it screws up, it breaks a dish, right?
How do you capture that?
I said, you can't run the same simulation
or the same experiment twice on a real robot.
Do we have to be able to bring that one-off failure
back into simulation in order to change our controllers,
study it, make sure it won't happen again?
Do we, is it enough to just try to add that
to our distribution and understand that on average,
we're gonna cover that situation again?
There's like really subtle questions at the corner cases
that I think we don't yet have satisfying answers for.
- How do you find the corner cases?
That's one kind of, is there,
do you think that's possible to create a systematized way
of discovering corner cases efficiently?
- Yes.
- And whatever the problem is?
- Yes, I mean, I think we have to get better at that.
I mean, control theory has, for decades,
talked about active experiment design.
- What's that?
- So people call it curiosity these days.
It's roughly this idea of trying to,
exploration or exploitation,
but in the active experiment design is even,
is more specific.
You could try to understand the uncertainty in your system,
design the experiment that will provide
the maximum information to reduce that uncertainty.
If there's a parameter you wanna learn about,
what is the optimal trajectory I could execute
to learn about that parameter, for instance?
Scaling that up to something that has a deep network
in the loop and a planning in the loop is tough.
We've done some work on,
you know, with Matt O'Kelley and Amansina,
we've worked on some falsification algorithms
that are trying to do rare event simulation
that try to just hammer on your simulator.
And if your simulator is good enough,
you can spend a lot of time,
you can write good algorithms
that try to spend most of their time in the corner cases.
So you basically imagine you're building an autonomous car
and you wanna put it in, I don't know,
downtown New Delhi all the time, right?
And accelerated testing.
If you can write sampling strategies,
which figure out where your controller's performing badly
in simulation and start generating
lots of examples around that.
You know, it's just the space of possible places
where that can be, where things can go wrong is very big.
So it's hard to write those algorithms.
- Yeah, rare event simulation
is just like a really compelling notion.
If it's possible.
- We joked and we call it the black swan generator.
- It's a black swan.
- 'Cause you don't just want the rare events,
you want the ones that are highly impactful.
- I mean, that's the most,
those are the most sort of profound questions
we ask of our world.
Like, what's the worst that can happen?
But what we're really asking
isn't some kind of like computer science,
worst case analysis.
We're asking like, what are the millions of ways
this can go wrong?
And that's like our curiosity.
We humans, I think are pretty bad at,
we just like run into it.
And I think there's a distributed sense
because there's now like 7.5 billion of us.
And so there's a lot of them,
and then a lot of them write blog posts
about the stupid thing they've done.
So we learn in a distributed way.
There's some--
- I think that's gonna be important for robots too.
- Yeah.
- I mean, that's another massive theme
at Toyota Research for robotics
is this fleet learning concept.
Is the idea that I as a human,
I don't have enough time to visit all of my states.
Right?
There's just a, it's very hard for one robot
to experience all the things.
But that's not actually the problem we have to solve.
Right?
We're gonna have fleets of robots
that can have very similar appendages.
And at some point, maybe collectively,
they have enough data that their computational processes
should be set up differently than ours.
Right?
- It's this vision of just,
I mean, all these dishwasher unloading robots.
I mean, that robot dropping a plate
and a human looking at the robot, probably pissed off.
- Yeah.
- But that's a special moment to record.
I think one thing in terms of fleet learning,
and I've seen that because I've talked to a lot of folks,
just like Tesla users or Tesla drivers,
they're another company that's using
this kind of fleet learning idea.
One hopeful thing I have about humans
is they really enjoy when a system improves, learns.
So they enjoy fleet learning.
And the reason it's hopeful for me
is they're willing to put up with something
that's kind of dumb right now.
And they're like, if it's improving,
they almost enjoy being part of the teaching.
It almost like if you have kids,
like you're teaching them something.
- Right.
- I think that's a beautiful thing
'cause that gives me hope
that we can put dumb robots out there.
I mean, the problem on the Tesla side with cars,
cars can kill you.
That makes the problem so much harder.
Dishwasher unloading is a little safe.
That's why home robotics is really exciting.
And just to clarify, I mean,
for people who might not know,
I mean, TRI, Toyota Research Institute,
so they're pretty well known
for like autonomous vehicle research,
but they're also interested in home robotics.
- Yep, there's a big group working on,
multiple groups working on home robotics.
It's a major part of the portfolio.
There's also a couple other projects
and advanced materials discovery,
using AI and machine learning to discover new materials
for car batteries and the like, for instance.
Yeah, and that's been actually incredibly successful team.
There's new projects starting up too.
So--
- Do you see a future of where like robots are in our home
and like robots that have like actuators
that look like arms in our home
or like more like humanoid type robots?
Or is this, are we gonna do the same thing
that you just mentioned that,
the dishwasher is no longer a robot.
We're going to just not even see them as robots.
But I mean, what's your vision of the home of the future?
10, 20 years from now, 50 years if you get crazy.
- Yeah, I think we already have Roombas cruising around.
We have, you know, Alexa's or Google Homes
on our kitchen counter.
It's only a matter of time till they spring arms
and start doing something useful like that.
So I do think it's coming.
I think lots of people have lots of motivations
for doing it.
It's been super interesting actually learning
about Toyota's vision for it,
which is about helping people age in place.
'Cause I think that's not necessarily the first entry,
the most lucrative entry point,
but it's the problem maybe that
we really need to solve no matter what.
And so I think there's a real opportunity.
It's a delicate problem.
How do you work with people, help people,
§14Transcript · 2:01:59–2:10:59
keep them active, engaged, you know,
but improve the quality of life
and help them age in place, for instance.
- It's interesting 'cause older folks are also,
I mean, there's a contrast there because
they're not always the folks
who are the most comfortable with technology, for example.
So there's a division that's interesting there
that you can do so much good with a robot
for older folks,
but there's a gap to fill of understanding.
I mean, it's actually kind of beautiful.
Robot is learning about the human
and the human is kind of learning about this new robot thing.
And it's also with,
at least with, like when I talk to my parents about robots,
there's a little bit of a blank slate there too.
Like you can, I mean, they don't know anything about robotics
so it's completely like wide open.
They don't have, they haven't,
my parents haven't seen Black Mirror.
So like it's a blank slate.
Here's a cool thing, like what can it do for me?
- Yeah.
- So it's an exciting space.
- I think it's a really important space.
I do feel like, you know, a few years ago,
drones were successful enough in academia.
They kind of broke out and started in industry
and autonomous cars have been happening.
It does feel like manipulation in logistics, of course,
first, but in the home shortly after,
it seems like one of the next big things
that's gonna really pop.
- So I don't think we talked about it,
but what's soft robotics?
So we talked about like rigid bodies.
If we can just linger on this whole touch thing.
- Yeah, so what's soft robotics?
So I told you that I really dislike the fact
that robots are afraid of touching the world
all over their body.
So there's a couple reasons for that.
If you look carefully at all the places
that robots actually do touch the world,
they're almost always soft.
They have some sort of pad on their fingers
or a rubber sole on their foot.
But if you look up and down the arm,
we're just pure aluminum or something.
So that makes it hard actually.
In fact, hitting the table with your rigid arm
or nearly rigid arm has some of the problems
that we talked about in terms of simulation.
I think it fundamentally changes the mechanics of contact
when you're soft, right?
You turn point contacts into patch contacts,
which can have torsional friction.
You can have distributed load.
If I wanna pick up an egg, right?
If I pick it up with two points,
then in order to put enough force
to sustain the weight of the egg,
I might have to put a lot of force to break the egg.
If I envelop it with contact all around,
then I can distribute my force across the shell of the egg
and have a better chance of not breaking it.
So soft robotics is for me a lot about
changing the mechanics of contact.
- Does it make the problem a lot harder?
(laughing)
- Quite the opposite.
It changes the computational problem.
I think because of the,
I think our world and our mathematics
has biased us towards rigid.
But it really should make things better in some ways, right?
It's a, I think the future is unwritten there.
But the other thing--
- I think ultimately, sorry to interrupt,
but I think ultimately it will make things simpler
if we embrace the softness of the world.
- It makes things smoother, right?
So the result of small actions is less discontinuous,
but it also means potentially less, you know,
instantaneously bad, for instance.
I won't necessarily contact something and send it flying off.
The other aspect of it
that just happens to dovetail really well
is that soft robotics tends to be a place
where we can embed a lot of sensors too.
So if you change your hardware and make it more soft,
then you can potentially have a tactile sensor,
which is measuring the deformation.
So there's a team at TRI that's working on soft hands,
and you get so much more information.
If you can put a camera behind the skin roughly
and get fantastic tactile information,
which is, it's super important.
Like in manipulation,
one of the things that really is frustrating
is if you work super hard on your head-mounted,
on your perception system for your head-mounted cameras,
and then you've identified an object,
you reach down to touch it,
and the last thing that happens,
right before the most important time,
you stick your hand and you're occluding
your head-mounted sensors, right?
So in all the part that really matters,
all of your off-board sensors are occluded.
And really, if you don't have tactile information,
then you're blind in an important way.
So it happens that soft robotics and tactile sensing
tend to go hand in hand.
- I think we've kind of talked about it,
but you taught a course on under-actuated robotics.
I believe that was the name of it, actually.
- That's right.
- Can you talk about it in that context?
What is under-actuated robotics?
- Right, so under-actuated robotics is my graduate course.
It's online mostly now,
in the sense that the lectures--
- Several versions of it, I think.
- Right, the YouTube--
- It's really great, I recommend it highly.
- Look on YouTube for the 2020 versions until March,
and then you have to go back to 2019, thanks to COVID.
No, I've poured my heart into that class.
And lecture one is basically explaining
what the word under-actuated means.
So people are very kind to show up,
and then maybe have to learn
what the title of the course means
over the course of the first lecture.
- That first lecture's really good.
You should watch it.
- Thanks.
It's a strange name,
but I thought it captured the essence
of what control was good at doing,
and what control was bad at doing.
So what do I mean by under-actuated?
So a mechanical system
has many degrees of freedom, for instance.
I think of a joint as a degree of freedom.
And it has some number of actuators, motors.
So if you have a robot that's bolted to the table
that has five degrees of freedom, and five motors,
then you have a fully actuated robot.
If you take away one of those motors,
then you have an under-actuated robot.
Now why on earth?
I have a good friend who likes to tease me.
He said, "Russ, if you had more research funding,
"would you work on fully actuated robots?"
- Yeah. (laughs)
- The answer's no.
The world gives us under-actuated robots,
whether we like it or not.
I'm a human.
I'm an under-actuated robot,
even though I have more muscles
than my big degrees of freedom,
because I have, in some places,
multiple muscles attached to the same joint.
But still, there's a really important degree of freedom
that I have, which is the location
of my center of mass in space, for instance.
All right, I can jump into the air,
and there's no motor that connects my center of mass
to the ground in that case.
So I have to think about the implications
of not having control over everything.
The passive dynamic walkers are the extreme view of that,
where you've taken away all the motors,
and you have to let physics do the work.
But it shows up in all of the walking robots,
where you have to use some of the actuators
to push and pull even the degrees of freedom
that you don't have an actuator on.
- That's referring to walking if you're falling forward.
Is there a way to walk that's fully actuated?
- So it's a subtle point.
When you're in contact, and you have your feet
on the ground, there are still limits to what you can do.
Unless I have suction cups on my feet,
I cannot accelerate my center of mass
towards the ground faster than gravity,
'cause I can't get a force pushing me down.
But I can still do most of the things that I want to.
So you can get away with basically thinking
of the system as fully actuated,
unless you suddenly needed to accelerate down super fast.
But as soon as I take a step,
I get into the more nuanced territory.
And to get to really dynamic robots,
or airplanes, or other things,
I think you have to embrace the under-actuated dynamics.
§15Transcript · 2:11:02–2:19:44
Manipulation, people think,
is manipulation under-actuated?
Even if my arm is fully actuated, I have a motor,
if my goal is to control the position
and orientation of this cup,
then I don't have an actuator for that directly.
So I have to use my actuators over here
to control this thing.
Now it gets even worse,
like what if I have to button my shirt?
What are the degrees of freedom of my shirt?
I suddenly, that's a hard question to think about.
It kind of makes me queasy
as thinking about my state-space control ideas.
But actually those are the problems
that make me so excited about manipulation right now,
is that it breaks some of the,
it breaks a lot of the foundational control stuff
that I've been thinking about.
- What are some interesting insights you could say
about trying to solve an under-actuated,
a control in an under-actuated system?
- So I think the philosophy there
is let physics do more of the work.
The technical approach has been optimization.
So you typically formulate your decision-making for control
as an optimization problem,
and you use the language of optimal control
and sometimes often numerical optimal control
in order to make those decisions
and balance these complicated equations of,
and in order to control.
You don't have to use optimal control
to do under-actuated systems,
but that has been the technical approach
that has borne the most fruit in our,
at least in our line of work.
- And there's some, so in under-actuated systems,
when you say let physics do some of the work,
so there's a kind of feedback loop
that observes the state that the physics brought you to.
So like you've, there's a perception there,
there's a feedback somehow.
Do you ever loop in like complicated perception systems
into this whole picture?
- Right, right around the time of the DARPA challenge,
we had a complicated perception system
in the DARPA challenge.
We also started to embrace perception
for our flying vehicles at the time.
We had a really good project
on trying to make airplanes fly
at high speeds through forests.
Sertac Karaman was on that project,
and we had, it was a really fun team to work on.
He's carried it farther, much farther forward since then.
- And that's using cameras for perception?
- So that was using cameras.
That was, at the time, we felt like LIDAR was too heavy
and too power heavy to be carried on a light UAV,
and we were using cameras.
And that was a big part of it,
was just how do you do even stereo matching
at a fast enough rate with a small camera,
a small onboard compute.
Since then, we have now,
so the deep learning revolution unquestionably changed
what we can do with perception for robotics and control.
So in manipulation, we can address,
we can use perception in, I think, a much deeper way.
And we get into not only,
I think the first use of it naturally
would be to ask your deep learning system
to look at the cameras and produce the state,
which is like the pose of my thing, for instance.
But I think we've quickly found out
that that's not always the right thing to do.
- Why is that?
- Because what's the state of my shirt?
Imagine--
- Is it very noisy, you mean?
- If the first step of me trying to button my shirt
is estimate the full state of my shirt,
including what's happening in the back,
you know, whatever, whatever,
that's just not the right specification.
There are aspects of the state
that are very important to the task.
There are many that are unobservable
and not important to the task.
So you really need,
it begs new questions about state representation.
Another example that we've been playing with in lab
has been just the idea of chopping onions, okay?
Or carrots, turns out to be better.
So the onions stink up the lab.
And they're hard to see in a camera.
But--
(Lex laughing)
- The details matter, yeah.
- Details matter, you know?
So--
- Chop and carrot.
- If I'm moving around a particular object, right?
Then I think about,
oh, it's got a position or an orientation in space,
that's the description I want.
Now, when I'm chopping an onion, okay,
the first chop comes down,
I have now 100 pieces of onion.
Does my control system really need to understand
the position and orientation
and even the shape of the 100 pieces of onion
in order to make a decision?
Probably not, you know?
And if I keep going, I'm just getting,
more and more is my state space getting bigger as I cut?
It's not right.
So somehow there's a-- - What do you do?
- I think there's a richer idea of state.
It's not the state that is given to us
by Lagrangian mechanics.
There is a proper Lagrangian state of the system,
but the relevant state for this
is some latent state,
is what we call it in machine learning.
But there's some different state representation.
- Some compressed representation?
- And that's what I worry about saying compressed
because I don't mind that it's low dimensional or not,
but it has to be something that's easier to think about.
- By us humans.
- Or my algorithms.
- Or the algorithms being like control, optimal control.
- So for instance, if the contact mechanics
of all of those onion pieces
and all the permutations of possible touches
between those onion pieces,
you can give me a high dimensional state representation,
I'm okay if it's linear.
But if I have to think about all the possible
shattering combinatorics of that,
then my robot's gonna sit there thinking
and the soup's gonna get cold or something.
- So since you taught the course,
it kinda entered my mind,
the idea of underactuated as really compelling
to see the world in this kind of way.
Do you ever, if we talk about onions
or you talk about the world with people in it in general,
do you see the world as basically an underactuated system?
Do you often look at the world in this way?
Or is this overreach?
- Underactuated is a way of life, man.
- Exactly, I guess that's what I'm asking.
- I do think it's everywhere.
I think in some places,
we already have natural tools to deal with it.
It rears its head.
I mean, in linear systems, it's not a problem.
An underactuated linear system
is really not sufficiently distinct
from a fully actuated linear system.
It's a subtle point about when that becomes a bottleneck
in what we know how to do with control.
It happens to be a bottleneck,
although we've gotten incredibly good solutions now.
But for a long time,
I felt that that was the key bottleneck in Legged Robots.
And roughly now, the underactuated course is,
me trying to tell people everything I can
about how to make Atlas do a backflip.
I have a second course now
that I teach in the other semesters,
which is on manipulation.
And that's where we get into now more of the,
that's a newer class.
I'm hoping to put it online this fall completely.
And that's gonna have much more aspects
about these perception problems
and the state representation questions,
and then how do you do control.
And the thing that's a little bit sad is that,
for me at least, is there's a lot of manipulation tasks
that people wanna do and should wanna do.
They could start a company with it and be very successful
that don't actually require you to think that much
about dynamics at all, even,
but certainly underactuated dynamics.
Once I have, if I reach out and grab something,
if I can sort of assume it's rigidly attached to my hand,
then I can do a lot of interesting,
meaningful things with it
without really ever thinking about the dynamics
of that object.
So we've built systems that kind of
reduce the need for that,
enveloping grasps and the like.
But I think the really good problems in manipulation,
so manipulation, by the way,
§16Transcript · 2:19:46–2:30:15
is more than just pick and place.
That's like, a lot of people think of that, just grasping.
I don't mean that, I mean buttoning my shirt.
I mean tying shoelaces.
- Tying shoelaces. - How do you program
a robot to tie shoelaces?
And not just one shoe, but every shoe, right?
That's a really good problem.
It's tempting to write down
like the infinite dimensional state of the laces.
That's probably not needed to write a good controller.
I know we could hand design a controller that would do it,
but I don't want that.
I wanna understand the principles
that would allow me to solve another problem
that's kind of like that.
But I think if we can stay pure in our approach,
then the challenge of tying anybody's shoes
is a great challenge.
- That's a great challenge.
I mean, and the soft touch comes into play there.
That's really interesting.
Let me ask another ridiculous question on this topic.
How important is touch?
We haven't talked much about humans,
but I have this argument with my dad
where like I think you can fall in love with a robot
based on language alone.
And he believes that touch is essential.
Touch and smell, he says.
So in terms of robots, you know, connecting with humans,
and we can go philosophical
in terms of like a deep, meaningful connection, like love,
but even just like collaborating in an interesting way,
how important is touch?
From an engineering perspective and a philosophical one.
- I think it's super important.
Even just in a practical sense,
if we forget about the emotional part of it,
but for robots to interact safely
while they're doing meaningful mechanical work
in the close contact with or vicinity of people
that need help, I think we have to have them,
we have to build them differently.
They have to be afraid, not afraid of touching the world.
So I think Baymax is just awesome.
That's just like the movie of "Big Hero 6"
and the concept of Baymax, that's just awesome.
I think we should, and we have some folks at Toyota
that are trying to, Toyota Research,
that are trying to build Baymax roughly.
And I think it's just a fantastically good project.
I think it will change the way people physically interact.
The same way, I mean, you gave a couple examples earlier,
but if the robot that was walking around my home
looked more like a teddy bear
and a little less like the Terminator,
that could change completely the way people perceive it
and interact with it.
And maybe they'll even wanna teach it, like you said.
You could not quite gamify it,
but somehow instead of people judging it
and looking at it as if it's not doing as well as a human,
they're gonna try to help out the cute teddy bear.
Who knows?
But I think we're building robots wrong
and being more soft and more contact is important.
- Yeah, like all the magical moments
I can remember with robots.
Well, first of all, just visiting your lab and seeing Atlas,
but also Spot Mini.
When I first saw Spot Mini in person
and hung out with him, her, it,
I don't have trouble engendering robots.
I feel robotics people really always say it.
I kinda like the idea that it's a her or him.
It's a magical moment, but there's no touching.
I guess the question I have, have you ever been,
like, have you had a human robot experience
where a robot touched you?
And it was like, wait, was there a moment
that you've forgotten that a robot is a robot?
And the anthropomorphization stepped in
and for a second you forgot that it's not human?
- I mean, I think when you're in on the details,
then we of course anthropomorphized our work with Atlas,
but in verbal communication and the like,
I think we were pretty aware of it as a machine
that needed to be respected.
I actually, I worry more about the smaller robots
that could still move quickly if programmed wrong
and we have to be careful actually about safety
and the like right now.
And that, if we build our robots correctly,
I think then a lot of those concerns could go away.
And we're seeing that trend.
We're seeing the lower cost, lighter weight arms now
that could be fundamentally safe.
I mean, I do think touch is so fundamental.
Ted Adelson is great.
He's a perceptual scientist at MIT
and he studied vision most of his life.
And he said, "When I had kids,
"I expected to be fascinated
"by their perceptual development."
But what really, what he noticed was,
felt more impressive, more dominant
was the way that they would touch everything
and lick everything and pick things up,
stick it on their tongue and whatever.
And he said, watching his daughter
convinced him that actually he needed to study
tactile sensing more.
So there's something very important.
I think it's a little bit also of the passive
versus active part of the world, right?
You can passively perceive the world,
but it's fundamentally different
if you can do an experiment, right?
And if you can change the world.
And you can learn a lot more than a passive observer.
So you can, in dialogue, that was your initial example,
you could have an active experiment exchange.
But I think if you're just a camera watching YouTube,
I think that's a very different problem
than if you're a robot that can apply force and touch.
I think it's important.
- Yeah, I think it's just an exciting area of research.
I think you're probably right
that this hasn't been under-researched.
To me, as a person who's captivated
by the idea of human robot interaction,
it feels like such a rich opportunity to explore touch.
Not even from a safety perspective,
but like you said, the emotional too.
I mean, safety comes first,
but the next step is like a real human connection.
Even in the industrial setting,
it just feels like it's nice for the robot.
- I don't know, you might disagree with this,
'cause I think it's important to see robots as tools often,
but I don't know.
I think they're just always going to be more effective
once you humanize them.
It's convenient now to think of them as tools
'cause we wanna focus on the safety,
but I think ultimately to create a good experience
for the worker, for the person,
there has to be a human element.
I don't know, for me.
It feels like an industrial robotic arm
would be better if it has a human element.
I think like we think robotics had that idea
with Baxter and having eyes and so on.
I don't know, I'm a big believer in that.
- It's not my area, but I am also a big believer.
- Do you have an emotional connection to Atlas?
(laughing)
Like, do you miss him?
- I mean, yes.
I don't know if I more so than if I had
a different science project that I'd worked on super hard.
But yeah, I mean, the robot,
we basically had to do heart surgery on the robot
in the final competition 'cause we melted the core.
Yeah, there was something about watching that robot
hanging there, we know we had to compete with it
in an hour and it was getting its guts ripped out.
- Those are all historic moments.
I think if you look back like 100 years from now,
yeah, I think those are important moments in robotics.
I mean, these are the early days.
You look at like the early days
of a lot of scientific disciplines,
they look ridiculous, it's full of failure.
But it feels like robotics will be important
in the coming 100 years.
And these are the early days.
So I think a lot of people look at a brilliant person
such as yourself and are curious
about the intellectual journey they've took.
Is there maybe three books, technical, fiction,
philosophical that had a big impact on your life
that you would recommend perhaps others reading?
- Yeah, so I actually didn't read that much as a kid,
but I read fairly voraciously now.
There are some recent books that if you're interested
in this kind of topic, like "AI Superpowers" by Kai-Fu Lee
is just a fantastic read, you must read that.
Yuval Harari, I think that can open your mind.
- "Sapiens." - "Sapiens" is the first one.
"Homo Deus" is the second, yeah.
We mentioned "The Black Swan" by Taleb.
I think that's a good sort of mind opener.
I actually, so there's maybe
a more controversial recommendation I could give.
- Great, we love controversial.
- In some sense, it's so classical it might surprise you.
But I actually recently read Mortimer Adler's
"How to Read a Book."
§17Transcript · 2:30:17–2:39:59
Not so long, it was a while ago.
But some people hate that book.
I loved it.
I think we're in this time right now where,
boy, we're just inundated with research papers
that you could read on archive with limited peer review
and just this wealth of information.
I don't know, I think the passion
of what you can get out of a book,
a really good book or a really good paper if you find it,
the attitude, the realization that you're only gonna find
a few that really are worth all your time.
But then once you find them, you should just dig in
and understand it very deeply and it's worth
marking it up and having the hard copy,
writing in the side notes, side margins.
I think that was really,
I read it at the right time where I was just feeling
just overwhelmed with really low quality stuff, I guess.
And similarly, I'm giving more than three now.
I'm sorry if I've exceeded my quota.
- But on that topic just real quick is,
so basically finding a few companions
to keep for the rest of your life
in terms of papers and books and so on.
And those are the ones, like not doing,
what is it, FOMO, fear of missing out,
constantly trying to update yourself,
but really deeply making a life journey
of studying a particular paper essentially, set of papers.
- Yeah, I think when you really find something,
which a book that resonates with you
might not be the same book that resonates with me,
but when you really find one that resonates with you,
I think the dialogue that happens,
and that's what I loved that Adler was saying,
I think Socrates and Plato say,
the written word is never gonna capture
the beauty of dialogue, right?
But Adler says, no, no,
a really good book is a dialogue between you and the author,
and it crosses time and space.
I don't know, I think it's a very romantic,
there's a bunch of specific advice,
which you can just gloss over,
but the romantic view of how to read
and really appreciate it is so good.
And similarly, teaching,
I thought a lot about teaching,
so Isaac Asimov, great science fiction writer,
has also actually spent a lot of his career
writing nonfiction, right?
His memoir is fantastic.
He was passionate about explaining things, right?
He wrote all kinds of books on all kinds of topics
in science, he was known as the great explainer.
And I do really resonate with his style
and just his way of talking about,
by communicating and explaining to something
is really the way that you learn something.
I think about problems very differently
because of the way I've been given the opportunity
to teach them at MIT, and have questions asked,
the fear of the lecture, the experience of the lecture
and the questions I get and the interactions
just forces me to be rock solid on these ideas
in a way that if I didn't have that,
I don't know I would be in a different intellectual space.
- Also, video, does that scare you
that your lectures are online,
and people like me in sweatpants can sit,
sipping coffee and watch you give lectures?
- I think it's great.
I do think that something's changed right now,
which is, right now we're giving lectures over Zoom,
I mean, giving seminars over Zoom and everything.
I'm trying to figure out, I think it's a new medium.
- Do you think it's-- - I'm trying to figure out
how to exploit it. - How to exploit it.
Yeah, I've been quite cynical
about the human to human connection over that medium,
but I think that's because it hasn't been explored fully,
and teaching is a different thing.
- Every lecture is a, I'm sorry, every seminar even,
I think every talk I give,
is an opportunity to give that differently.
I can deliver content directly into your browser.
You have a WebGL engine right there.
I can throw 3D content into your browser
while you're listening to me, right?
And I can assume that you have a,
at least a powerful enough laptop or something
to watch Zoom while I'm doing that,
while I'm giving a lecture.
That's a new communication tool
that I didn't have last year, right?
And I think robotics can potentially benefit a lot
from teaching that way.
We'll see, it's gonna be an experiment this fall.
- It's interesting. - I'm thinking a lot
about it.
- Yeah, and also the length of lectures,
or the length of, there's something,
so I guarantee you, 80% of people
who started listening to our conversation
are still listening to now, which is crazy to me.
There's a patience and an interest in long-form content,
but at the same time, there's a magic
to forcing yourself to condense,
an idea to a short as possible,
shortest possible like clip.
It can be a part of a longer thing,
but like just a really beautifully condensed an idea.
There's a lot of opportunity there
that's easier to do in remote with,
I don't know, with editing too.
Editing is an interesting thing.
Well, most professors don't get,
when they give a lecture,
they don't get to go back and edit out parts,
like crisp it up a little bit.
That's also, it can do magic.
Like if you remove like five to 10 minutes
from an hour lecture, it can actually,
it can make something special of a lecture.
I've seen that in myself and in others too,
'cause I edit other people's lectures to extract clips.
It's like there's certain tangents that are like,
that lose, they're not interesting.
They're mumbling, they're just not,
they're not clarifying, they're not helpful at all.
And once you remove them, it's just, I don't know.
Editing can be magic.
- It can take a lot of time.
- Yeah, it takes, it depends like what is teaching,
you have to ask.
- Yeah.
- Yeah, 'cause I find the editing process
is also beneficial as for teaching,
but also for your own learning.
I don't know if, have you watched yourself?
- Yeah, sure.
- Have you watched those videos?
- I mean, not all of them.
- It could be painful to see how to improve.
- So do you find that, I know you segment your podcast.
Do you think that helps people
with the attention span aspect of it?
Or is it--
- Segment like sections like--
- Yeah, we're talking about this topic, whatever.
- Nope, nope, that just helps me.
It's actually bad.
So, and you've been incredible.
So I'm learning, like I'm afraid of conversation.
This is even today, I'm terrified of talking to you.
I mean, it's something I'm trying to remove from myself.
A guy, I mean, I learned from a lot of people,
but really there's been a few people
who's been inspirational to me in terms of conversation.
Whatever people think of him, Joe Rogan
has been inspirational to me
because comedians have been too.
Being able to just have fun and enjoy themselves
and lose themselves in conversation,
that requires you to be a great storyteller,
to be able to pull a lot of different pieces
of information together,
but mostly just to enjoy yourself in conversations.
And I'm trying to learn that.
These notes are, you see me looking down,
that's like a safety blanket
that I'm trying to let go of more and more.
- Cool.
- So that's, people love just regular conversation.
That's what they, the structure is like, whatever.
I would say maybe like 10 to,
so there's a bunch of,
there's probably a couple thousand PhD students
listening to this right now, right?
And they might know what we're talking about,
but there is somebody, I guarantee you right now in Russia,
some kid who's just like, who's just smoked some weed,
sitting back and just enjoying
the hell out of this conversation.
Not really understanding,
he kind of watched some Boston Dynamics videos.
He's just enjoying it.
And I salute you, sir.
No, but just like, there's so much variety of people
that just have curiosity about engineering,
about sciences, about mathematics.
And also like I should,
I mean, enjoying it is one thing,
but I also often notice it inspires people to,
there's a lot of people who are like
in their undergraduate studies trying to figure out what,
trying to figure out what to pursue.
And these conversations can really spark
the direction of their life.
§18Transcript · 2:40:02–2:48:39
And in terms of robotics, I hope it does,
'cause I'm excited about the possibilities
of what robotics brings.
On that topic, do you have advice?
Like what advice would you give to a young person about life?
- A young person about life
or a young person about life in robotics?
- It could be in robotics.
It could be in life in general.
It could be career.
It could be relationship advice.
It could be running advice.
Just like, that's one of the things I see,
like we talked to like 20 year olds.
They're like, how do I do this thing?
What do I do?
If they come up to you, what would you tell them?
- I think it's an interesting time to be a kid these days.
Everything points to this being
sort of a winner take all economy and the like.
I think the people that will really excel, in my opinion,
are gonna be the ones that can think deeply about problems.
You have to be able to ask questions agilely
and use the internet for everything it's good for
and stuff like this.
And I think a lot of people will develop those skills.
I think the leaders, thought leaders,
robotics leaders, whatever,
are gonna be the ones that can do more
and they can think very deeply and critically.
And that's a harder thing to learn.
I think one path to learning that is through mathematics,
through engineering.
I would encourage people to start math early.
I mean, I didn't really start.
I mean, I was always in the better math classes
that I could take,
but I wasn't pursuing super advanced mathematics
or anything like that until I got to MIT.
I think MIT lit me up
and really started the life that I'm living now.
But yeah, I really want kids to dig deep,
really understand things, building things too.
I mean, pull things apart, put them back together.
Like that's just such a good way
to really understand things
and expect it to be a long journey, right?
You don't have to know everything.
You're never gonna know everything.
- So think deeply and stick with it.
- Enjoy the ride,
but just make sure you're not, yeah.
Just make sure you're stopping
to think about why things work.
- It's true.
It's easy to lose yourself
in the distractions of the world.
- We're overwhelmed with content right now,
but you have to stop and pick some of it
and really understand.
- Yeah, on the book point,
I've read "Animal Farm" by George Orwell
a ridiculous number of times.
So for me, like that book,
I don't know if it's a good book in general,
but for me, it connects deeply somehow.
It somehow connects.
So I was born in the Soviet Union.
So it connects to me to the entirety of the history
of the Soviet Union and to World War II
and to the love and hatred and suffering that went on there.
And the corrupting nature of power and greed
and just somehow,
that book has taught me more about life
than like anything else.
Even though it's just like a silly,
like childlike book about pigs.
I don't know why, it just connects and inspires.
And the same, there's a few,
yeah, there's a few technical books too
and algorithms that just, yeah, you return to often.
I'm with you.
Yeah, there's, I don't know.
And I've been losing that because of the internet.
I've been like going on,
I've been going on archive and blog posts and GitHub
and the new thing and you lose your ability
to really master an idea.
- Right.
- Wow.
- Exactly right.
- What's a fond memory from childhood?
Baby, Russ, Tedrick.
- Well, I guess I just said that,
at least my current life began when I got to MIT.
If I have to go farther than that.
- Yeah, was there a life before MIT?
- Oh, absolutely.
But let me actually tell you what happened
when I first got to MIT.
'Cause that, I think might be relevant here.
But I had taken a computer engineering degree at Michigan.
I enjoyed it immensely, learned a bunch of stuff.
I liked computers, I liked programming.
But when I did get to MIT and started working
with Sebastian Sung, theoretical physicist,
computational neuroscientist,
the culture here was just different.
It demanded more of me, certainly mathematically
and in the critical thinking.
And I remember the day that I borrowed one of the books
from my advisor's office and walked down
to the Charles River and was like,
I'm getting my butt kicked.
And I think that's gonna happen to everybody
who's doing this kind of stuff.
I think, I expected you to ask me the meaning of life.
I think that the, somehow I think that's gotta be part of it.
This--
- Doing hard things?
- Yeah.
- Did you consider quitting at any point?
Did you consider this isn't for me?
- No, never that.
I mean, I was working hard, but I was loving it.
I think there's this magical thing where you,
I'm lucky to surround myself with people that basically,
almost every day I'll see something,
I'll be told something or something that I realize,
wow, I don't understand that.
And if I could just understand that,
there's something else to learn,
that if I could just learn that thing,
I would connect another piece of the puzzle.
And I think that is just such an important aspect
and being willing to understand what you can and can't do
and loving the journey of going
and learning those other things.
I think that's the best part.
- I don't think there's a better way to end it, Russ.
You've been an inspiration to me since I showed up at MIT.
Your work has been an inspiration to the world.
This conversation was amazing.
I can't wait to see what you do next
with robotics, home robots.
I hope to see your work in my home one day.
So thanks so much for talking today.
It's been awesome.
- Cheers.
- Thanks for listening to this conversation
with Russ Tedrick.
And thank you to our sponsors,
Magic Spoon Cereal, BetterHelp and ExpressVPN.
Please consider supporting this podcast
by going to magicspoon.com/lex
and using code LEX at checkout.
Go into betterhelp.com/lex
and signing up at expressvpn.com/lexpod.
Click the links, buy the stuff, get the discount.
It really is the best way to support this podcast.
If you enjoy this thing, subscribe on YouTube,
review it with Five Stars and Apple Podcast,
support it on Patreon or connect with me on Twitter
at Lex Friedman spelled somehow without the E,
just F-R-I-D-M-A-N.
And now let me leave you with some words
from Neil deGrasse Tyson,
talking about robots in space
and the emphasis we humans put
on human-based space exploration.
"Robots are important.
"If I don my pure scientist hat,
"I would say just send robots.
"I'll stay down here and get the data.
"But nobody's ever given a parade for a robot.
"Nobody's ever named a high school after a robot.
"So when I don my public educator hat,
"I have to recognize the elements of exploration
"that excite people.
"It's not only the discoveries and the beautiful photos
"that come down from the heavens.
"It's the vicarious participation in discovery itself."
Thank you for listening and hope to see you next time.
(upbeat music)
(upbeat music)