Russ Tedrake: Underactuated Robotics, Control, Dynamics and Touch

Lex Fridman (01:26.240)

support it on Patreon, or connect with me

Russ Tedrake (01:28.280)

on Twitter at lexfreedman.

Lex Fridman (01:31.280)

As usual, I'll do a few minutes of ads now

Lex Fridman (01:33.640)

and never any ads in the middle

Lex Fridman (01:34.880)

that can break the flow of the conversation.

Russ Tedrake (01:37.880)

This episode is supported by Magic Spoon,

Lex Fridman (01:40.880)

low carb keto friendly cereal.

Russ Tedrake (01:43.460)

I've been on a mix of keto or carnivore diet

Lex Fridman (01:45.800)

for a very long time now.

Russ Tedrake (01:47.320)

That means eating very little carbs.

Lex Fridman (01:50.520)

I used to love cereal.

Russ Tedrake (01:52.200)

Obviously, most have crazy amounts of sugar,

Lex Fridman (01:54.960)

which is terrible for you, so I quit years ago,

Lex Fridman (01:58.000)

but Magic Spoon is a totally new thing.

Lex Fridman (02:00.420)

Zero sugar, 11 grams of protein,

Lex Fridman (02:03.000)

and only three net grams of carbs.

Lex Fridman (02:05.720)

It tastes delicious.

Russ Tedrake (02:07.240)

It has a bunch of flavors, they're all good,

Lex Fridman (02:09.660)

but if you know what's good for you,

Russ Tedrake (02:11.200)

you'll go with cocoa, my favorite flavor

Lex Fridman (02:13.940)

and the flavor of champions.

Russ Tedrake (02:15.820)

Click the magicspoon.com slash lex link in the description,

Lex Fridman (02:19.460)

use code lex at checkout to get the discount

Lex Fridman (02:22.160)

and to let them know I sent you.

Lex Fridman (02:24.400)

So buy all of their cereal.

Russ Tedrake (02:26.680)

It's delicious and good for you.

Lex Fridman (02:28.640)

You won't regret it.

Russ Tedrake (02:30.560)

This show is also sponsored by BetterHelp,

Lex Fridman (02:33.160)

spelled H E L P Help.

Russ Tedrake (02:36.040)

Check it out at betterhelp.com slash lex.

Lex Fridman (02:39.440)

They figure out what you need

Lex Fridman (02:40.600)

and match you with a licensed professional therapist

Lex Fridman (02:43.240)

in under 48 hours.

Russ Tedrake (02:44.960)

It's not a crisis line, it's not self help,

Lex Fridman (02:47.640)

it is professional counseling done securely online.

Russ Tedrake (02:51.040)

As you may know, I'm a bit from the David Goggins line

Lex Fridman (02:53.720)

of creatures and still have some demons to contend with,

Russ Tedrake (02:57.080)

usually on long runs or all nighters full of self doubt.

Lex Fridman (03:01.580)

I think suffering is essential for creation,

Lex Fridman (03:04.360)

but you can suffer beautifully

Lex Fridman (03:06.040)

in a way that doesn't destroy you.

Russ Tedrake (03:08.200)

For most people, I think a good therapist can help in this.

Lex Fridman (03:11.540)

So it's at least worth a try.

Russ Tedrake (03:13.400)

Check out the reviews, they're all good.

Lex Fridman (03:15.620)

It's easy, private, affordable, available worldwide.

Russ Tedrake (03:19.220)

You can communicate by text anytime

Lex Fridman (03:21.640)

and schedule weekly audio and video sessions.

Russ Tedrake (03:25.080)

Check it out at betterhelp.com slash lex.

Lex Fridman (03:28.500)

This show is also sponsored by ExpressVPN.

Russ Tedrake (03:31.840)

Get it at expressvpn.com slash lex pod

Lex Fridman (03:34.860)

to get a discount and to support this podcast.

Lex Fridman (03:37.680)

Have you ever watched The Office?

Lex Fridman (03:39.680)

If you have, you probably know it's based

Russ Tedrake (03:41.900)

on a UK series also called The Office.

Lex Fridman (03:45.120)

Not to stir up trouble, but I personally think

Russ Tedrake (03:48.080)

the British version is actually more brilliant

Lex Fridman (03:50.320)

than the American one, but both are amazing.

Russ Tedrake (03:53.120)

Anyway, there are actually nine other countries

Lex Fridman (03:56.120)

with their own version of The Office.

Russ Tedrake (03:58.400)

You can get access to them with no geo restriction

Lex Fridman (04:01.180)

when you use ExpressVPN.

Russ Tedrake (04:03.600)

It lets you control where you want sites

Lex Fridman (04:05.560)

to think you're located.

Russ Tedrake (04:07.340)

You can choose from nearly 100 different countries,

Lex Fridman (04:10.360)

giving you access to content

Russ Tedrake (04:12.120)

that isn't available in your region.

Lex Fridman (04:14.020)

So again, get it on any device at expressvpn.com slash lex pod

Russ Tedrake (04:19.800)

to get an extra three months free

Lex Fridman (04:22.080)

and to support this podcast.

Lex Fridman (04:25.000)

And now here's my conversation with Russ Tedrick.

Lex Fridman (04:29.560)

What is the most beautiful motion

Lex Fridman (04:31.480)

of an animal or robot that you've ever seen?

Lex Fridman (04:36.160)

I think the most beautiful motion of a robot

Russ Tedrake (04:38.280)

has to be the passive dynamic walkers.

Lex Fridman (04:41.120)

I think there's just something fundamentally beautiful.

Russ Tedrake (04:43.320)

The ones in particular that Steve Collins built

Lex Fridman (04:45.360)

with Andy Ruina at Cornell, a 3D walking machine.

Lex Fridman (04:50.520)

So it was not confined to a boom or a plane

Lex Fridman (04:54.680)

that you put it on top of a small ramp,

Russ Tedrake (04:57.460)

give it a little push, it's powered only by gravity.

Lex Fridman (05:00.500)

No controllers, no batteries whatsoever.

Russ Tedrake (05:04.320)

It just falls down the ramp.

Lex Fridman (05:06.160)

And at the time it looked more natural, more graceful,

Russ Tedrake (05:09.520)

more human like than any robot we'd seen to date

Lex Fridman (05:13.460)

powered only by gravity.

Lex Fridman (05:15.240)

How does it work?

Lex Fridman (05:17.160)

Well, okay, the simplest model, it's kind of like a slinky.

Russ Tedrake (05:19.480)

It's like an elaborate slinky.

Lex Fridman (05:21.560)

One of the simplest models we used to think about it

Russ Tedrake (05:23.840)

is actually a rimless wheel.

Lex Fridman (05:25.360)

So imagine taking a bicycle wheel, but take the rim off.

Lex Fridman (05:30.100)

So it's now just got a bunch of spokes.

Lex Fridman (05:32.640)

If you give that a push,

Russ Tedrake (05:33.720)

it still wants to roll down the ramp,

Lex Fridman (05:35.840)

but every time its foot, its spoke comes around

Lex Fridman (05:38.180)

and hits the ground, it loses a little energy.

Lex Fridman (05:41.880)

Every time it takes a step forward,

Russ Tedrake (05:43.280)

it gains a little energy.

Lex Fridman (05:45.800)

Those things can come into perfect balance.

Lex Fridman (05:48.200)

And actually they want to, it's a stable phenomenon.

Lex Fridman (05:51.240)

If it's going too slow, it'll speed up.

Russ Tedrake (05:53.720)

If it's going too fast, it'll slow down

Lex Fridman (05:55.880)

and it comes into a stable periodic motion.

Russ Tedrake (05:59.480)

Now you can take that rimless wheel,

Lex Fridman (06:02.120)

which doesn't look very much like a human walking,

Russ Tedrake (06:05.040)

take all the extra spokes away, put a hinge in the middle.

Lex Fridman (06:08.080)

Now it's two legs.

Russ Tedrake (06:09.720)

That's called our compass gait walker.

Lex Fridman (06:11.880)

That can still, you give it a little push,

Russ Tedrake (06:13.800)

it starts falling down a ramp.

Lex Fridman (06:15.520)

It looks a little bit more like walking.

Russ Tedrake (06:17.240)

At least it's a biped.

Lex Fridman (06:19.700)

But what Steve and Andy,

Lex Fridman (06:21.400)

and Tad McGeer started the whole exercise,

Lex Fridman (06:23.480)

but what Steve and Andy did was they took it

Russ Tedrake (06:25.200)

to this beautiful conclusion

Lex Fridman (06:28.700)

where they built something that had knees, arms, a torso.

Russ Tedrake (06:32.440)

The arms swung naturally, give it a little push.

Lex Fridman (06:36.320)

And that looked like a stroll through the park.

Lex Fridman (06:38.720)

How do you design something like that?

Lex Fridman (06:40.240)

I mean, is that art or science?

Russ Tedrake (06:42.360)

It's on the boundary.

Lex Fridman (06:43.800)

I think there's a science to getting close to the solution.

Russ Tedrake (06:47.640)

I think there's certainly art in the way

Lex Fridman (06:49.040)

that they made a beautiful robot.

Lex Fridman (06:52.000)

But then the finesse, because they were working

Lex Fridman (06:57.080)

with a system that wasn't perfectly modeled,

Russ Tedrake (06:58.980)

wasn't perfectly controlled,

Lex Fridman (07:01.060)

there's all these little tricks

Russ Tedrake (07:02.800)

that you have to tune the suction cups at the knees,

Lex Fridman (07:05.480)

for instance, so that they stick,

Lex Fridman (07:07.960)

but then they release at just the right time.

Lex Fridman (07:09.640)

Or there's all these little tricks of the trade,

Russ Tedrake (07:12.360)

which really are art, but it was a point.

Lex Fridman (07:14.440)

I mean, it made the point.

Russ Tedrake (07:16.200)

We were, at that time, the walking robot,

Lex Fridman (07:18.800)

the best walking robot in the world was Honda's Asmo.

Russ Tedrake (07:21.840)

Absolutely marvel of modern engineering.

Lex Fridman (07:24.120)

Is this 90s?

Russ Tedrake (07:25.240)

This was in 97 when they first released.

Lex Fridman (07:27.440)

It sort of announced P2, and then it went through.

Russ Tedrake (07:29.920)

It was Asmo by then in 2004.

Lex Fridman (07:32.360)

And it looks like this very cautious walking,

Russ Tedrake (07:37.840)

like you're walking on hot coals or something like that.

Lex Fridman (07:41.320)

I think it gets a bad rap.

Russ Tedrake (07:43.760)

Asmo is a beautiful machine.

Lex Fridman (07:45.340)

It does walk with its knees bent.

Russ Tedrake (07:47.000)

Our Atlas walking had its knees bent.

Lex Fridman (07:49.740)

But actually, Asmo was pretty fantastic.

Lex Fridman (07:52.340)

But it wasn't energy efficient.

Lex Fridman (07:54.320)

Neither was Atlas when we worked on Atlas.

Russ Tedrake (07:58.220)

None of our robots that have been that complicated

Lex Fridman (08:00.520)

have been very energy efficient.

Lex Fridman (08:04.040)

But there's a thing that happens when you do control,

Lex Fridman (08:09.680)

when you try to control a system of that complexity.

Russ Tedrake (08:12.480)

You try to use your motors to basically counteract gravity.

Lex Fridman (08:17.360)

Take whatever the world's doing to you and push back,

Russ Tedrake (08:20.680)

erase the dynamics of the world,

Lex Fridman (08:23.520)

and impose the dynamics you want

Russ Tedrake (08:25.040)

because you can make them simple and analyzable,

Lex Fridman (08:28.220)

mathematically simple.

Lex Fridman (08:30.760)

And this was a very sort of beautiful example

Lex Fridman (08:34.400)

that you don't have to do that.

Russ Tedrake (08:36.380)

You can just let go.

Lex Fridman (08:37.480)

Let physics do most of the work, right?

Lex Fridman (08:40.280)

And you just have to give it a little bit of energy.

Lex Fridman (08:42.200)

This one only walked down a ramp.

Russ Tedrake (08:43.560)

It would never walk on the flat.

Lex Fridman (08:45.340)

To walk on the flat,

Russ Tedrake (08:46.180)

you have to give a little energy at some point.

Lex Fridman (08:48.480)

But maybe instead of trying to take the forces imparted

Russ Tedrake (08:51.960)

to you by the world and replacing them,

Lex Fridman (08:55.200)

what we should be doing is letting the world push us around

Lex Fridman (08:58.200)

and we go with the flow.

Lex Fridman (08:59.360)

Very zen, very zen robot.

Russ Tedrake (09:01.280)

Yeah, but okay, so that sounds very zen,

Lex Fridman (09:03.440)

but I can also imagine how many like failed versions

Russ Tedrake (09:10.220)

they had to go through.

Lex Fridman (09:11.640)

Like how many, like, I would say it's probably,

Russ Tedrake (09:14.040)

would you say it's in the thousands

Lex Fridman (09:15.320)

that they've had to have the system fall down

Lex Fridman (09:17.920)

before they figured out how to get it?

Lex Fridman (09:19.840)

I don't know if it's thousands, but it's a lot.

Russ Tedrake (09:22.560)

It takes some patience.

Lex Fridman (09:23.560)

There's no question.

Lex Fridman (09:25.040)

So in that sense, control might help a little bit.

Lex Fridman (09:28.320)

Oh, I think everybody, even at the time,

Russ Tedrake (09:32.100)

said that the answer is to do with that with control.

Lex Fridman (09:35.020)

But it was just pointing out

Russ Tedrake (09:36.340)

that maybe the way we're doing control right now

Lex Fridman (09:39.120)

isn't the way we should.

Russ Tedrake (09:41.040)

Got it.

Lex Fridman (09:41.880)

So what about on the animal side,

Lex Fridman (09:43.800)

the ones that figured out how to move efficiently?

Lex Fridman (09:46.200)

Is there anything you find inspiring or beautiful

Lex Fridman (09:49.440)

in the movement of any particular animal?

Lex Fridman (09:51.160)

I do have a favorite example.

Russ Tedrake (09:51.980)

Okay.

Lex Fridman (09:52.820)

So it sort of goes with the passive walking idea.

Lex Fridman (09:57.160)

So is there, you know, how energy efficient are animals?

Lex Fridman (10:01.400)

Okay, there's a great series of experiments

Russ Tedrake (10:03.840)

by George Lauder at Harvard and Mike Tranofilo at MIT.

Lex Fridman (10:07.520)

They were studying fish swimming in a water tunnel.

Russ Tedrake (10:10.640)

Okay.

Lex Fridman (10:11.820)

And one of these, the type of fish they were studying

Russ Tedrake (10:15.240)

were these rainbow trout,

Lex Fridman (10:17.240)

because there was a phenomenon well understood

Russ Tedrake (10:20.360)

that rainbow trout, when they're swimming upstream

Lex Fridman (10:22.180)

in mating season, they kind of hang out behind the rocks.

Lex Fridman (10:25.120)

And it looks like, I mean,

Lex Fridman (10:26.080)

that's tiring work swimming upstream.

Russ Tedrake (10:28.080)

They're hanging out behind the rocks.

Lex Fridman (10:29.180)

Maybe there's something energetically interesting there.

Lex Fridman (10:31.980)

So they tried to recreate that.

Lex Fridman (10:33.400)

They put in this water tunnel, a rock basically,

Russ Tedrake (10:36.440)

a cylinder that had the same sort of vortex street,

Lex Fridman (10:40.560)

the eddies coming off the back of the rock

Russ Tedrake (10:42.480)

that you would see in a stream.

Lex Fridman (10:44.240)

And they put a real fish behind this

Lex Fridman (10:46.080)

and watched how it swims.

Lex Fridman (10:48.000)

And the amazing thing is that if you watch from above

Lex Fridman (10:51.960)

what the fish swims when it's not behind a rock,

Lex Fridman (10:53.800)

it has a particular gate.

Russ Tedrake (10:56.120)

You can identify the fish the same way you look

Lex Fridman (10:58.240)

at a human walking down the street.

Russ Tedrake (10:59.840)

You sort of have a sense of how a human walks.

Lex Fridman (11:02.420)

The fish has a characteristic gate.

Russ Tedrake (11:05.360)

You put that fish behind the rock, its gate changes.

Lex Fridman (11:09.160)

And what they saw was that it was actually resonating

Lex Fridman (11:12.720)

and kind of surfing between the vortices.

Lex Fridman (11:16.560)

Now, here was the experiment that really was the clincher.

Russ Tedrake (11:20.140)

Because there was still, it wasn't clear how much of that

Lex Fridman (11:22.160)

was mechanics of the fish,

Lex Fridman (11:24.000)

how much of that is control, the brain.

Lex Fridman (11:26.940)

So the clincher experiment,

Lex Fridman (11:28.480)

and maybe one of my favorites to date,

Lex Fridman (11:29.800)

although there are many good experiments.

Russ Tedrake (11:33.700)

They took, this was now a dead fish.

Lex Fridman (11:38.380)

They took a dead fish.

Russ Tedrake (11:40.200)

They put a string that went,

Lex Fridman (11:41.640)

that tied the mouth of the fish to the rock

Lex Fridman (11:44.160)

so it couldn't go back and get caught in the grates.

Lex Fridman (11:47.160)

And then they asked what would that dead fish do

Lex Fridman (11:49.180)

when it was hanging out behind the rock?

Lex Fridman (11:51.160)

And so what you'd expect, it sort of flopped around

Russ Tedrake (11:52.920)

like a dead fish in the vortex wake

Lex Fridman (11:56.120)

until something sort of amazing happens.

Lex Fridman (11:57.800)

And this video is worth putting in, right?

Lex Fridman (12:02.880)

What happens?

Lex Fridman (12:04.040)

The dead fish basically starts swimming upstream, right?

Lex Fridman (12:07.520)

It's completely dead, no brain, no motors, no control.

Lex Fridman (12:12.160)

But it's somehow the mechanics of the fish

Lex Fridman (12:14.600)

resonate with the vortex street

Lex Fridman (12:16.360)

and it starts swimming upstream.

Lex Fridman (12:18.280)

It's one of the best examples ever.

Lex Fridman (12:20.520)

Who do you give credit for that to?

Lex Fridman (12:23.740)

Is that just evolution constantly just figuring out

Russ Tedrake (12:27.980)

by killing a lot of generations of animals,

Lex Fridman (12:30.920)

like the most efficient motion?

Russ Tedrake (12:33.360)

Is that, or maybe the physics of our world completely like,

Lex Fridman (12:38.660)

is like if evolution applied not only to animals,

Lex Fridman (12:40.920)

but just the entirety of it somehow drives to efficiency,

Lex Fridman (12:45.220)

like nature likes efficiency?

Russ Tedrake (12:47.020)

I don't know if that question even makes any sense.

Lex Fridman (12:49.980)

I understand the question.

Russ Tedrake (12:51.020)

That's reasonable.

Lex Fridman (12:51.860)

I mean, do they co evolve?

Russ Tedrake (12:54.460)

Yeah, somehow co, yeah.

Lex Fridman (12:55.620)

Like I don't know if an environment can evolve, but.

Russ Tedrake (13:00.020)

I mean, there are experiments that people do,

Lex Fridman (13:02.340)

careful experiments that show that animals can adapt

Russ Tedrake (13:05.940)

to unusual situations and recover efficiency.

Lex Fridman (13:08.660)

So there seems like at least in one direction,

Russ Tedrake (13:11.100)

I think there is reason to believe

Lex Fridman (13:12.740)

that the animal's motor system and probably its mechanics

Russ Tedrake (13:18.100)

adapt in order to be more efficient.

Lex Fridman (13:20.060)

But efficiency isn't the only goal, of course.

Russ Tedrake (13:23.140)

Sometimes it's too easy to think about only efficiency,

Lex Fridman (13:26.220)

but we have to do a lot of other things first, not get eaten.

Lex Fridman (13:30.540)

And then all other things being equal, try to save energy.

Lex Fridman (13:34.140)

By the way, let's draw a distinction

Russ Tedrake (13:36.100)

between control and mechanics.

Lex Fridman (13:38.160)

Like how would you define each?

Russ Tedrake (13:40.820)

Yeah.

Lex Fridman (13:41.720)

I mean, I think part of the point is that

Russ Tedrake (13:43.940)

we shouldn't draw a line as clearly as we tend to.

Lex Fridman (13:47.860)

But on a robot, we have motors

Lex Fridman (13:51.460)

and we have the links of the robot, let's say.

Lex Fridman (13:54.840)

If the motors are turned off,

Lex Fridman (13:56.260)

the robot has some passive dynamics, okay?

Lex Fridman (13:59.780)

Gravity does the work.

Lex Fridman (14:01.380)

You can put springs, I would call that mechanics, right?

Lex Fridman (14:03.700)

If we have springs and dampers,

Russ Tedrake (14:04.940)

which our muscles are springs and dampers and tendons.

Lex Fridman (14:08.540)

But then you have something that's doing active work,

Russ Tedrake (14:10.440)

putting energy in, which are your motors on the robot.

Lex Fridman (14:13.240)

The controller's job is to send commands to the motor

Lex Fridman (14:16.580)

that add new energy into the system, right?

Lex Fridman (14:19.960)

So the mechanics and control interplay somewhere,

Russ Tedrake (14:22.820)

the divide is around, you know,

Lex Fridman (14:24.820)

did you decide to send some commands to your motor

Russ Tedrake (14:27.560)

or did you just leave the motors off,

Lex Fridman (14:28.980)

let them do their work?

Russ Tedrake (14:30.580)

Would you say is most of nature

Lex Fridman (14:35.140)

on the dynamic side or the control side?

Lex Fridman (14:39.820)

So like, if you look at biological systems,

Lex Fridman (14:43.580)

we're living in a pandemic now,

Russ Tedrake (14:45.100)

like, do you think a virus is a,

Lex Fridman (14:47.840)

do you think it's a dynamic system

Lex Fridman (14:50.100)

or is there a lot of control, intelligence?

Lex Fridman (14:54.100)

I think it's both, but I think we maybe have underestimated

Lex Fridman (14:57.040)

how important the dynamics are, right?

Lex Fridman (15:02.020)

I mean, even our bodies, the mechanics of our bodies,

Russ Tedrake (15:04.300)

certainly with exercise, they evolve.

Lex Fridman (15:06.140)

But so I actually, I lost a finger in early 2000s

Lex Fridman (15:11.060)

and it's my fifth metacarpal.

Lex Fridman (15:14.460)

And it turns out you use that a lot

Russ Tedrake (15:16.620)

in ways you don't expect when you're opening jars,

Lex Fridman (15:19.340)

even when I'm just walking around,

Russ Tedrake (15:20.620)

if I bump it on something, there's a bone there

Lex Fridman (15:23.220)

that was used to taking contact.

Russ Tedrake (15:26.780)

My fourth metacarpal wasn't used to taking contact,

Lex Fridman (15:28.820)

it used to hurt, it still does a little bit.

Lex Fridman (15:31.100)

But actually my bone has remodeled, right?

Lex Fridman (15:34.180)

Over a couple of years, the geometry,

Russ Tedrake (15:39.580)

the mechanics of that bone changed

Lex Fridman (15:42.100)

to address the new circumstances.

Lex Fridman (15:44.340)

So the idea that somehow it's only our brain

Lex Fridman (15:46.820)

that's adapting or evolving is not right.

Russ Tedrake (15:50.140)

Maybe sticking on evolution for a bit,

Lex Fridman (15:52.560)

because it's tended to create some interesting things.

Russ Tedrake (15:56.720)

Bipedal walking, why the heck did evolution give us,

Lex Fridman (16:01.720)

I think we're, are we the only mammals that walk on two feet?

Russ Tedrake (16:05.040)

No, I mean, there's a bunch of animals that do it a bit.

Lex Fridman (16:09.040)

A bit.

Russ Tedrake (16:09.880)

I think we are the most successful bipeds.

Lex Fridman (16:12.280)

I think I read somewhere that the reason

Russ Tedrake (16:17.760)

the evolution made us walk on two feet

Lex Fridman (16:22.760)

is because there's an advantage

Russ Tedrake (16:24.720)

to being able to carry food back to the tribe

Lex Fridman (16:27.200)

or something like that.

Lex Fridman (16:28.040)

So like you can carry, it's kind of this communal,

Lex Fridman (16:31.960)

cooperative thing, so like to carry stuff back

Russ Tedrake (16:35.080)

to a place of shelter and so on to share with others.

Lex Fridman (16:40.080)

Do you understand at all the value of walking on two feet

Lex Fridman (16:44.520)

from both a robotics and a human perspective?

Lex Fridman (16:48.000)

Yeah, there are some great books written

Russ Tedrake (16:50.280)

about evolution of, walking evolution of the human body.

Lex Fridman (16:54.560)

I think it's easy though to make bad evolutionary arguments.

Russ Tedrake (17:00.600)

Sure, most of them are probably bad,

Lex Fridman (17:03.740)

but what else can we do?

Russ Tedrake (17:06.200)

I mean, I think a lot of what dominated our evolution

Lex Fridman (17:11.120)

probably was not the things that worked well

Russ Tedrake (17:15.080)

sort of in the steady state, you know,

Lex Fridman (17:18.560)

when things are good, but for instance,

Russ Tedrake (17:22.800)

people talk about what we should eat now

Lex Fridman (17:25.040)

because our ancestors were meat eaters or whatever.

Russ Tedrake (17:28.320)

Oh yeah, I love that, yeah.

Lex Fridman (17:30.240)

But probably, you know, the reason

Russ Tedrake (17:32.520)

that one pre Homo sapiens species versus another survived

Lex Fridman (17:39.640)

was not because of whether they ate well

Russ Tedrake (17:43.440)

when there was lots of food.

Lex Fridman (17:45.300)

But when the ice age came, you know,

Russ Tedrake (17:47.920)

probably one of them happened to be in the wrong place.

Lex Fridman (17:50.940)

One of them happened to forage a food that was okay

Russ Tedrake (17:54.200)

even when the glaciers came or something like that, I mean.

Lex Fridman (17:58.240)

There's a million variables that contributed

Lex Fridman (18:00.560)

and we can't, and our, actually the amount of information

Lex Fridman (18:04.080)

we're working with and telling these stories,

Russ Tedrake (18:06.680)

these evolutionary stories is very little.

Lex Fridman (18:10.220)

So yeah, just like you said, it seems like,

Russ Tedrake (18:13.080)

if you study history, it seems like history turns

Lex Fridman (18:15.680)

on like these little events that otherwise

Russ Tedrake (18:20.280)

would seem meaningless, but in a grant,

Lex Fridman (18:23.320)

like when you, in retrospect, were turning points.

Russ Tedrake (18:27.560)

Absolutely.

Lex Fridman (18:28.400)

And that's probably how like somebody got hit in the head

Russ Tedrake (18:31.280)

with a rock because somebody slept with the wrong person

Lex Fridman (18:35.160)

back in the cave days and somebody get angry

Lex Fridman (18:38.500)

and that turned, you know, warring tribes

Lex Fridman (18:41.920)

combined with the environment, all those millions of things

Lex Fridman (18:45.360)

and the meat eating, which I get a lot of criticism

Lex Fridman (18:47.680)

because I don't know what your dietary processes are like,

Lex Fridman (18:51.480)

but these days I've been eating only meat,

Lex Fridman (18:55.040)

which is, there's a large community of people who say,

Russ Tedrake (18:59.080)

yeah, probably make evolutionary arguments

Lex Fridman (19:01.080)

and say you're doing a great job.

Russ Tedrake (19:02.720)

There's probably an even larger community of people,

Lex Fridman (19:05.760)

including my mom, who says it's deeply unhealthy,

Russ Tedrake (19:08.520)

it's wrong, but I just feel good doing it.

Lex Fridman (19:10.760)

But you're right, these evolutionary arguments

Russ Tedrake (19:12.980)

can be flawed, but is there anything interesting

Lex Fridman (19:15.420)

to pull out for?

Russ Tedrake (19:17.320)

There's a great book, by the way,

Lex Fridman (19:19.360)

well, a series of books by Nicholas Taleb

Russ Tedrake (19:21.280)

about Fooled by Randomness and Black Swan.

Lex Fridman (19:24.800)

Highly recommend them, but yeah,

Russ Tedrake (19:26.840)

they make the point nicely that probably

Lex Fridman (19:29.160)

it was a few random events that, yes,

Russ Tedrake (19:34.360)

maybe it was someone getting hit by a rock, as you say.

Lex Fridman (19:39.520)

That said, do you think, I don't know how to ask this

Russ Tedrake (19:42.700)

question or how to talk about this,

Lex Fridman (19:44.080)

but there's something elegant and beautiful

Russ Tedrake (19:45.680)

about moving on two feet, obviously biased

Lex Fridman (19:48.800)

because I'm human, but from a robotics perspective, too,

Russ Tedrake (19:53.280)

you work with robots on two feet,

Lex Fridman (19:56.440)

is it all useful to build robots that are on two feet

Russ Tedrake (1:00:02.280)

in order to test.

Lex Fridman (1:00:04.220)

Maybe that's all we'll ever be able to,

Lex Fridman (1:00:05.480)

but I think, you know,

Lex Fridman (1:00:07.320)

cause there's an argument that the things that'll get you

Russ Tedrake (1:00:10.520)

are the things that are really nuanced in the world.

Lex Fridman (1:00:14.040)

And there'd be very hard to, for instance,

Russ Tedrake (1:00:15.700)

put back in a simulation.

Lex Fridman (1:00:16.880)

Yeah, I guess the edge cases.

Lex Fridman (1:00:19.880)

What was the hardest thing?

Lex Fridman (1:00:21.840)

Like, so you said walking over rough terrain,

Russ Tedrake (1:00:24.680)

like just taking footsteps.

Lex Fridman (1:00:27.120)

I mean, people, it's so dramatic and painful

Russ Tedrake (1:00:31.360)

in a certain kind of way to watch these videos

Lex Fridman (1:00:33.520)

from the DRC of robots falling.

Russ Tedrake (1:00:37.600)

Yep.

Lex Fridman (1:00:38.440)

It's just so heartbreaking.

Russ Tedrake (1:00:39.440)

I don't know.

Lex Fridman (1:00:40.280)

Maybe it's because for me at least,

Russ Tedrake (1:00:42.400)

we anthropomorphize the robot.

Lex Fridman (1:00:45.120)

Of course, it's also funny for some reason,

Russ Tedrake (1:00:48.400)

like humans falling is funny for, I don't,

Lex Fridman (1:00:51.920)

it's some dark reason.

Russ Tedrake (1:00:53.400)

I'm not sure why it is so,

Lex Fridman (1:00:55.300)

but it's also like tragic and painful.

Lex Fridman (1:00:57.880)

And so speaking of which, I mean,

Lex Fridman (1:01:00.380)

what made the robots fall and fail in your view?

Lex Fridman (1:01:05.000)

So I can tell you exactly what happened on our,

Lex Fridman (1:01:06.960)

we, I contributed one of those.

Russ Tedrake (1:01:08.360)

Our team contributed one of those spectacular falls.

Lex Fridman (1:01:10.960)

Every one of those falls has a complicated story.

Russ Tedrake (1:01:15.560)

I mean, at one time,

Lex Fridman (1:01:16.920)

the power effectively went out on the robot

Russ Tedrake (1:01:20.200)

because it had been sitting at the door

Lex Fridman (1:01:21.720)

waiting for a green light to be able to proceed

Lex Fridman (1:01:24.400)

and its batteries, you know,

Lex Fridman (1:01:26.280)

and therefore it just fell backwards

Lex Fridman (1:01:28.080)

and smashed its head against the ground.

Lex Fridman (1:01:29.280)

And it was hilarious,

Lex Fridman (1:01:30.120)

but it wasn't because of bad software, right?

Lex Fridman (1:01:34.100)

But for ours, so the hardest part of the challenge,

Russ Tedrake (1:01:37.120)

the hardest task in my view was getting out of the Polaris.

Lex Fridman (1:01:40.400)

It was actually relatively easy to drive the Polaris.

Lex Fridman (1:01:43.760)

Can you tell the story?

Lex Fridman (1:01:44.600)

Sorry to interrupt.

Russ Tedrake (1:01:45.440)

The story of the car.

Lex Fridman (1:01:50.040)

People should watch this video.

Russ Tedrake (1:01:51.240)

I mean, the thing you've come up with is just brilliant,

Lex Fridman (1:01:53.900)

but anyway, sorry, what's...

Russ Tedrake (1:01:55.920)

Yeah, we kind of joke.

Lex Fridman (1:01:56.920)

We call it the big robot, little car problem

Russ Tedrake (1:01:59.040)

because somehow the race organizers decided

Lex Fridman (1:02:03.440)

to give us a 400 pound humanoid.

Lex Fridman (1:02:05.360)

And then they also provided the vehicle,

Lex Fridman (1:02:07.480)

which was a little Polaris.

Lex Fridman (1:02:08.640)

And the robot didn't really fit in the car.

Lex Fridman (1:02:11.760)

So you couldn't drive the car with your feet

Russ Tedrake (1:02:14.520)

under the steering column.

Lex Fridman (1:02:15.720)

We actually had to straddle the main column of the,

Lex Fridman (1:02:21.280)

and have basically one foot in the passenger seat,

Lex Fridman (1:02:23.580)

one foot in the driver's seat,

Lex Fridman (1:02:25.280)

and then drive with our left hand.

Lex Fridman (1:02:28.880)

But the hard part was we had to then park the car,

Russ Tedrake (1:02:31.300)

get out of the car.

Lex Fridman (1:02:33.080)

It didn't have a door, that was okay.

Lex Fridman (1:02:34.320)

But it's just getting up from crouched, from sitting,

Lex Fridman (1:02:38.720)

when you're in this very constrained environment.

Russ Tedrake (1:02:41.880)

First of all, I remember after watching those videos,

Lex Fridman (1:02:44.320)

I was much more cognizant of how hard it is for me

Russ Tedrake (1:02:47.840)

to get in and out of the car,

Lex Fridman (1:02:49.600)

and out of the car, especially.

Russ Tedrake (1:02:51.760)

It's actually a really difficult control problem.

Lex Fridman (1:02:54.240)

Yeah.

Russ Tedrake (1:02:55.480)

I'm very cognizant of it when I'm like injured

Lex Fridman (1:02:58.360)

for whatever reason.

Russ Tedrake (1:02:59.200)

Oh, that's really hard.

Lex Fridman (1:03:00.120)

Yeah.

Lex Fridman (1:03:01.440)

So how did you approach this problem?

Lex Fridman (1:03:03.560)

So we had, you think of NASA's operations,

Lex Fridman (1:03:08.160)

and they have these checklists,

Lex Fridman (1:03:09.800)

prelaunched checklists and the like.

Russ Tedrake (1:03:11.080)

We weren't far off from that.

Lex Fridman (1:03:12.380)

We had this big checklist.

Lex Fridman (1:03:13.500)

And on the first day of the competition,

Lex Fridman (1:03:16.320)

we were running down our checklist.

Lex Fridman (1:03:17.520)

And one of the things we had to do,

Lex Fridman (1:03:19.120)

we had to turn off the controller,

Russ Tedrake (1:03:21.320)

the piece of software that was running

Lex Fridman (1:03:23.320)

that would drive the left foot of the robot

Russ Tedrake (1:03:25.560)

in order to accelerate on the gas.

Lex Fridman (1:03:28.120)

And then we turned on our balancing controller.

Lex Fridman (1:03:30.840)

And the nerves, jitters of the first day of the competition,

Lex Fridman (1:03:34.280)

someone forgot to check that box

Lex Fridman (1:03:35.660)

and turn that controller off.

Lex Fridman (1:03:37.560)

So we used a lot of motion planning

Russ Tedrake (1:03:40.880)

to figure out a sort of configuration of the robot

Lex Fridman (1:03:45.320)

that we could get up and over.

Russ Tedrake (1:03:47.200)

We relied heavily on our balancing controller.

Lex Fridman (1:03:50.320)

And basically, when the robot was in one

Russ Tedrake (1:03:53.760)

of its most precarious sort of configurations,

Lex Fridman (1:03:57.560)

trying to sneak its big leg out of the side,

Russ Tedrake (1:04:01.800)

the other controller that thought it was still driving

Lex Fridman (1:04:05.000)

told its left foot to go like this.

Lex Fridman (1:04:06.760)

And that wasn't good.

Lex Fridman (1:04:11.000)

But it turned disastrous for us

Russ Tedrake (1:04:13.320)

because what happened was a little bit of push here.

Lex Fridman (1:04:16.980)

Actually, we have videos of us running into the robot

Russ Tedrake (1:04:21.080)

with a 10 foot pole and it kind of will recover.

Lex Fridman (1:04:24.680)

But this is a case where there's no space to recover.

Lex Fridman (1:04:27.800)

So a lot of our secondary balancing mechanisms

Lex Fridman (1:04:30.180)

about like take a step to recover,

Russ Tedrake (1:04:32.160)

they were all disabled because we were in the car

Lex Fridman (1:04:33.760)

and there was no place to step.

Lex Fridman (1:04:35.320)

So we were relying on our just lowest level reflexes.

Lex Fridman (1:04:38.380)

And even then, I think just hitting the foot on the seat,

Russ Tedrake (1:04:42.200)

on the floor, we probably could have recovered from it.

Lex Fridman (1:04:44.960)

But the thing that was bad that happened

Russ Tedrake (1:04:46.400)

is when we did that and we jostled a little bit,

Lex Fridman (1:04:49.440)

the tailbone of our robot was only a little off the seat,

Russ Tedrake (1:04:53.720)

it hit the seat.

Lex Fridman (1:04:55.480)

And the other foot came off the ground just a little bit.

Lex Fridman (1:04:58.260)

And nothing in our plans had ever told us what to do

Lex Fridman (1:05:02.280)

if your butt's on the seat and your feet are in the air.

Russ Tedrake (1:05:05.120)

Feet in the air.

Lex Fridman (1:05:06.040)

And then the thing is once you get off the script,

Russ Tedrake (1:05:10.080)

things can go very wrong

Lex Fridman (1:05:11.040)

because even our state estimation,

Russ Tedrake (1:05:12.760)

our system that was trying to collect all the data

Lex Fridman (1:05:15.200)

from the sensors and understand

Russ Tedrake (1:05:16.760)

what's happening with the robot,

Lex Fridman (1:05:18.480)

it didn't know about this situation.

Lex Fridman (1:05:20.080)

So it was predicting things that were just wrong.

Lex Fridman (1:05:22.800)

And then we did a violent shake and fell off

Russ Tedrake (1:05:26.560)

in our face first out of the robot.

Lex Fridman (1:05:29.180)

But like into the destination.

Russ Tedrake (1:05:32.520)

That's true, we fell in, we got our point for egress.

Lex Fridman (1:05:36.320)

But so is there any hope for, that's interesting,

Russ Tedrake (1:05:39.280)

is there any hope for Atlas to be able to do something

Lex Fridman (1:05:43.280)

when it's just on its butt and feet in the air?

Russ Tedrake (1:05:46.320)

Absolutely.

Lex Fridman (1:05:47.200)

So you can, what do you?

Russ Tedrake (1:05:48.520)

No, so that is one of the big challenges.

Lex Fridman (1:05:50.920)

And I think it's still true, you know,

Russ Tedrake (1:05:53.840)

Boston Dynamics and Antimal and there's this incredible work

Lex Fridman (1:05:59.120)

on legged robots happening around the world.

Russ Tedrake (1:06:04.540)

Most of them still are very good at the case

Lex Fridman (1:06:07.620)

where you're making contact with the world at your feet.

Lex Fridman (1:06:10.080)

And they have typically point feet relatively,

Lex Fridman (1:06:12.200)

they have balls on their feet, for instance.

Russ Tedrake (1:06:14.480)

If those robots get in a situation

Lex Fridman (1:06:16.600)

where the elbow hits the wall or something like this,

Russ Tedrake (1:06:19.880)

that's a pretty different situation.

Lex Fridman (1:06:21.240)

Now they have layers of mechanisms that will make,

Russ Tedrake (1:06:24.080)

I think the more mature solutions have ways

Lex Fridman (1:06:27.680)

in which the controller won't do stupid things.

Lex Fridman (1:06:31.240)

But a human, for instance, is able to leverage

Lex Fridman (1:06:34.720)

incidental contact in order to accomplish a goal.

Russ Tedrake (1:06:36.760)

In fact, I might, if you push me,

Lex Fridman (1:06:37.800)

I might actually put my hand out

Lex Fridman (1:06:39.720)

and make a new brand new contact.

Lex Fridman (1:06:42.220)

The feet of the robot are doing this on quadrupeds,

Lex Fridman (1:06:44.940)

but we mostly in robotics are afraid of contact

Lex Fridman (1:06:49.120)

on the rest of our body, which is crazy.

Russ Tedrake (1:06:53.180)

There's this whole field of motion planning,

Lex Fridman (1:06:56.040)

collision free motion planning.

Lex Fridman (1:06:58.040)

And we write very complex algorithms

Lex Fridman (1:06:59.800)

so that the robot can dance around

Lex Fridman (1:07:01.640)

and make sure it doesn't touch the world.

Lex Fridman (1:07:05.840)

So people are just afraid of contact

Russ Tedrake (1:07:07.720)

because contact the scene is a difficult.

Lex Fridman (1:07:09.880)

It's still a difficult control problem and sensing problem.

Russ Tedrake (1:07:13.380)

Now you're a serious person, I'm a little bit of an idiot

Lex Fridman (1:07:21.180)

and I'm going to ask you some dumb questions.

Lex Fridman (1:07:24.140)

So I do martial arts.

Lex Fridman (1:07:27.140)

So like jiu jitsu, I wrestled my whole life.

Lex Fridman (1:07:30.380)

So let me ask the question, like whenever people learn

Lex Fridman (1:07:35.380)

that I do any kind of AI or like I mentioned robots

Lex Fridman (1:07:38.500)

and things like that, they say,

Lex Fridman (1:07:40.040)

when are we going to have robots that can win

Lex Fridman (1:07:45.020)

in a wrestling match or in a fight against a human?

Lex Fridman (1:07:49.880)

So we just mentioned sitting on your butt,

Russ Tedrake (1:07:52.160)

if you're in the air, that's a common position.

Lex Fridman (1:07:53.940)

Jiu jitsu, when you're on the ground,

Russ Tedrake (1:07:55.420)

you're a down opponent.

Lex Fridman (1:07:59.100)

Like how difficult do you think is the problem?

Lex Fridman (1:08:03.800)

And when will we have a robot that can defeat a human

Lex Fridman (1:08:06.880)

in a wrestling match?

Lex Fridman (1:08:08.580)

And we're talking about a lot, like, I don't know

Lex Fridman (1:08:11.100)

if you're familiar with wrestling, but essentially.

Russ Tedrake (1:08:15.340)

Not very.

Lex Fridman (1:08:16.180)

It's basically the art of contact.

Russ Tedrake (1:08:19.580)

It's like, it's because you're picking contact points

Lex Fridman (1:08:24.580)

and then using like leverage like to off balance

Russ Tedrake (1:08:29.300)

to trick people, like you make them feel

Lex Fridman (1:08:33.940)

like you're doing one thing

Lex Fridman (1:08:35.620)

and then they change their balance

Lex Fridman (1:08:38.840)

and then you switch what you're doing

Lex Fridman (1:08:41.620)

and then results in a throw or whatever.

Lex Fridman (1:08:44.100)

So like, it's basically the art of multiple contacts.

Russ Tedrake (1:08:48.540)

So.

Lex Fridman (1:08:49.380)

Awesome, that's a nice description of it.

Lex Fridman (1:08:50.820)

So there's also an opponent in there, right?

Lex Fridman (1:08:53.040)

So if.

Russ Tedrake (1:08:54.180)

Very dynamic.

Lex Fridman (1:08:55.060)

Right, if you are wrestling a human

Lex Fridman (1:08:58.520)

and are in a game theoretic situation with a human,

Lex Fridman (1:09:02.900)

that's still hard, but just to speak to the, you know,

Russ Tedrake (1:09:08.220)

quickly reasoning about contact part of it, for instance.

Lex Fridman (1:09:11.340)

Yeah, maybe even throwing the game theory out of it,

Russ Tedrake (1:09:13.380)

almost like, yeah, almost like a non dynamic opponent.

Lex Fridman (1:09:17.700)

Right, there's reasons to be optimistic,

Lex Fridman (1:09:20.060)

but I think our best understanding of those problems

Lex Fridman (1:09:22.660)

are still pretty hard.

Russ Tedrake (1:09:24.820)

I have been increasingly focused on manipulation,

Lex Fridman (1:09:29.860)

partly where that's a case where the contact

Russ Tedrake (1:09:31.720)

has to be much more rich.

Lex Fridman (1:09:35.800)

And there are some really impressive examples

Russ Tedrake (1:09:38.260)

of deep learning policies, controllers

Lex Fridman (1:09:41.820)

that can appear to do good things through contact.

Russ Tedrake (1:09:47.860)

We've even got new examples of, you know,

Lex Fridman (1:09:51.380)

deep learning models of predicting what's gonna happen

Russ Tedrake (1:09:53.940)

to objects as they go through contact.

Lex Fridman (1:09:56.220)

But I think the challenge you just offered there

Lex Fridman (1:09:59.780)

still eludes us, right?

Lex Fridman (1:10:01.500)

The ability to make a decision

Russ Tedrake (1:10:03.620)

based on those models quickly.

Lex Fridman (1:10:07.560)

You know, I have to think though, it's hard for humans too,

Russ Tedrake (1:10:10.140)

when you get that complicated.

Lex Fridman (1:10:11.380)

I think probably you had maybe a slow motion version

Russ Tedrake (1:10:16.100)

of where you learned the basic skills

Lex Fridman (1:10:17.980)

and you've probably gotten better at it

Lex Fridman (1:10:20.700)

and there's much more subtle to you.

Lex Fridman (1:10:24.660)

But it might still be hard to actually, you know,

Russ Tedrake (1:10:27.940)

really on the fly take a, you know, model of your humanoid

Lex Fridman (1:10:32.140)

and figure out how to plan the optimal sequence.

Russ Tedrake (1:10:35.260)

That might be a problem we never solve.

Lex Fridman (1:10:36.660)

Well, the, I mean, one of the most amazing things to me

Russ Tedrake (1:10:40.360)

about the, we can talk about martial arts.

Lex Fridman (1:10:43.740)

We could also talk about dancing.

Russ Tedrake (1:10:45.340)

Doesn't really matter.

Lex Fridman (1:10:46.740)

Too human, I think it's the most interesting study

Russ Tedrake (1:10:50.540)

of contact.

Lex Fridman (1:10:51.380)

It's not even the dynamic element of it.

Russ Tedrake (1:10:53.040)

It's the, like when you get good at it, it's so effortless.

Lex Fridman (1:10:58.740)

Like I can just, I'm very cognizant

Russ Tedrake (1:11:00.900)

of the entirety of the learning process

Lex Fridman (1:11:03.380)

being essentially like learning how to move my body

Russ Tedrake (1:11:07.660)

in a way that I could throw very large weights

Lex Fridman (1:11:12.220)

around effortlessly, like, and I can feel the learning.

Russ Tedrake (1:11:18.500)

Like I'm a huge believer in drilling of techniques

Lex Fridman (1:11:21.540)

and you can just like feel your, I don't,

Russ Tedrake (1:11:23.580)

you're not feeling, you're feeling, sorry,

Lex Fridman (1:11:26.780)

you're learning it intellectually a little bit,

Lex Fridman (1:11:29.800)

but a lot of it is the body learning it somehow,

Lex Fridman (1:11:32.820)

like instinctually and whatever that learning is,

Russ Tedrake (1:11:36.100)

that's really, I'm not even sure if that's equivalent

Lex Fridman (1:11:40.780)

to like a deep learning, learning a controller.

Russ Tedrake (1:11:44.760)

I think it's something more,

Lex Fridman (1:11:46.820)

it feels like there's a lot of distributed learning

Russ Tedrake (1:11:49.720)

going on.

Lex Fridman (1:11:50.560)

Yeah, I think there's hierarchy and composition

Russ Tedrake (1:11:56.440)

probably in the systems that we don't capture very well yet.

Lex Fridman (1:12:00.840)

You have layers of control systems.

Russ Tedrake (1:12:02.440)

You have reflexes at the bottom layer

Lex Fridman (1:12:03.960)

and you have a system that's capable

Russ Tedrake (1:12:07.440)

of planning a vacation to some distant country,

Lex Fridman (1:12:11.320)

which is probably, you probably don't have a controller,

Russ Tedrake (1:12:14.240)

a policy for every possible destination you'll ever pick.

Lex Fridman (1:12:18.260)

Right?

Lex Fridman (1:12:20.380)

But there's something magical in the in between

Lex Fridman (1:12:23.460)

and how do you go from these low level feedback loops

Russ Tedrake (1:12:26.340)

to something that feels like a pretty complex

Lex Fridman (1:12:30.020)

set of outcomes.

Russ Tedrake (1:12:32.740)

You know, my guess is, I think there's evidence

Lex Fridman (1:12:34.760)

that you can plan at some of these levels, right?

Lex Fridman (1:12:37.620)

So Josh Tenenbaum just showed it in his talk the other day.

Lex Fridman (1:12:41.740)

He's got a game he likes to talk about.

Russ Tedrake (1:12:43.320)

I think he calls it the pick three game or something,

Lex Fridman (1:12:46.700)

where he puts a bunch of clutter down in front of a person

Lex Fridman (1:12:50.740)

and he says, okay, pick three objects.

Lex Fridman (1:12:52.380)

And it might be a telephone or a shoe

Russ Tedrake (1:12:55.700)

or a Kleenex box or whatever.

Lex Fridman (1:12:59.880)

And apparently you pick three items and then you pick,

Russ Tedrake (1:13:01.820)

he says, okay, pick the first one up with your right hand,

Lex Fridman (1:13:04.100)

the second one up with your left hand.

Russ Tedrake (1:13:06.360)

Now using those objects, now as tools,

Lex Fridman (1:13:08.860)

pick up the third object.

Russ Tedrake (1:13:11.060)

Right, so that's down at the level of physics

Lex Fridman (1:13:15.700)

and mechanics and contact mechanics

Russ Tedrake (1:13:17.140)

that I think we do learning or we do have policies for,

Lex Fridman (1:13:21.880)

we do control for, almost feedback,

Lex Fridman (1:13:24.740)

but somehow we're able to still,

Lex Fridman (1:13:26.300)

I mean, I've never picked up a telephone

Russ Tedrake (1:13:28.420)

with a shoe and a water bottle before.

Lex Fridman (1:13:30.220)

And somehow, and it takes me a little longer to do that

Russ Tedrake (1:13:33.140)

the first time, but most of the time

Lex Fridman (1:13:35.180)

we can sort of figure that out.

Lex Fridman (1:13:37.260)

So yeah, I think the amazing thing is this ability

Lex Fridman (1:13:41.940)

to be flexible with our models,

Russ Tedrake (1:13:44.100)

plan when we need to use our well oiled controllers

Lex Fridman (1:13:48.700)

when we don't, when we're in familiar territory.

Russ Tedrake (1:13:53.280)

Having models, I think the other thing you just said

Lex Fridman (1:13:55.560)

was something about, I think your awareness

Russ Tedrake (1:13:58.140)

of what's happening is even changing

Lex Fridman (1:13:59.860)

as you improve your expertise, right?

Lex Fridman (1:14:02.380)

So maybe you have a very approximate model

Lex Fridman (1:14:04.980)

of the mechanics to begin with.

Lex Fridman (1:14:06.240)

And as you gain expertise,

Lex Fridman (1:14:09.300)

you get a more refined version of that model.

Russ Tedrake (1:14:11.920)

You're aware of muscles or balance components

Lex Fridman (1:14:17.100)

that you just weren't even aware of before.

Lex Fridman (1:14:19.700)

So how do you scaffold that?

Lex Fridman (1:14:21.740)

Yeah, plus the fear of injury,

Russ Tedrake (1:14:24.180)

the ambition of goals, of excelling,

Lex Fridman (1:14:28.780)

and fear of mortality.

Lex Fridman (1:14:32.020)

Let's see, what else is in there?

Lex Fridman (1:14:33.340)

As the motivations, overinflated ego in the beginning,

Lex Fridman (1:14:38.040)

and then a crash of confidence in the middle.

Lex Fridman (1:14:42.900)

All of those seem to be essential for the learning process.

Lex Fridman (1:14:46.700)

And if all that's good,

Lex Fridman (1:14:48.140)

then you're probably optimizing energy efficiency.

Russ Tedrake (1:14:50.500)

Yeah, right, so we have to get that right.

Lex Fridman (1:14:53.080)

So there was this idea that you would have robots

Russ Tedrake (1:14:58.580)

play soccer better than human players by 2050.

Lex Fridman (1:15:03.780)

That was the goal.

Russ Tedrake (1:15:05.300)

Basically, it was the goal to beat world champion team,

Lex Fridman (1:15:10.140)

to become a world cup, beat like a world cup level team.

Lex Fridman (1:15:13.340)

So are we gonna see that first?

Lex Fridman (1:15:15.900)

Or a robot, if you're familiar,

Russ Tedrake (1:15:19.580)

there's an organization called UFC for mixed martial arts.

Lex Fridman (1:15:23.440)

Are we gonna see a world cup championship soccer team

Russ Tedrake (1:15:27.100)

that have robots, or a UFC champion mixed martial artist

Lex Fridman (1:15:32.660)

as a robot?

Russ Tedrake (1:15:33.860)

I mean, it's very hard to say one thing is harder,

Lex Fridman (1:15:37.140)

some problem is harder than the other.

Lex Fridman (1:15:38.580)

What probably matters is who started the organization that,

Lex Fridman (1:15:44.980)

I mean, I think RoboCup has a pretty serious following,

Lex Fridman (1:15:47.140)

and there is a history now of people playing that game,

Lex Fridman (1:15:50.860)

learning about that game, building robots to play that game,

Russ Tedrake (1:15:53.620)

building increasingly more human robots.

Lex Fridman (1:15:55.820)

It's got momentum.

Lex Fridman (1:15:57.020)

So if you want to have mixed martial arts compete,

Lex Fridman (1:16:00.900)

you better start your organization now, right?

Russ Tedrake (1:16:05.460)

I think almost independent of which problem

Lex Fridman (1:16:07.740)

is technically harder,

Russ Tedrake (1:16:08.660)

because they're both hard and they're both different.

Lex Fridman (1:16:11.400)

That's a good point.

Russ Tedrake (1:16:12.240)

I mean, those videos are just hilarious,

Lex Fridman (1:16:14.700)

like especially the humanoid robots

Russ Tedrake (1:16:17.140)

trying to play soccer.

Lex Fridman (1:16:21.260)

I mean, they're kind of terrible right now.

Russ Tedrake (1:16:23.420)

I mean, I guess there is robo sumo wrestling.

Lex Fridman (1:16:26.020)

There's like the robo one competitions,

Russ Tedrake (1:16:28.740)

where they do have these robots that go on the table

Lex Fridman (1:16:31.140)

and basically fight.

Lex Fridman (1:16:32.100)

So maybe I'm wrong, maybe.

Lex Fridman (1:16:33.720)

First of all, do you have a year in mind for RoboCup,

Lex Fridman (1:16:37.140)

just from a robotics perspective?

Lex Fridman (1:16:39.100)

Seems like a super exciting possibility

Russ Tedrake (1:16:42.060)

that like in the physical space,

Lex Fridman (1:16:46.340)

this is what's interesting.

Russ Tedrake (1:16:47.620)

I think the world is captivated.

Lex Fridman (1:16:50.560)

I think it's really exciting.

Russ Tedrake (1:16:52.620)

It inspires just a huge number of people

Lex Fridman (1:16:56.400)

when a machine beats a human at a game

Russ Tedrake (1:17:01.460)

that humans are really damn good at.

Lex Fridman (1:17:03.460)

So you're talking about chess and go,

Lex Fridman (1:17:05.740)

but that's in the world of digital.

Lex Fridman (1:17:09.820)

I don't think machines have beat humans

Russ Tedrake (1:17:13.320)

at a game in the physical space yet,

Lex Fridman (1:17:16.020)

but that would be just.

Lex Fridman (1:17:17.700)

You have to make the rules very carefully, right?

Lex Fridman (1:17:20.340)

I mean, if Atlas kicked me in the shins, I'm down

Lex Fridman (1:17:22.980)

and game over.

Lex Fridman (1:17:25.440)

So it's very subtle on what's fair.

Russ Tedrake (1:17:31.220)

I think the fighting one is a weird one.

Lex Fridman (1:17:33.020)

Yeah, because you're talking about a machine

Russ Tedrake (1:17:35.180)

that's much stronger than you.

Lex Fridman (1:17:36.500)

But yeah, in terms of soccer, basketball, all those kinds.

Lex Fridman (1:17:39.740)

Even soccer, right?

Lex Fridman (1:17:40.580)

I mean, as soon as there's contact or whatever,

Lex Fridman (1:17:43.500)

and there are some things that the robot will do better.

Lex Fridman (1:17:46.540)

I think if you really set yourself up to try to see

Russ Tedrake (1:17:51.540)

could robots win the game of soccer

Lex Fridman (1:17:53.140)

as the rules were written, the right thing

Russ Tedrake (1:17:56.300)

for the robot to do is to play very differently

Lex Fridman (1:17:58.060)

than a human would play.

Russ Tedrake (1:17:59.680)

You're not gonna get the perfect soccer player robot.

Lex Fridman (1:18:04.060)

You're gonna get something that exploits the rules,

Russ Tedrake (1:18:07.900)

exploits its super actuators, its super low bandwidth

Lex Fridman (1:18:13.420)

feedback loops or whatever, and it's gonna play the game

Russ Tedrake (1:18:15.340)

differently than you want it to play.

Lex Fridman (1:18:17.540)

And I bet there's ways, I bet there's loopholes, right?

Russ Tedrake (1:18:21.380)

We saw that in the DARPA challenge that it's very hard

Lex Fridman (1:18:27.060)

to write a set of rules that someone can't find

Russ Tedrake (1:18:30.660)

a way to exploit.

Lex Fridman (1:18:32.860)

Let me ask another ridiculous question.

Russ Tedrake (1:18:35.020)

I think this might be the last ridiculous question,

Lex Fridman (1:18:37.980)

but I doubt it.

Russ Tedrake (1:18:39.220)

I aspire to ask as many ridiculous questions

Lex Fridman (1:18:44.540)

of a brilliant MIT professor.

Russ Tedrake (1:18:48.060)

Okay, I don't know if you've seen the black mirror.

Lex Fridman (1:18:53.660)

It's funny, I never watched the episode.

Russ Tedrake (1:18:56.740)

I know when it happened though, because I gave a talk

Lex Fridman (1:19:00.620)

to some MIT faculty one day on a unassuming Monday

Russ Tedrake (1:19:05.380)

or whatever I was telling him about the state of robotics.

Lex Fridman (1:19:08.500)

And I showed some video from Boston Dynamics

Russ Tedrake (1:19:10.740)

of the quadruped spot at the time.

Lex Fridman (1:19:13.940)

It was the early version of spot.

Lex Fridman (1:19:15.900)

And there was a look of horror that went across the room.

Lex Fridman (1:19:19.300)

And I said, I've shown videos like this a lot of times,

Lex Fridman (1:19:23.220)

what happened?

Lex Fridman (1:19:24.060)

And it turns out that this video had gone,

Russ Tedrake (1:19:26.780)

this black mirror episode had changed

Lex Fridman (1:19:28.380)

the way people watched the videos I was putting out.

Russ Tedrake (1:19:33.180)

The way they see these kinds of robots.

Lex Fridman (1:19:34.740)

So I talked to so many people who are just terrified

Russ Tedrake (1:19:37.780)

because of that episode probably of these kinds of robots.

Lex Fridman (1:19:41.020)

I almost wanna say that they almost enjoy being terrified.

Russ Tedrake (1:19:44.540)

I don't even know what it is about human psychology

Lex Fridman (1:19:47.100)

that kind of imagine doomsday,

Russ Tedrake (1:19:49.220)

the destruction of the universe or our society

Lex Fridman (1:19:52.780)

and kind of like enjoy being afraid.

Russ Tedrake (1:19:57.340)

I don't wanna simplify it, but it feels like

Lex Fridman (1:19:59.300)

they talk about it so often.

Russ Tedrake (1:20:01.020)

It almost, there does seem to be an addictive quality to it.

Lex Fridman (1:20:06.380)

I talked to a guy, a guy named Joe Rogan,

Russ Tedrake (1:20:09.500)

who's kind of the flag bearer

Lex Fridman (1:20:11.580)

for being terrified at these robots.

Lex Fridman (1:20:14.660)

Do you have two questions?

Lex Fridman (1:20:17.340)

One, do you have an understanding

Lex Fridman (1:20:18.620)

of why people are afraid of robots?

Lex Fridman (1:20:21.700)

And the second question is in black mirror,

Russ Tedrake (1:20:24.940)

just to tell you the episode,

Lex Fridman (1:20:26.380)

I don't even remember it that much anymore,

Lex Fridman (1:20:28.180)

but these robots, I think they can shoot

Lex Fridman (1:20:31.100)

like a pellet or something.

Russ Tedrake (1:20:32.820)

They basically have, it's basically a spot with a gun.

Lex Fridman (1:20:36.540)

And how far are we away from having robots

Lex Fridman (1:20:41.940)

that go rogue like that?

Lex Fridman (1:20:44.100)

Basically spot that goes rogue for some reason

Lex Fridman (1:20:48.460)

and somehow finds a gun.

Lex Fridman (1:20:51.300)

Right, so, I mean, I'm not a psychologist.

Russ Tedrake (1:20:56.420)

I think, I don't know exactly why

Lex Fridman (1:20:59.860)

people react the way they do.

Russ Tedrake (1:21:01.700)

I think we have to be careful about the way robots influence

Lex Fridman (1:21:06.700)

our society and the like.

Russ Tedrake (1:21:07.980)

I think that's something, that's a responsibility

Lex Fridman (1:21:09.860)

that roboticists need to embrace.

Russ Tedrake (1:21:13.260)

I don't think robots are gonna come after me

Lex Fridman (1:21:15.460)

with a kitchen knife or a pellet gun right away.

Lex Fridman (1:21:18.460)

And I mean, if they were programmed in such a way,

Lex Fridman (1:21:21.420)

but I used to joke with Atlas that all I had to do

Russ Tedrake (1:21:25.940)

was run for five minutes and its battery would run out.

Lex Fridman (1:21:28.340)

But actually they've got to be careful

Lex Fridman (1:21:30.620)

and actually they've got a very big battery

Lex Fridman (1:21:32.460)

in there by the end.

Lex Fridman (1:21:33.300)

So it was over an hour.

Lex Fridman (1:21:37.220)

I think the fear is a bit cultural though.

Russ Tedrake (1:21:39.420)

Cause I mean, you notice that, like, I think in my age,

Lex Fridman (1:21:45.140)

in the US, we grew up watching Terminator, right?

Russ Tedrake (1:21:48.260)

If I had grown up at the same time in Japan,

Lex Fridman (1:21:50.500)

I probably would have been watching Astro Boy.

Lex Fridman (1:21:52.740)

And there's a very different reaction to robots

Lex Fridman (1:21:55.860)

in different countries, right?

Lex Fridman (1:21:57.460)

So I don't know if it's a human innate fear of metal marvels

Lex Fridman (1:22:02.620)

or if it's something that we've done to ourselves

Russ Tedrake (1:22:06.420)

with our sci fi.

Lex Fridman (1:22:09.860)

Yeah, the stories we tell ourselves through movies,

Russ Tedrake (1:22:12.580)

through just through popular media.

Lex Fridman (1:22:16.780)

But if I were to tell, you know, if you were my therapist

Lex Fridman (1:22:21.100)

and I said, I'm really terrified that we're going

Lex Fridman (1:22:24.900)

to have these robots very soon that will hurt us.

Lex Fridman (1:22:30.900)

Like, how do you approach making me feel better?

Lex Fridman (1:22:36.620)

Like, why shouldn't people be afraid?

Russ Tedrake (1:22:39.580)

There's a, I think there's a video

Lex Fridman (1:22:41.380)

that went viral recently.

Russ Tedrake (1:22:44.500)

Everything, everything was spot in Boston,

Lex Fridman (1:22:46.900)

which goes viral in general.

Lex Fridman (1:22:48.380)

But usually it's like really cool stuff.

Lex Fridman (1:22:50.060)

Like they're doing flips and stuff

Russ Tedrake (1:22:51.420)

or like sad stuff, the Atlas being hit with a broomstick

Lex Fridman (1:22:56.140)

or something like that.

Lex Fridman (1:22:57.300)

But there's a video where I think one of the new productions

Lex Fridman (1:23:02.420)

bought robots, which are awesome.

Russ Tedrake (1:23:04.620)

It was like patrolling somewhere in like in some country.

Lex Fridman (1:23:08.540)

And like people immediately were like saying like,

Russ Tedrake (1:23:11.920)

this is like the dystopian future,

Lex Fridman (1:23:14.580)

like the surveillance state.

Russ Tedrake (1:23:16.380)

For some reason, like you can just have a camera,

Lex Fridman (1:23:18.940)

like something about spot being able to walk on four feet

Russ Tedrake (1:23:23.420)

with like really terrified people.

Lex Fridman (1:23:25.940)

So like, what do you say to those people?

Russ Tedrake (1:23:31.060)

I think there is a legitimate fear there

Lex Fridman (1:23:33.820)

because so much of our future is uncertain.

Lex Fridman (1:23:37.840)

But at the same time, technically speaking,

Lex Fridman (1:23:40.140)

it seems like we're not there yet.

Lex Fridman (1:23:41.920)

So what do you say?

Lex Fridman (1:23:42.820)

I mean, I think technology is complicated.

Russ Tedrake (1:23:48.580)

It can be used in many ways.

Lex Fridman (1:23:49.940)

I think there are purely software attacks

Russ Tedrake (1:23:56.360)

that somebody could use to do great damage.

Lex Fridman (1:23:59.000)

Maybe they have already, you know,

Russ Tedrake (1:24:01.480)

I think wheeled robots could be used in bad ways too.

Lex Fridman (1:24:08.340)

Drones.

Russ Tedrake (1:24:09.180)

Drones, right, I don't think that, let's see.

Lex Fridman (1:24:16.340)

I don't want to be building technology

Russ Tedrake (1:24:19.920)

just because I'm compelled to build technology

Lex Fridman (1:24:21.860)

and I don't think about it.

Lex Fridman (1:24:23.580)

But I would consider myself a technological optimist,

Lex Fridman (1:24:27.740)

I guess, in the sense that I think we should continue

Russ Tedrake (1:24:32.220)

to create and evolve and our world will change.

Lex Fridman (1:24:37.220)

And if we will introduce new challenges,

Russ Tedrake (1:24:40.780)

we'll screw something up maybe,

Lex Fridman (1:24:42.900)

but I think also we'll invent ourselves

Russ Tedrake (1:24:46.220)

out of those challenges and life will go on.

Lex Fridman (1:24:49.380)

So it's interesting because you didn't mention

Russ Tedrake (1:24:51.580)

like this is technically too hard.

Lex Fridman (1:24:54.540)

I don't think robots are, I think people attribute

Russ Tedrake (1:24:57.380)

a robot that looks like an animal

Lex Fridman (1:24:59.140)

as maybe having a level of self awareness

Russ Tedrake (1:25:02.140)

or consciousness or something that they don't have yet.

Lex Fridman (1:25:05.460)

Right, so it's not, I think our ability

Russ Tedrake (1:25:09.380)

to anthropomorphize those robots is probably,

Lex Fridman (1:25:13.700)

we're assuming that they have a level of intelligence

Russ Tedrake (1:25:16.540)

that they don't yet have.

Lex Fridman (1:25:17.940)

And that might be part of the fear.

Lex Fridman (1:25:20.060)

So in that sense, it's too hard.

Lex Fridman (1:25:22.260)

But, you know, there are many scary things in the world.

Russ Tedrake (1:25:25.540)

Right, so I think we're right to ask those questions.

Lex Fridman (1:25:29.860)

We're right to think about the implications of our work.

Russ Tedrake (1:25:33.600)

Right, in the short term as we're working on it for sure,

Lex Fridman (1:25:39.720)

is there something long term that scares you

Lex Fridman (1:25:43.840)

about our future with AI and robots?

Lex Fridman (1:25:47.680)

A lot of folks from Elon Musk to Sam Harris

Russ Tedrake (1:25:52.400)

to a lot of folks talk about the existential threats

Lex Fridman (1:25:56.860)

about artificial intelligence.

Russ Tedrake (1:25:58.880)

Oftentimes, robots kind of inspire that the most

Lex Fridman (1:26:03.680)

because of the anthropomorphism.

Lex Fridman (1:26:05.840)

Do you have any fears?

Lex Fridman (1:26:07.400)

It's an important question.

Russ Tedrake (1:26:12.120)

I actually, I think I like Rod Brooks answer

Lex Fridman (1:26:14.920)

maybe the best on this, I think.

Lex Fridman (1:26:17.080)

And it's not the only answer he's given over the years,

Lex Fridman (1:26:19.320)

but maybe one of my favorites is he says,

Russ Tedrake (1:26:24.360)

it's not gonna be, he's got a book,

Lex Fridman (1:26:25.920)

Flesh and Machines, I believe, it's not gonna be

Russ Tedrake (1:26:29.960)

the robots versus the people,

Lex Fridman (1:26:31.880)

we're all gonna be robot people.

Russ Tedrake (1:26:34.240)

Because, you know, we already have smartphones,

Lex Fridman (1:26:38.000)

some of us have serious technology implanted

Russ Tedrake (1:26:41.120)

in our bodies already, whether we have a hearing aid

Lex Fridman (1:26:43.780)

or a pacemaker or anything like this,

Russ Tedrake (1:26:47.800)

people with amputations might have prosthetics.

Lex Fridman (1:26:50.880)

And that's a trend I think that is likely to continue.

Russ Tedrake (1:26:57.340)

I mean, this is now wild speculation.

Lex Fridman (1:27:01.420)

But I mean, when do we get to cognitive implants

Lex Fridman (1:27:05.500)

and the like, and.

Lex Fridman (1:27:06.620)

Yeah, with neural link, brain computer interfaces,

Russ Tedrake (1:27:09.500)

that's interesting.

Lex Fridman (1:27:10.340)

So there's a dance between humans and robots

Russ Tedrake (1:27:12.620)

that's going to be, it's going to be impossible

Lex Fridman (1:27:17.220)

to be scared of the other out there, the robot,

Russ Tedrake (1:27:23.380)

because the robot will be part of us, essentially.

Lex Fridman (1:27:26.060)

It'd be so intricately sort of part of our society that.

Russ Tedrake (1:27:30.180)

Yeah, and it might not even be implanted part of us,

Lex Fridman (1:27:33.060)

but just, it's so much a part of our, yeah, our society.

Lex Fridman (1:27:37.220)

So in that sense, the smartphone is already the robot

Lex Fridman (1:27:39.380)

we should be afraid of, yeah.

Russ Tedrake (1:27:41.660)

I mean, yeah, and all the usual fears arise

Lex Fridman (1:27:45.460)

of the misinformation, the manipulation,

Russ Tedrake (1:27:51.860)

all those kinds of things that,

Lex Fridman (1:27:56.180)

the problems are all the same.

Russ Tedrake (1:27:57.860)

They're human problems, essentially, it feels like.

Lex Fridman (1:28:00.700)

Yeah, I mean, I think the way we interact

Russ Tedrake (1:28:03.420)

with each other online is changing the value we put on,

Lex Fridman (1:28:07.420)

you know, personal interaction.

Lex Fridman (1:28:08.940)

And that's a crazy big change that's going to happen

Lex Fridman (1:28:11.260)

and rip through our, has already been ripping

Lex Fridman (1:28:13.080)

through our society, right?

Lex Fridman (1:28:14.200)

And that has implications that are massive.

Russ Tedrake (1:28:18.060)

I don't know if they should be scared of it

Lex Fridman (1:28:19.300)

or go with the flow, but I don't see, you know,

Russ Tedrake (1:28:24.700)

some battle lines between humans and robots

Lex Fridman (1:28:26.500)

being the first thing to worry about.

Russ Tedrake (1:28:29.580)

I mean, I do want to just, as a kind of comment,

Lex Fridman (1:28:33.340)

maybe you can comment about your just feelings

Russ Tedrake (1:28:35.460)

about Boston Dynamics in general, but you know,

Lex Fridman (1:28:38.660)

I love science, I love engineering,

Russ Tedrake (1:28:40.300)

I think there's so many beautiful ideas in it.

Lex Fridman (1:28:42.540)

And when I look at Boston Dynamics

Russ Tedrake (1:28:45.300)

or legged robots in general,

Lex Fridman (1:28:47.620)

I think they inspire people, curiosity and feelings

Russ Tedrake (1:28:54.620)

in general, excitement about engineering

Lex Fridman (1:28:57.460)

more than almost anything else in popular culture.

Lex Fridman (1:29:00.620)

And I think that's such an exciting,

Lex Fridman (1:29:03.660)

like responsibility and possibility for robotics.

Lex Fridman (1:29:06.820)

And Boston Dynamics is riding that wave pretty damn well.

Lex Fridman (1:29:10.460)

Like they found it, they've discovered that hunger

Lex Fridman (1:29:13.980)

and curiosity in the people and they're doing magic with it.

Lex Fridman (1:29:17.540)

I don't care if the, I mean, I guess is that their company,

Lex Fridman (1:29:19.820)

they have to make money, right?

Lex Fridman (1:29:21.340)

But they're already doing incredible work

Lex Fridman (1:29:24.300)

and inspiring the world about technology.

Lex Fridman (1:29:26.940)

I mean, do you have thoughts about Boston Dynamics

Lex Fridman (1:29:30.700)

and maybe others, your own work in robotics

Lex Fridman (1:29:34.620)

and inspiring the world in that way?

Russ Tedrake (1:29:36.600)

I completely agree, I think Boston Dynamics

Lex Fridman (1:29:40.240)

is absolutely awesome.

Russ Tedrake (1:29:42.640)

I think I show my kids those videos, you know,

Lex Fridman (1:29:46.160)

and the best thing that happens is sometimes

Lex Fridman (1:29:48.640)

they've already seen them, you know, right?

Lex Fridman (1:29:50.740)

I think, I just think it's a pinnacle of success

Russ Tedrake (1:29:55.360)

in robotics that is just one of the best things

Lex Fridman (1:29:58.760)

that's happened, absolutely completely agree.

Russ Tedrake (1:30:01.660)

One of the heartbreaking things to me is how many

Lex Fridman (1:30:06.220)

robotics companies fail, how hard it is to make money

Russ Tedrake (1:30:11.300)

with a robotics company.

Lex Fridman (1:30:13.100)

Like iRobot like went through hell just to arrive

Russ Tedrake (1:30:17.220)

at a Roomba to figure out one product.

Lex Fridman (1:30:19.740)

And then there's so many home robotics companies

Russ Tedrake (1:30:23.900)

like Jibo and Anki, Anki, the cutest toy that's a great robot

Lex Fridman (1:30:32.720)

I thought went down, I'm forgetting a bunch of them,

Lex Fridman (1:30:36.320)

but a bunch of robotics companies fail,

Lex Fridman (1:30:37.980)

Rod's company, Rethink Robotics.

Russ Tedrake (1:30:42.340)

Like, do you have anything hopeful to say

Lex Fridman (1:30:47.260)

about the possibility of making money with robots?

Russ Tedrake (1:30:50.340)

Oh, I think you can't just look at the failures.

Lex Fridman (1:30:54.220)

I mean, Boston Dynamics is a success.

Russ Tedrake (1:30:55.940)

There's lots of companies that are still doing amazingly

Lex Fridman (1:30:58.500)

good work in robotics.

Lex Fridman (1:31:01.140)

I mean, this is the capitalist ecology or something, right?

Lex Fridman (1:31:05.360)

I think you have many companies, you have many startups

Lex Fridman (1:31:07.700)

and they push each other forward and many of them fail

Lex Fridman (1:31:11.380)

and some of them get through and that's sort of

Russ Tedrake (1:31:13.820)

the natural way of those things.

Lex Fridman (1:31:17.040)

I don't know that is robotics really that much worse.

Russ Tedrake (1:31:20.460)

I feel the pain that you feel too.

Lex Fridman (1:31:22.300)

Every time I read one of these, sometimes it's friends

Lex Fridman (1:31:26.480)

and I definitely wish it went better or went differently.

Lex Fridman (1:31:33.580)

But I think it's healthy and good to have bursts of ideas,

Russ Tedrake (1:31:38.340)

bursts of activities, ideas, if they are really aggressive,

Lex Fridman (1:31:41.880)

they should fail sometimes.

Lex Fridman (1:31:45.180)

Certainly that's the research mantra, right?

Lex Fridman (1:31:46.940)

If you're succeeding at every problem you attempt,

Russ Tedrake (1:31:50.780)

then you're not choosing aggressively enough.

Lex Fridman (1:31:53.380)

Is it exciting to you, the new spot?

Russ Tedrake (1:31:55.980)

Oh, it's so good.

Lex Fridman (1:31:57.620)

When are you getting them as a pet or it?

Russ Tedrake (1:32:00.140)

Yeah, I mean, I have to dig up 75K right now.

Lex Fridman (1:32:03.220)

I mean, it's so cool that there's a price tag,

Russ Tedrake (1:32:05.740)

you can go and then actually buy it.

Lex Fridman (1:32:08.620)

I have a Skydio R1, love it.

Lex Fridman (1:32:11.500)

So no, I would absolutely be a customer.

Lex Fridman (1:32:18.580)

I wonder what your kids would think about it.

Russ Tedrake (1:32:20.060)

I actually, Zach from Boston Dynamics would let my kid drive

Lex Fridman (1:32:25.660)

in one of their demos one time.

Lex Fridman (1:32:27.140)

And that was just so good, so good.

Lex Fridman (1:32:31.100)

And again, I'll forever be grateful for that.

Lex Fridman (1:32:34.220)

And there's something magical about the anthropomorphization

Lex Fridman (1:32:37.260)

of that arm, it adds another level of human connection.

Russ Tedrake (1:32:42.580)

I'm not sure we understand from a control aspect,

Lex Fridman (1:32:47.480)

the value of anthropomorphization.

Russ Tedrake (1:32:51.540)

I think that's an understudied

Lex Fridman (1:32:53.980)

and under understood engineering problem.

Russ Tedrake (1:32:57.060)

There's been a, like psychologists have been studying it.

Lex Fridman (1:33:00.160)

I think it's part like manipulating our mind

Russ Tedrake (1:33:02.860)

to believe things is a valuable engineering.

Lex Fridman (1:33:06.740)

Like this is another degree of freedom

Russ Tedrake (1:33:08.820)

that can be controlled.

Lex Fridman (1:33:09.820)

I like that, yeah, I think that's right.

Russ Tedrake (1:33:11.380)

I think there's something that humans seem to do

Lex Fridman (1:33:16.020)

or maybe my dangerous introspection is,

Russ Tedrake (1:33:20.340)

I think we are able to make very simple models

Lex Fridman (1:33:23.820)

that assume a lot about the world very quickly.

Lex Fridman (1:33:27.780)

And then it takes us a lot more time, like you're wrestling.

Lex Fridman (1:33:31.220)

You probably thought you knew what you were doing

Russ Tedrake (1:33:33.080)

with wrestling and you were fairly functional

Lex Fridman (1:33:35.340)

as a complete wrestler.

Lex Fridman (1:33:36.900)

And then you slowly got more expertise.

Lex Fridman (1:33:39.340)

So maybe it's natural that our first level of defense

Russ Tedrake (1:33:45.740)

against seeing a new robot is to think of it

Lex Fridman (1:33:48.040)

in our existing models of how humans and animals behave.

Lex Fridman (1:33:52.420)

And it's just, as you spend more time with it,

Lex Fridman (1:33:55.060)

then you'll develop more sophisticated models

Russ Tedrake (1:33:56.980)

that will appreciate the differences.

Lex Fridman (1:34:00.340)

Exactly.

Lex Fridman (1:34:01.620)

Can you say what does it take to control a robot?

Lex Fridman (1:34:05.700)

Like what is the control problem of a robot?

Lex Fridman (1:34:08.580)

And in general, what is a robot in your view?

Lex Fridman (1:34:10.980)

Like how do you think of this system?

Lex Fridman (1:34:15.020)

What is a robot?

Lex Fridman (1:34:16.020)

What is a robot?

Russ Tedrake (1:34:17.580)

I think robotics.

Lex Fridman (1:34:18.400)

I told you ridiculous questions.

Russ Tedrake (1:34:20.020)

No, no, it's good.

Lex Fridman (1:34:21.500)

I mean, there's standard definitions

Russ Tedrake (1:34:22.980)

of combining computation with some ability

Lex Fridman (1:34:27.460)

to do mechanical work.

Russ Tedrake (1:34:29.060)

I think that gets us pretty close.

Lex Fridman (1:34:30.980)

But I think robotics has this problem

Russ Tedrake (1:34:34.180)

that once things really work,

Lex Fridman (1:34:37.200)

we don't call them robots anymore.

Russ Tedrake (1:34:38.920)

Like my dishwasher at home is pretty sophisticated,

Lex Fridman (1:34:44.100)

beautiful mechanisms.

Russ Tedrake (1:34:45.600)

There's actually a pretty good computer,

Lex Fridman (1:34:46.940)

probably a couple of chips in there doing amazing things.

Russ Tedrake (1:34:49.580)

We don't think of that as a robot anymore,

Lex Fridman (1:34:51.620)

which isn't fair.

Russ Tedrake (1:34:52.460)

Because then what roughly it means

Lex Fridman (1:34:53.940)

that robotics always has to solve the next problem

Lex Fridman (1:34:58.340)

and doesn't get to celebrate its past successes.

Lex Fridman (1:35:00.580)

I mean, even factory room floor robots

Russ Tedrake (1:35:05.660)

are super successful.

Lex Fridman (1:35:06.860)

They're amazing.

Lex Fridman (1:35:08.260)

But that's not the ones,

Lex Fridman (1:35:09.500)

I mean, people think of them as robots,

Lex Fridman (1:35:10.880)

but they don't,

Lex Fridman (1:35:11.720)

if you ask what are the successes of robotics,

Russ Tedrake (1:35:14.500)

somehow it doesn't come to your mind immediately.

Lex Fridman (1:35:17.860)

So the definition of robot is a system

Russ Tedrake (1:35:20.560)

with some level of automation that fails frequently.

Lex Fridman (1:35:23.500)

Something like, it's the computation plus mechanical work

Lex Fridman (1:35:28.420)

and an unsolved problem.

Lex Fridman (1:35:30.540)

It's an unsolved problem, yeah.

Lex Fridman (1:35:32.300)

So from a perspective of control and mechanics,

Lex Fridman (1:35:37.020)

dynamics, what is a robot?

Lex Fridman (1:35:40.700)

So there are many different types of robots.

Lex Fridman (1:35:42.380)

The control that you need for a Jibo robot,

Russ Tedrake (1:35:47.620)

you know, some robot that's sitting on your countertop

Lex Fridman (1:35:50.620)

and interacting with you, but not touching you,

Russ Tedrake (1:35:53.580)

for instance, is very different than what you need

Lex Fridman (1:35:55.820)

for an autonomous car or an autonomous drone.

Russ Tedrake (1:35:59.460)

It's very different than what you need for a robot

Lex Fridman (1:36:01.020)

that's gonna walk or pick things up with its hands, right?

Russ Tedrake (1:36:04.740)

My passion has always been for the places

Lex Fridman (1:36:09.140)

where you're interacting more,

Russ Tedrake (1:36:10.540)

you're doing more dynamic interactions with the world.

Lex Fridman (1:36:13.700)

So walking, now manipulation.

Lex Fridman (1:36:18.740)

And the control problems there are beautiful.

Lex Fridman (1:36:21.700)

I think contact is one thing that differentiates them

Russ Tedrake (1:36:25.940)

from many of the control problems we've solved classically,

Lex Fridman (1:36:29.240)

right, like modern control grew up stabilizing fighter jets

Russ Tedrake (1:36:32.780)

that were passively unstable,

Lex Fridman (1:36:34.060)

and there's like amazing success stories from control

Russ Tedrake (1:36:37.020)

all over the place.

Lex Fridman (1:36:39.140)

Power grid, I mean, there's all kinds of,

Russ Tedrake (1:36:41.340)

it's everywhere that we don't even realize,

Lex Fridman (1:36:44.640)

just like AI is now.

Lex Fridman (1:36:47.540)

So you mentioned contact, like what's contact?

Lex Fridman (1:36:51.500)

So an airplane is an extremely complex system

Russ Tedrake (1:36:54.980)

or a spacecraft landing or whatever,

Lex Fridman (1:36:57.380)

but at least it has the luxury

Russ Tedrake (1:36:59.340)

of things change relatively continuously.

Lex Fridman (1:37:03.640)

That's an oversimplification.

Lex Fridman (1:37:04.940)

But if I make a small change

Lex Fridman (1:37:07.060)

in the command I send to my actuator,

Russ Tedrake (1:37:10.140)

then the path that the robot will take

Lex Fridman (1:37:12.680)

tends to change only by a small amount.

Lex Fridman (1:37:16.820)

And there's a feedback mechanism here.

Lex Fridman (1:37:18.860)

That's what we're talking about.

Lex Fridman (1:37:19.700)

And there's a feedback mechanism.

Lex Fridman (1:37:20.980)

And thinking about this as locally,

Russ Tedrake (1:37:23.780)

like a linear system, for instance,

Lex Fridman (1:37:25.820)

I can use more linear algebra tools

Russ Tedrake (1:37:29.220)

to study systems like that,

Lex Fridman (1:37:31.340)

generalizations of linear algebra to these smooth systems.

Lex Fridman (1:37:36.400)

What is contact?

Lex Fridman (1:37:37.380)

The robot has something very discontinuous

Russ Tedrake (1:37:41.540)

that happens when it makes or breaks,

Lex Fridman (1:37:43.620)

when it starts touching the world.

Lex Fridman (1:37:45.420)

And even the way it touches or the order of contacts

Lex Fridman (1:37:48.080)

can change the outcome in potentially unpredictable ways.

Russ Tedrake (1:37:53.080)

Not unpredictable, but complex ways.

Lex Fridman (1:37:56.880)

I do think there's a little bit of,

Russ Tedrake (1:38:01.440)

a lot of people will say that contact is hard in robotics,

Lex Fridman (1:38:04.580)

even to simulate.

Lex Fridman (1:38:06.360)

And I think there's a little bit of a,

Lex Fridman (1:38:08.720)

there's truth to that,

Lex Fridman (1:38:09.640)

but maybe a misunderstanding around that.

Lex Fridman (1:38:13.560)

So what is limiting is that when we think about our robots

Lex Fridman (1:38:19.600)

and we write our simulators,

Lex Fridman (1:38:21.400)

we often make an assumption that objects are rigid.

Lex Fridman (1:38:26.000)

And when it comes down, that their mass moves all,

Lex Fridman (1:38:30.720)

stays in a constant position relative to each other itself.

Lex Fridman (1:38:37.080)

And that leads to some paradoxes

Lex Fridman (1:38:39.360)

when you go to try to talk about

Russ Tedrake (1:38:40.560)

rigid body mechanics and contact.

Lex Fridman (1:38:43.200)

And so for instance, if I have a three legged stool

Russ Tedrake (1:38:48.200)

with just imagine it comes to a point at the leg.

Lex Fridman (1:38:51.840)

So it's only touching the world at a point.

Russ Tedrake (1:38:54.400)

If I draw my physics,

Lex Fridman (1:38:56.920)

my high school physics diagram of the system,

Russ Tedrake (1:39:00.280)

then there's a couple of things

Lex Fridman (1:39:01.600)

that I'm given by elementary physics.

Russ Tedrake (1:39:03.800)

I know if the system, if the table is at rest,

Lex Fridman (1:39:06.320)

if it's not moving, zero velocities,

Russ Tedrake (1:39:09.520)

that means that the normal force,

Lex Fridman (1:39:11.120)

all the forces are in balance.

Lex Fridman (1:39:13.280)

So the force of gravity is being countered

Lex Fridman (1:39:16.400)

by the forces that the ground is pushing on my table legs.

Russ Tedrake (1:39:21.240)

I also know since it's not rotating

Lex Fridman (1:39:23.880)

that the moments have to balance.

Lex Fridman (1:39:25.800)

And since it's a three dimensional table,

Lex Fridman (1:39:29.560)

it could fall in any direction.

Russ Tedrake (1:39:31.120)

It actually tells me uniquely

Lex Fridman (1:39:33.040)

what those three normal forces have to be.

Russ Tedrake (1:39:37.080)

If I have four legs on my table,

Lex Fridman (1:39:39.600)

four legged table and they were perfectly machined

Russ Tedrake (1:39:43.280)

to be exactly the right same height

Lex Fridman (1:39:45.360)

and they're set down and the table's not moving,

Russ Tedrake (1:39:48.040)

then the basic conservation laws don't tell me,

Lex Fridman (1:39:51.960)

there are many solutions for the forces

Russ Tedrake (1:39:54.040)

that the ground could be putting on my legs

Lex Fridman (1:39:56.600)

that would still result in the table not moving.

Russ Tedrake (1:40:00.200)

Now, the reason that seems fine, I could just pick one.

Lex Fridman (1:40:03.920)

But it gets funny now because if you think about friction,

Lex Fridman (1:40:07.840)

what we think about with friction is our standard model

Lex Fridman (1:40:11.000)

says the amount of force that the table will push back

Russ Tedrake (1:40:15.880)

if I were to now try to push my table sideways,

Lex Fridman (1:40:18.000)

I guess I have a table here,

Russ Tedrake (1:40:20.880)

is proportional to the normal force.

Lex Fridman (1:40:24.040)

So if I'm barely touching and I push, I'll slide,

Lex Fridman (1:40:27.200)

but if I'm pushing more and I push, I'll slide less.

Lex Fridman (1:40:30.440)

It's called coulomb friction is our standard model.

Russ Tedrake (1:40:33.720)

Now, if you don't know what the normal force is

Lex Fridman (1:40:35.520)

on the four legs and you push the table,

Russ Tedrake (1:40:38.840)

then you don't know what the friction forces are gonna be.

Lex Fridman (1:40:43.440)

And so you can't actually tell,

Russ Tedrake (1:40:45.560)

the laws just aren't explicit yet

Lex Fridman (1:40:47.960)

about which way the table's gonna go.

Russ Tedrake (1:40:49.680)

It could veer off to the left,

Lex Fridman (1:40:51.360)

it could veer off to the right, it could go straight.

Lex Fridman (1:40:54.720)

So the rigid body assumption of contact

Lex Fridman (1:40:58.440)

leaves us with some paradoxes,

Russ Tedrake (1:40:59.840)

which are annoying for writing simulators

Lex Fridman (1:41:02.840)

and for writing controllers.

Russ Tedrake (1:41:04.240)

We still do that sometimes because soft contact

Lex Fridman (1:41:07.720)

is potentially harder numerically or whatever,

Lex Fridman (1:41:11.400)

and the best simulators do both

Lex Fridman (1:41:12.920)

or do some combination of the two.

Lex Fridman (1:41:15.240)

But anyways, because of these kinds of paradoxes,

Lex Fridman (1:41:17.360)

there's all kinds of paradoxes in contact,

Russ Tedrake (1:41:20.720)

mostly due to these rigid body assumptions.

Lex Fridman (1:41:23.560)

It becomes very hard to write the same kind of control laws

Russ Tedrake (1:41:27.880)

that we've been able to be successful with

Lex Fridman (1:41:29.600)

for fighter jets.

Russ Tedrake (1:41:32.000)

Like fighter jets, we haven't been as successful

Lex Fridman (1:41:34.560)

writing those controllers for manipulation.

Lex Fridman (1:41:37.440)

And so you don't know what's going to happen

Lex Fridman (1:41:39.160)

at the point of contact, at the moment of contact.

Russ Tedrake (1:41:41.480)

There are situations absolutely

Lex Fridman (1:41:42.880)

where our laws don't tell us.

Lex Fridman (1:41:45.760)

So the standard approach, that's okay.

Lex Fridman (1:41:47.440)

I mean, instead of having a differential equation,

Russ Tedrake (1:41:51.160)

you end up with a differential inclusion, it's called.

Lex Fridman (1:41:53.640)

It's a set valued equation.

Russ Tedrake (1:41:56.080)

It says that I'm in this configuration,

Lex Fridman (1:41:58.320)

I have these forces applied on me.

Lex Fridman (1:42:00.000)

And there's a set of things that could happen, right?

Lex Fridman (1:42:03.480)

And you can...

Lex Fridman (1:42:04.320)

And those aren't continuous, I mean, what...

Lex Fridman (1:42:07.480)

So when you're saying like non smooth,

Lex Fridman (1:42:10.360)

they're not only not smooth, but this is discontinuous?

Lex Fridman (1:42:14.520)

The non smooth comes in

Russ Tedrake (1:42:15.800)

when I make or break a new contact first,

Lex Fridman (1:42:18.760)

or when I transition from stick to slip.

Lex Fridman (1:42:21.200)

So you typically have static friction,

Lex Fridman (1:42:23.520)

and then you'll start sliding,

Lex Fridman (1:42:24.840)

and that'll be a discontinuous change in philosophy.

Lex Fridman (1:42:28.920)

In philosophy, for instance,

Russ Tedrake (1:42:31.360)

especially if you come to rest or...

Lex Fridman (1:42:33.360)

That's so fascinating.

Lex Fridman (1:42:34.480)

Okay, so what do you do?

Lex Fridman (1:42:37.720)

Sorry, I interrupted you.

Russ Tedrake (1:42:38.920)

It's fine.

Lex Fridman (1:42:41.600)

What's the hope under so much uncertainty

Lex Fridman (1:42:44.160)

about what's going to happen?

Lex Fridman (1:42:45.440)

What are you supposed to do?

Russ Tedrake (1:42:46.360)

I mean, control has an answer for this.

Lex Fridman (1:42:48.520)

Robust control is one approach,

Lex Fridman (1:42:50.240)

but roughly you can write controllers

Lex Fridman (1:42:52.640)

which try to still perform the right task

Russ Tedrake (1:42:55.920)

despite all the things that could possibly happen.

Lex Fridman (1:42:58.120)

The world might want the table to go this way and this way,

Lex Fridman (1:43:00.000)

but if I write a controller that pushes a little bit more

Lex Fridman (1:43:03.640)

and pushes a little bit,

Russ Tedrake (1:43:04.480)

I can certainly make the table go in the direction I want.

Lex Fridman (1:43:08.000)

It just puts a little bit more of a burden

Lex Fridman (1:43:10.000)

on the control system, right?

Lex Fridman (1:43:12.120)

And this discontinuities do change the control system

Russ Tedrake (1:43:15.440)

because the way we write it down right now,

Lex Fridman (1:43:21.200)

every different control configuration,

Russ Tedrake (1:43:24.320)

including sticking or sliding

Lex Fridman (1:43:26.200)

or parts of my body that are in contact or not,

Russ Tedrake (1:43:29.160)

looks like a different system.

Lex Fridman (1:43:30.840)

And I think of them,

Russ Tedrake (1:43:31.880)

I reason about them separately or differently

Lex Fridman (1:43:34.680)

and the combinatorics of that blow up, right?

Lex Fridman (1:43:38.000)

So I just don't have enough time to compute

Lex Fridman (1:43:41.440)

all the possible contact configurations of my humanoid.

Russ Tedrake (1:43:45.000)

Interestingly, I mean, I'm a humanoid.

Lex Fridman (1:43:49.000)

I have lots of degrees of freedom, lots of joints.

Russ Tedrake (1:43:52.400)

I've only been around for a handful of years.

Lex Fridman (1:43:54.960)

It's getting up there,

Lex Fridman (1:43:55.800)

but I haven't had time in my life

Lex Fridman (1:43:59.200)

to visit all of the states in my system,

Russ Tedrake (1:44:03.080)

certainly all the contact configurations.

Lex Fridman (1:44:05.240)

So if step one is to consider

Russ Tedrake (1:44:08.320)

every possible contact configuration that I'll ever be in,

Lex Fridman (1:44:12.160)

that's probably not a problem I need to solve, right?

Lex Fridman (1:44:17.040)

Just as a small tangent, what's a contact configuration?

Lex Fridman (1:44:20.560)

What like, just so we can enumerate

Lex Fridman (1:44:24.920)

what are we talking about?

Lex Fridman (1:44:26.280)

How many are there?

Russ Tedrake (1:44:27.600)

The simplest example maybe would be,

Lex Fridman (1:44:30.000)

imagine a robot with a flat foot.

Lex Fridman (1:44:32.720)

And we think about the phases of gait

Lex Fridman (1:44:35.440)

where the heel strikes and then the front toe strikes,

Lex Fridman (1:44:40.000)

and then you can heel up, toe off.

Lex Fridman (1:44:43.720)

Those are each different contact configurations.

Russ Tedrake (1:44:46.720)

I only had two different contacts,

Lex Fridman (1:44:48.320)

but I ended up with four different contact configurations.

Russ Tedrake (1:44:51.440)

Now, of course, my robot might actually have bumps on it

Lex Fridman (1:44:57.400)

or other things,

Lex Fridman (1:44:58.240)

so it could be much more subtle than that, right?

Lex Fridman (1:45:00.640)

But it's just even with one sort of box

Russ Tedrake (1:45:03.160)

interacting with the ground already in the plane

Lex Fridman (1:45:06.240)

has that many, right?

Lex Fridman (1:45:07.120)

And if I was just even a 3D foot,

Lex Fridman (1:45:09.440)

then it probably my left toe might touch

Russ Tedrake (1:45:11.240)

just before my right toe and things get subtle.

Lex Fridman (1:45:14.360)

Now, if I'm a dexterous hand

Lex Fridman (1:45:16.480)

and I go to talk about just grabbing a water bottle,

Lex Fridman (1:45:22.280)

if I have to enumerate every possible order

Russ Tedrake (1:45:26.720)

that my hand came into contact with the bottle,

Lex Fridman (1:45:31.000)

then I'm dead in the water.

Russ Tedrake (1:45:32.960)

Any approach that we were able to get away with that

Lex Fridman (1:45:35.400)

in walking because we mostly touched the ground

Russ Tedrake (1:45:38.480)

within a small number of points, for instance,

Lex Fridman (1:45:40.840)

and we haven't been able to get dexterous hands that way.

Lex Fridman (1:45:43.800)

So you've mentioned that people think

Lex Fridman (1:45:50.200)

that contact is really hard

Lex Fridman (1:45:52.520)

and that that's the reason that robotic manipulation

Lex Fridman (1:45:58.160)

is problem is really hard.

Lex Fridman (1:46:00.560)

Is there any flaws in that thinking?

Lex Fridman (1:46:06.560)

So I think simulating contact is one aspect.

Russ Tedrake (1:46:10.560)

I know people often say that we don't,

Lex Fridman (1:46:12.880)

that one of the reasons that we have a limit in robotics

Russ Tedrake (1:46:16.320)

is because we do not simulate contact accurately

Lex Fridman (1:46:19.040)

in our simulators.

Lex Fridman (1:46:20.840)

And I think that is the extent to which that's true

Lex Fridman (1:46:25.600)

is partly because our simulators,

Russ Tedrake (1:46:27.880)

we haven't got mature enough simulators.

Lex Fridman (1:46:31.240)

There are some things that are still hard, difficult,

Russ Tedrake (1:46:34.120)

that we should change,

Lex Fridman (1:46:38.200)

but we actually, we know what the governing equations are.

Russ Tedrake (1:46:41.520)

They have some foibles like this indeterminacy,

Lex Fridman (1:46:44.720)

but we should be able to simulate them accurately.

Russ Tedrake (1:46:48.600)

We have incredible open source community in robotics,

Lex Fridman (1:46:51.440)

but it actually just takes a professional engineering team

Russ Tedrake (1:46:54.360)

a lot of work to write a very good simulator like that.

Lex Fridman (1:46:59.080)

Now, where does, I believe you've written, Drake.

Russ Tedrake (1:47:03.280)

There's a team of people.

Lex Fridman (1:47:04.520)

I certainly spent a lot of hours on it myself.

Lex Fridman (1:47:07.320)

But what is Drake and what does it take to create

Lex Fridman (1:47:12.060)

a simulation environment for the kind of difficult control

Lex Fridman (1:47:18.200)

problems we're talking about?

Lex Fridman (1:47:20.740)

Right, so Drake is the simulator that I've been working on.

Russ Tedrake (1:47:24.640)

There are other good simulators out there.

Lex Fridman (1:47:26.780)

I don't like to think of Drake as just a simulator

Russ Tedrake (1:47:29.680)

because we write our controllers in Drake,

Lex Fridman (1:47:31.780)

we write our perception systems a little bit in Drake,

Lex Fridman (1:47:34.360)

but we write all of our low level control

Lex Fridman (1:47:37.040)

and even planning and optimization.

Lex Fridman (1:47:40.840)

So it has optimization capabilities as well?

Lex Fridman (1:47:42.480)

Absolutely, yeah.

Russ Tedrake (1:47:43.640)

I mean, Drake is three things roughly.

Lex Fridman (1:47:46.000)

It's an optimization library, which is sits on,

Russ Tedrake (1:47:49.800)

it provides a layer of abstraction in C++ and Python

Lex Fridman (1:47:54.240)

for commercial solvers.

Russ Tedrake (1:47:55.920)

You can write linear programs, quadratic programs,

Lex Fridman (1:48:00.760)

semi definite programs, sums of squares programs,

Russ Tedrake (1:48:03.340)

the ones we've used, mixed integer programs,

Lex Fridman (1:48:05.660)

and it will do the work to curate those

Lex Fridman (1:48:07.960)

and send them to whatever the right solver is for instance,

Lex Fridman (1:48:10.360)

and it provides a level of abstraction.

Russ Tedrake (1:48:13.720)

The second thing is a system modeling language,

Lex Fridman (1:48:18.360)

a bit like LabVIEW or Simulink,

Russ Tedrake (1:48:20.880)

where you can make block diagrams out of complex systems,

Lex Fridman (1:48:24.840)

or it's like ROS in that sense,

Russ Tedrake (1:48:26.640)

where you might have lots of ROS nodes

Lex Fridman (1:48:29.040)

that are each doing some part of your system,

Lex Fridman (1:48:31.960)

but to contrast it with ROS, we try to write,

Lex Fridman (1:48:36.560)

if you write a Drake system, then you have to,

Russ Tedrake (1:48:40.120)

it asks you to describe a little bit more about the system.

Lex Fridman (1:48:43.000)

If you have any state, for instance, in the system,

Russ Tedrake (1:48:46.240)

any variables that are gonna persist,

Lex Fridman (1:48:47.680)

you have to declare them.

Russ Tedrake (1:48:49.120)

Parameters can be declared and the like,

Lex Fridman (1:48:51.620)

but the advantage of doing that is that you can,

Russ Tedrake (1:48:54.160)

if you like, run things all on one process,

Lex Fridman (1:48:57.460)

but you can also do control design against it.

Russ Tedrake (1:49:00.200)

You can do, I mean, simple things like rewinding

Lex Fridman (1:49:03.120)

and playing back your simulations, for instance,

Russ Tedrake (1:49:07.960)

these things, you get some rewards

Lex Fridman (1:49:09.600)

for spending a little bit more upfront cost

Russ Tedrake (1:49:11.380)

in describing each system.

Lex Fridman (1:49:13.320)

And I was inspired to do that

Russ Tedrake (1:49:16.920)

because I think the complexity of Atlas, for instance,

Lex Fridman (1:49:21.260)

is just so great.

Lex Fridman (1:49:22.600)

And I think, although, I mean,

Lex Fridman (1:49:24.140)

ROS has been an incredible, absolutely huge fan

Russ Tedrake (1:49:27.520)

of what it's done for the robotics community,

Lex Fridman (1:49:30.720)

but the ability to rapidly put different pieces together

Lex Fridman (1:49:35.480)

and have a functioning thing is very good.

Lex Fridman (1:49:38.960)

But I do think that it's hard to think clearly

Russ Tedrake (1:49:42.880)

about a bag of disparate parts,

Lex Fridman (1:49:45.000)

Mr. Potato Head kind of software stack.

Lex Fridman (1:49:48.160)

And if you can ask a little bit more

Lex Fridman (1:49:53.060)

out of each of those parts,

Russ Tedrake (1:49:54.200)

then you can understand the way they work better.

Lex Fridman (1:49:56.120)

You can try to verify them and the like,

Russ Tedrake (1:50:00.160)

or you can do learning against them.

Lex Fridman (1:50:02.680)

And then one of those systems, the last thing,

Russ Tedrake (1:50:04.760)

I said the first two things that Drake is,

Lex Fridman (1:50:06.480)

but the last thing is that there is a set

Russ Tedrake (1:50:09.680)

of multi body equations, rigid body equations,

Lex Fridman (1:50:12.560)

that is trying to provide a system that simulates physics.

Lex Fridman (1:50:16.760)

And we also have renderers and other things,

Lex Fridman (1:50:20.060)

but I think the physics component of Drake is special

Russ Tedrake (1:50:23.300)

in the sense that we have done excessive amount

Lex Fridman (1:50:27.740)

of engineering to make sure

Russ Tedrake (1:50:29.840)

that we've written the equations correctly.

Lex Fridman (1:50:31.580)

Every possible tumbling satellite or spinning top

Russ Tedrake (1:50:34.160)

or anything that we could possibly write as a test is tested.

Lex Fridman (1:50:38.300)

We are making some, I think, fundamental improvements

Russ Tedrake (1:50:42.000)

on the way you simulate contact.

Lex Fridman (1:50:44.240)

Just what does it take to simulate contact?

Russ Tedrake (1:50:47.600)

I mean, it just seems,

Lex Fridman (1:50:50.920)

I mean, there's something just beautiful

Russ Tedrake (1:50:52.400)

to the way you were like explaining contact

Lex Fridman (1:50:55.240)

and you were like tapping your fingers

Russ Tedrake (1:50:56.720)

on the table while you're doing it, just.

Lex Fridman (1:51:00.720)

Easily, right?

Russ Tedrake (1:51:01.560)

Easily, just like, just not even like,

Lex Fridman (1:51:04.800)

it was like helping you think, I guess.

Lex Fridman (1:51:10.640)

So you have this like awesome demo

Lex Fridman (1:51:12.280)

of loading or unloading a dishwasher,

Russ Tedrake (1:51:16.720)

just picking up a plate,

Lex Fridman (1:51:18.840)

or grasping it like for the first time.

Russ Tedrake (1:51:26.120)

That's just seems like so difficult.

Lex Fridman (1:51:29.440)

What, how do you simulate any of that?

Lex Fridman (1:51:33.600)

So it was really interesting that what happened was

Lex Fridman (1:51:35.840)

that we started getting more professional

Russ Tedrake (1:51:39.200)

about our software development

Lex Fridman (1:51:40.520)

during the DARPA Robotics Challenge.

Russ Tedrake (1:51:43.360)

I learned the value of software engineering

Lex Fridman (1:51:46.040)

and how these, how to bridle complexity.

Russ Tedrake (1:51:48.640)

I guess that's what I want to somehow fight against

Lex Fridman (1:51:52.800)

and bring some of the clear thinking of controls

Russ Tedrake (1:51:54.760)

into these complex systems we're building for robots.

Lex Fridman (1:52:00.460)

Shortly after the DARPA Robotics Challenge,

Russ Tedrake (1:52:02.940)

Toyota opened a research institute,

Lex Fridman (1:52:04.600)

TRI, Toyota Research Institute.

Russ Tedrake (1:52:08.200)

They put one of their, there's three locations.

Lex Fridman (1:52:10.880)

One of them is just down the street from MIT.

Lex Fridman (1:52:13.040)

And I helped ramp that up right up

Lex Fridman (1:52:17.520)

as a part of my, the end of my sabbatical, I guess.

Lex Fridman (1:52:23.480)

So TRI has given me, the TRI robotics effort

Lex Fridman (1:52:29.480)

has made this investment in simulation in Drake.

Lex Fridman (1:52:32.640)

And Michael Sherman leads a team there

Lex Fridman (1:52:34.480)

of just absolutely top notch dynamics experts

Russ Tedrake (1:52:37.800)

that are trying to write those simulators

Lex Fridman (1:52:40.120)

that can pick up the dishes.

Lex Fridman (1:52:41.960)

And there's also a team working on manipulation there

Lex Fridman (1:52:44.780)

that is taking problems like loading the dishwasher.

Lex Fridman (1:52:48.980)

And we're using that to study these really hard corner cases

Lex Fridman (1:52:53.180)

kind of problems in manipulation.

Lex Fridman (1:52:55.280)

So for me, this, you know, simulating the dishes,

Lex Fridman (1:52:59.760)

we could actually write a controller.

Russ Tedrake (1:53:01.580)

If we just cared about picking up dishes in the sink once,

Lex Fridman (1:53:05.040)

we could write a controller

Russ Tedrake (1:53:05.880)

without any simulation whatsoever,

Lex Fridman (1:53:07.760)

and we could call it done.

Lex Fridman (1:53:10.040)

But we want to understand like,

Lex Fridman (1:53:12.140)

what is the path you take to actually get to a robot

Russ Tedrake (1:53:17.040)

that could perform that for any dish in anybody's kitchen

Lex Fridman (1:53:22.120)

with enough confidence

Lex Fridman (1:53:23.280)

that it could be a commercial product, right?

Lex Fridman (1:53:26.520)

And it has deep learning perception in the loop.

Russ Tedrake (1:53:29.360)

It has complex dynamics in the loop.

Lex Fridman (1:53:31.040)

It has controller, it has a planner.

Lex Fridman (1:53:33.240)

And how do you take all of that complexity

Lex Fridman (1:53:36.320)

and put it through this engineering discipline

Lex Fridman (1:53:39.020)

and verification and validation process

Lex Fridman (1:53:42.440)

to actually get enough confidence to deploy?

Russ Tedrake (1:53:46.440)

I mean, the DARPA challenge made me realize

Lex Fridman (1:53:49.840)

that that's not something you throw over the fence

Lex Fridman (1:53:52.000)

and hope that somebody will harden it for you,

Lex Fridman (1:53:54.080)

that there are really fundamental challenges

Russ Tedrake (1:53:57.380)

in closing that last gap.

Lex Fridman (1:53:59.840)

They're doing the validation and the testing.

Russ Tedrake (1:54:03.520)

I think it might even change the way we have to think about

Lex Fridman (1:54:06.780)

the way we write systems.

Lex Fridman (1:54:09.580)

What happens if you have the robot running lots of tests

Lex Fridman (1:54:15.560)

and it screws up, it breaks a dish, right?

Lex Fridman (1:54:19.040)

How do you capture that?

Lex Fridman (1:54:19.960)

I said, you can't run the same simulation

Russ Tedrake (1:54:23.580)

or the same experiment twice on a real robot.

Lex Fridman (1:54:27.920)

Do we have to be able to bring that one off failure

Russ Tedrake (1:54:31.520)

back into simulation

Lex Fridman (1:54:32.640)

in order to change our controllers, study it,

Lex Fridman (1:54:35.120)

make sure it won't happen again?

Lex Fridman (1:54:37.240)

Do we, is it enough to just try to add that

Russ Tedrake (1:54:40.600)

to our distribution and understand that on average,

Lex Fridman (1:54:43.800)

we're gonna cover that situation again?

Russ Tedrake (1:54:45.920)

There's like really subtle questions at the corner cases

Lex Fridman (1:54:49.960)

that I think we don't yet have satisfying answers for.

Lex Fridman (1:54:53.240)

Like how do you find the corner cases?

Lex Fridman (1:54:55.120)

That's one kind of, is there,

Lex Fridman (1:54:57.160)

do you think that's possible to create a systematized way

Lex Fridman (1:55:01.260)

of discovering corner cases efficiently?

Russ Tedrake (1:55:04.720)

Yes.

Lex Fridman (1:55:05.560)

In whatever the problem is?

Russ Tedrake (1:55:07.600)

Yes, I mean, I think we have to get better at that.

Lex Fridman (1:55:10.760)

I mean, control theory has for decades

Russ Tedrake (1:55:14.920)

talked about active experiment design.

Lex Fridman (1:55:17.840)

What's that?

Lex Fridman (1:55:19.560)

So people call it curiosity these days.

Lex Fridman (1:55:22.080)

It's roughly this idea of trying to exploration

Russ Tedrake (1:55:24.800)

or exploitation, but in the active experiment design

Lex Fridman (1:55:27.600)

is even, is more specific.

Russ Tedrake (1:55:29.640)

You could try to understand the uncertainty in your system,

Lex Fridman (1:55:34.120)

design the experiment that will provide

Russ Tedrake (1:55:36.480)

the maximum information to reduce that uncertainty.

Lex Fridman (1:55:40.120)

If there's a parameter you wanna learn about,

Lex Fridman (1:55:42.360)

what is the optimal trajectory I could execute

Lex Fridman (1:55:45.440)

to learn about that parameter, for instance.

Russ Tedrake (1:55:49.520)

Scaling that up to something that has a deep network

Lex Fridman (1:55:51.720)

in the loop and a planning in the loop is tough.

Russ Tedrake (1:55:55.660)

We've done some work on, you know,

Lex Fridman (1:55:58.200)

with Matt Okely and Aman Sinha,

Russ Tedrake (1:56:00.280)

we've worked on some falsification algorithms

Lex Fridman (1:56:03.600)

that are trying to do rare event simulation

Russ Tedrake (1:56:05.600)

that try to just hammer on your simulator.

Lex Fridman (1:56:08.120)

And if your simulator is good enough,

Russ Tedrake (1:56:10.000)

you can spend a lot of time,

Lex Fridman (1:56:13.840)

or you can write good algorithms

Russ Tedrake (1:56:15.840)

that try to spend most of their time in the corner cases.

Lex Fridman (1:56:19.920)

So you basically imagine you're building an autonomous car

Lex Fridman (1:56:25.880)

and you wanna put it in, I don't know,

Lex Fridman (1:56:27.360)

downtown New Delhi all the time, right?

Lex Fridman (1:56:29.400)

And accelerated testing.

Lex Fridman (1:56:31.640)

If you can write sampling strategies,

Russ Tedrake (1:56:33.340)

which figure out where your controller's

Lex Fridman (1:56:35.400)

performing badly in simulation

Lex Fridman (1:56:37.440)

and start generating lots of examples around that.

Lex Fridman (1:56:40.600)

You know, it's just the space of possible places

Russ Tedrake (1:56:44.060)

where that can be, where things can go wrong is very big.

Lex Fridman (1:56:48.040)

So it's hard to write those algorithms.

Russ Tedrake (1:56:49.800)

Yeah, rare event simulation

Lex Fridman (1:56:51.720)

is just a really compelling notion, if it's possible.

Russ Tedrake (1:56:55.760)

We joked and we call it the black swan generator.

Lex Fridman (1:56:58.600)

It's a black swan.

Russ Tedrake (1:57:00.080)

Because you don't just want the rare events,

Lex Fridman (1:57:01.680)

you want the ones that are highly impactful.

Russ Tedrake (1:57:04.020)

I mean, that's the most,

Lex Fridman (1:57:06.560)

those are the most sort of profound questions

Russ Tedrake (1:57:08.780)

we ask of our world.

Lex Fridman (1:57:10.120)

Like, what's the worst that can happen?

Lex Fridman (1:57:16.720)

But what we're really asking

Lex Fridman (1:57:18.080)

isn't some kind of like computer science,

Russ Tedrake (1:57:20.800)

worst case analysis.

Lex Fridman (1:57:22.560)

We're asking like, what are the millions of ways

Lex Fridman (1:57:25.600)

this can go wrong?

Lex Fridman (1:57:27.360)

And that's like our curiosity.

Lex Fridman (1:57:29.500)

And we humans, I think are pretty bad at,

Lex Fridman (1:57:34.900)

we just like run into it.

Lex Fridman (1:57:36.980)

And I think there's a distributed sense

Lex Fridman (1:57:38.580)

because there's now like 7.5 billion of us.

Lex Fridman (1:57:41.620)

And so there's a lot of them.

Lex Fridman (1:57:42.860)

And then a lot of them write blog posts

Russ Tedrake (1:57:45.060)

about the stupid thing they've done.

Lex Fridman (1:57:46.540)

So we learn in a distributed way.

Russ Tedrake (1:57:49.980)

There's some.

Lex Fridman (1:57:50.820)

I think that's gonna be important for robots too.

Russ Tedrake (1:57:53.380)

I mean, that's another massive theme

Lex Fridman (1:57:55.940)

at Toyota Research for Robotics

Russ Tedrake (1:57:58.800)

is this fleet learning concept

Lex Fridman (1:58:00.540)

is the idea that I, as a human,

Lex Fridman (1:58:04.780)

I don't have enough time to visit all of my states, right?

Lex Fridman (1:58:07.880)

There's just a, it's very hard for one robot

Russ Tedrake (1:58:10.140)

to experience all the things.

Lex Fridman (1:58:12.640)

But that's not actually the problem we have to solve, right?

Russ Tedrake (1:58:16.540)

We're gonna have fleets of robots

Lex Fridman (1:58:17.700)

that can have very similar appendages.

Lex Fridman (1:58:20.660)

And at some point, maybe collectively,

Lex Fridman (1:58:24.160)

they have enough data

Russ Tedrake (1:58:26.220)

that their computational processes

Lex Fridman (1:58:29.340)

should be set up differently than ours, right?

Russ Tedrake (1:58:31.860)

It's this vision of just,

Lex Fridman (1:58:34.180)

I mean, all these dishwasher unloading robots.

Russ Tedrake (1:58:38.880)

I mean, that robot dropping a plate

Lex Fridman (1:58:42.580)

and a human looking at the robot probably pissed off.

Russ Tedrake (1:58:46.860)

Yeah.

Lex Fridman (1:58:47.820)

But that's a special moment to record.

Russ Tedrake (1:58:51.220)

I think one thing in terms of fleet learning,

Lex Fridman (1:58:54.500)

and I've seen that because I've talked to a lot of folks,

Russ Tedrake (1:58:57.740)

just like Tesla users or Tesla drivers,

Lex Fridman (1:59:01.220)

they're another company

Russ Tedrake (1:59:02.980)

that's using this kind of fleet learning idea.

Lex Fridman (1:59:05.300)

One hopeful thing I have about humans

Russ Tedrake (1:59:08.220)

is they really enjoy when a system improves, learns.

Lex Fridman (1:59:13.260)

So they enjoy fleet learning.

Lex Fridman (1:59:14.680)

And the reason it's hopeful for me

Lex Fridman (1:59:17.260)

is they're willing to put up with something

Russ Tedrake (1:59:20.300)

that's kind of dumb right now.

Lex Fridman (1:59:22.660)

And they're like, if it's improving,

Russ Tedrake (1:59:25.540)

they almost like enjoy being part of the, like teaching it.

Lex Fridman (1:59:29.460)

Almost like if you have kids,

Lex Fridman (1:59:30.960)

like you're teaching them something, right?

Lex Fridman (1:59:33.540)

I think that's a beautiful thing

Russ Tedrake (1:59:35.140)

because that gives me hope

Lex Fridman (1:59:36.300)

that we can put dumb robots out there.

Russ Tedrake (1:59:40.100)

I mean, the problem on the Tesla side with cars,

Lex Fridman (1:59:43.340)

cars can kill you.

Russ Tedrake (1:59:45.320)

That makes the problem so much harder.

Lex Fridman (1:59:47.740)

Dishwasher unloading is a little safe.

Russ Tedrake (1:59:50.580)

That's why home robotics is really exciting.

Lex Fridman (1:59:54.220)

And just to clarify, I mean, for people who might not know,

Russ Tedrake (1:59:57.580)

I mean, TRI, Toyota Research Institute.

Lex Fridman (20:00.120)

as opposed to four?

Lex Fridman (20:01.120)

Is there something useful about it?

Lex Fridman (20:02.320)

I think the most, I mean, the reason I spent a long time

Russ Tedrake (20:05.540)

working on bipedal walking was because it was hard

Lex Fridman (20:09.000)

and it challenged control theory in ways

Russ Tedrake (20:12.480)

that I thought were important.

Lex Fridman (20:13.920)

I wouldn't have ever tried to convince you

Russ Tedrake (20:18.520)

that you should start a company around bipeds

Lex Fridman (20:22.440)

or something like this.

Russ Tedrake (20:24.240)

There are people that make pretty compelling arguments.

Lex Fridman (20:26.120)

I think the most compelling one is that the world

Russ Tedrake (20:28.920)

is built for the human form, and if you want a robot

Lex Fridman (20:32.320)

to work in the world we have today,

Russ Tedrake (20:34.800)

then having a human form is a pretty good way to go.

Lex Fridman (20:39.680)

There are places that a biped can go that would be hard

Russ Tedrake (20:42.560)

for other form factors to go, even natural places,

Lex Fridman (20:47.640)

but at some point in the long run,

Russ Tedrake (20:51.360)

we'll be building our environments for our robots, probably,

Lex Fridman (20:54.220)

and so maybe that argument falls aside.

Lex Fridman (20:56.480)

So you famously run barefoot.

Lex Fridman (21:00.640)

Do you still run barefoot?

Russ Tedrake (21:02.120)

I still run barefoot.

Lex Fridman (21:03.080)

That's so awesome.

Russ Tedrake (21:04.760)

Much to my wife's chagrin.

Lex Fridman (21:07.800)

Do you want to make an evolutionary argument

Lex Fridman (21:09.320)

for why running barefoot is advantageous?

Lex Fridman (21:12.680)

What have you learned about human and robot movement

Lex Fridman (21:17.560)

in general from running barefoot?

Lex Fridman (21:21.160)

Human or robot and or?

Lex Fridman (21:23.640)

Well, you know, it happened the other way, right?

Lex Fridman (21:25.640)

So I was studying walking robots,

Lex Fridman (21:27.680)

and there's a great conference called

Lex Fridman (21:31.760)

the Dynamic Walking Conference where it brings together

Russ Tedrake (21:35.320)

both the biomechanics community

Lex Fridman (21:36.980)

and the walking robots community.

Lex Fridman (21:39.880)

And so I had been going to this for years

Lex Fridman (21:41.660)

and hearing talks by people who study barefoot running

Lex Fridman (21:45.080)

and other, the mechanics of running.

Lex Fridman (21:48.080)

So I did eventually read Born to Run.

Lex Fridman (21:50.280)

Most people read Born to Run in the first, right?

Lex Fridman (21:54.080)

The other thing I had going for me is actually

Russ Tedrake (21:55.720)

that I wasn't a runner before,

Lex Fridman (21:58.800)

and I learned to run after I had learned

Russ Tedrake (22:01.560)

about barefoot running, or I mean,

Lex Fridman (22:03.640)

started running longer distances.

Lex Fridman (22:05.440)

So I didn't have to unlearn.

Lex Fridman (22:07.360)

And I'm definitely, I'm a big fan of it for me,

Lex Fridman (22:11.080)

but I'm not going to,

Lex Fridman (22:12.360)

I tend to not try to convince other people.

Russ Tedrake (22:14.600)

There's people who run beautifully with shoes on,

Lex Fridman (22:17.240)

and that's good.

Lex Fridman (22:20.040)

But here's why it makes sense for me.

Lex Fridman (22:24.040)

It's all about the longterm game, right?

Lex Fridman (22:26.360)

So I think it's just too easy to run 10 miles,

Lex Fridman (22:29.440)

feel pretty good, and then you get home at night

Lex Fridman (22:31.560)

and you realize my knees hurt.

Lex Fridman (22:33.840)

I did something wrong, right?

Russ Tedrake (22:37.880)

If you take your shoes off,

Lex Fridman (22:39.780)

then if you hit hard with your foot at all,

Russ Tedrake (22:44.080)

then it hurts.

Lex Fridman (22:45.720)

You don't like run 10 miles

Lex Fridman (22:47.560)

and then realize you've done some damage.

Lex Fridman (22:50.800)

You have immediate feedback telling you

Russ Tedrake (22:52.940)

that you've done something that's maybe suboptimal,

Lex Fridman (22:55.420)

and you change your gait.

Russ Tedrake (22:56.520)

I mean, it's even subconscious.

Lex Fridman (22:57.720)

If I, right now, having run many miles barefoot,

Russ Tedrake (23:00.640)

if I put a shoe on, my gait changes

Lex Fridman (23:03.160)

in a way that I think is not as good.

Lex Fridman (23:05.840)

So it makes me land softer.

Lex Fridman (23:09.520)

And I think my goals for running

Russ Tedrake (23:13.160)

are to do it for as long as I can into old age,

Lex Fridman (23:16.860)

not to win any races.

Lex Fridman (23:19.000)

And so for me, this is a way to protect myself.

Lex Fridman (23:23.420)

Yeah, I think, first of all,

Russ Tedrake (23:25.680)

I've tried running barefoot many years ago,

Lex Fridman (23:29.540)

probably the other way,

Russ Tedrake (23:30.480)

just reading Born to Run.

Lex Fridman (23:33.920)

But just to understand,

Russ Tedrake (23:36.440)

because I felt like I couldn't put in the miles

Lex Fridman (23:39.520)

that I wanted to.

Lex Fridman (23:40.840)

And it feels like running for me,

Lex Fridman (23:44.260)

and I think for a lot of people,

Russ Tedrake (23:46.280)

was one of those activities that we do often

Lex Fridman (23:48.880)

and we never really try to learn to do correctly.

Russ Tedrake (23:53.340)

Like, it's funny, there's so many activities

Lex Fridman (23:55.920)

we do every day, like brushing our teeth, right?

Russ Tedrake (24:00.280)

I think a lot of us, at least me,

Lex Fridman (24:02.360)

probably have never deeply studied

Lex Fridman (24:04.320)

how to properly brush my teeth, right?

Lex Fridman (24:07.040)

Or wash, as now with the pandemic,

Russ Tedrake (24:08.960)

or how to properly wash our hands.

Lex Fridman (24:10.640)

We do it every day, but we haven't really studied,

Lex Fridman (24:13.800)

like, am I doing this correctly?

Lex Fridman (24:15.200)

But running felt like one of those things,

Russ Tedrake (24:17.120)

it was absurd not to study how to do correctly,

Lex Fridman (24:20.220)

because it's the source of so much pain and suffering.

Russ Tedrake (24:23.320)

Like, I hate running, but I do it.

Lex Fridman (24:25.680)

I do it because I hate it, but I feel good afterwards.

Lex Fridman (24:28.940)

But I think it feels like you need

Lex Fridman (24:30.280)

to learn how to do it properly.

Lex Fridman (24:31.440)

So that's where barefoot running came in,

Lex Fridman (24:33.540)

and then I quickly realized that my gait

Russ Tedrake (24:35.760)

was completely wrong.

Lex Fridman (24:38.040)

I was taking huge steps,

Lex Fridman (24:41.440)

and landing hard on the heel, all those elements.

Lex Fridman (24:45.840)

And so, yeah, from that I actually learned

Russ Tedrake (24:47.600)

to take really small steps, look.

Lex Fridman (24:50.520)

I already forgot the number,

Lex Fridman (24:52.280)

but I feel like it was 180 a minute or something like that.

Lex Fridman (24:55.600)

And I remember I actually just took songs

Russ Tedrake (25:00.080)

that are 180 beats per minute,

Lex Fridman (25:03.360)

and then like tried to run at that beat,

Lex Fridman (25:06.520)

and just to teach myself.

Lex Fridman (25:07.660)

It took a long time, and I feel like after a while,

Russ Tedrake (25:11.120)

you learn to run, you adjust properly,

Lex Fridman (25:14.320)

without going all the way to barefoot.

Lex Fridman (25:15.960)

But I feel like barefoot is the legit way to do it.

Lex Fridman (25:19.440)

I mean, I think a lot of people

Russ Tedrake (25:21.640)

would be really curious about it.

Lex Fridman (25:23.360)

Can you, if they're interested in trying,

Lex Fridman (25:25.560)

what would you, how would you recommend

Lex Fridman (25:27.840)

they start, or try, or explore?

Russ Tedrake (25:30.740)

Slowly.

Lex Fridman (25:31.580)

That's the biggest thing people do,

Russ Tedrake (25:33.720)

is they are excellent runners,

Lex Fridman (25:35.920)

and they're used to running long distances,

Russ Tedrake (25:37.620)

or running fast, and they take their shoes off,

Lex Fridman (25:39.240)

and they hurt themselves instantly trying to do

Russ Tedrake (25:42.520)

something that they were used to doing.

Lex Fridman (25:44.280)

I think I lucked out in the sense

Russ Tedrake (25:46.000)

that I couldn't run very far when I first started trying.

Lex Fridman (25:50.200)

And I run with minimal shoes too.

Russ Tedrake (25:51.840)

I mean, I will bring along a pair of,

Lex Fridman (25:54.360)

actually, like aqua socks or something like this,

Russ Tedrake (25:56.320)

I can just slip on, or running sandals,

Lex Fridman (25:58.320)

I've tried all of them.

Russ Tedrake (26:00.360)

What's the difference between a minimal shoe

Lex Fridman (26:02.600)

and nothing at all?

Lex Fridman (26:03.760)

What's, like, feeling wise, what does it feel like?

Lex Fridman (26:07.020)

There is a, I mean, I notice my gait changing, right?

Russ Tedrake (26:10.000)

So, I mean, your foot has as many muscles

Lex Fridman (26:15.080)

and sensors as your hand does, right?

Russ Tedrake (26:17.600)

Sensors, ooh, okay.

Lex Fridman (26:19.960)

And we do amazing things with our hands.

Lex Fridman (26:23.200)

And we stick our foot in a big, solid shoe, right?

Lex Fridman (26:26.000)

So there's, I think, you know, when you're barefoot,

Russ Tedrake (26:29.640)

you're just giving yourself more proprioception.

Lex Fridman (26:33.240)

And that's why you're more aware of some of the gait flaws

Lex Fridman (26:35.720)

and stuff like this.

Lex Fridman (26:37.080)

Now, you have less protection too, so.

Russ Tedrake (26:40.720)

Rocks and stuff.

Lex Fridman (26:42.400)

I mean, yeah, so I think people who are afraid

Russ Tedrake (26:45.160)

of barefoot running are worried about getting cuts

Lex Fridman (26:47.160)

or stepping on rocks.

Russ Tedrake (26:49.800)

First of all, even if that was a concern,

Lex Fridman (26:51.560)

I think those are all, like, very short term.

Russ Tedrake (26:54.240)

You know, if I get a scratch or something,

Lex Fridman (26:55.420)

it'll heal in a week.

Russ Tedrake (26:56.520)

If I blow out my knees, I'm done running forever.

Lex Fridman (26:58.240)

So I will trade the short term for the long term anytime.

Lex Fridman (27:01.720)

But even then, you know, and this, again,

Lex Fridman (27:04.760)

to my wife's chagrin, your feet get tough, right?

Russ Tedrake (27:07.760)

And, yeah, I can run over almost anything now.

Lex Fridman (27:13.760)

I mean, what, can you talk about,

Russ Tedrake (27:17.240)

is there, like, is there tips or tricks

Lex Fridman (27:21.940)

that you have, suggestions about,

Lex Fridman (27:24.820)

like, if I wanted to try it?

Lex Fridman (27:26.620)

You know, there is a good book, actually.

Russ Tedrake (27:29.580)

There's probably more good books since I read them.

Lex Fridman (27:32.700)

But Ken Bob, Barefoot Ken Bob Saxton.

Russ Tedrake (27:37.340)

He's an interesting guy.

Lex Fridman (27:38.820)

But I think his book captures the right way

Russ Tedrake (27:42.620)

to describe running, barefoot running,

Lex Fridman (27:44.180)

to somebody better than any other I've seen.

Lex Fridman (27:48.580)

So you run pretty good distances, and you bike,

Lex Fridman (27:52.540)

and is there, you know, if we talk about bucket list items,

Russ Tedrake (27:57.820)

is there something crazy on your bucket list,

Lex Fridman (28:00.220)

athletically, that you hope to do one day?

Russ Tedrake (28:04.620)

I mean, my commute is already a little crazy.

Lex Fridman (28:07.180)

What are we talking about here?

Lex Fridman (28:09.020)

What distance are we talking about?

Lex Fridman (28:11.420)

Well, I live about 12 miles from MIT,

Lex Fridman (28:14.680)

but you can find lots of different ways to get there.

Lex Fridman (28:16.620)

So, I mean, I've run there for many years, I've biked there.

Lex Fridman (28:20.540)

Old ways?

Lex Fridman (28:21.460)

Yeah, but normally I would try to run in

Lex Fridman (28:23.900)

and then bike home, bike in, run home.

Lex Fridman (28:25.980)

But you have run there and back before?

Russ Tedrake (28:28.140)

Sure.

Lex Fridman (28:28.980)

Barefoot?

Russ Tedrake (28:29.820)

Yeah, or with minimal shoes or whatever that.

Lex Fridman (28:32.260)

12, 12 times two?

Russ Tedrake (28:34.340)

Yeah.

Lex Fridman (28:35.180)

Okay.

Lex Fridman (28:36.020)

It became kind of a game of how can I get to work?

Lex Fridman (28:38.500)

I've rollerbladed, I've done all kinds of weird stuff,

Lex Fridman (28:41.020)

but my favorite one these days,

Lex Fridman (28:42.700)

I've been taking the Charles River to work.

Russ Tedrake (28:45.060)

So, I can put in the rowboat not so far from my house,

Lex Fridman (28:50.740)

but the Charles River takes a long way to get to MIT,

Lex Fridman (28:53.300)

so I can spend a long time getting there.

Lex Fridman (28:56.380)

And it's not about, I don't know, it's just about,

Russ Tedrake (29:01.620)

I've had people ask me,

Lex Fridman (29:02.560)

how can you justify taking that time?

Lex Fridman (29:05.820)

But for me, it's just a magical time to think,

Lex Fridman (29:10.140)

to compress, decompress.

Russ Tedrake (29:13.740)

Especially, I'll wake up, do a lot of work in the morning,

Lex Fridman (29:16.220)

and then I kind of have to just let that settle

Russ Tedrake (29:19.180)

before I'm ready for all my meetings.

Lex Fridman (29:20.700)

And then on the way home, it's a great time to sort of

Russ Tedrake (29:23.160)

let that settle.

Lex Fridman (29:24.580)

You lead a large group of people.

Russ Tedrake (29:31.860)

Is there days where you're like,

Lex Fridman (29:33.980)

oh shit, I gotta get to work in an hour?

Lex Fridman (29:36.620)

Like, I mean, is there a tension there?

Lex Fridman (29:45.420)

And like, if we look at the grand scheme of things,

Russ Tedrake (29:47.940)

just like you said, long term,

Lex Fridman (29:49.500)

that meeting probably doesn't matter.

Russ Tedrake (29:51.700)

Like, you can always say, I'll just, I'll run

Lex Fridman (29:54.660)

and let the meeting happen, how it happens.

Russ Tedrake (29:57.100)

Like, what, how do you, that zen, how do you,

Lex Fridman (2:00:00.100)

So they're, I mean, they're pretty well known

Russ Tedrake (2:00:03.980)

for like autonomous vehicle research,

Lex Fridman (2:00:06.140)

but they're also interested in home robotics.

Russ Tedrake (2:00:10.260)

Yep, there's a big group working on,

Lex Fridman (2:00:12.780)

multiple groups working on home robotics.

Russ Tedrake (2:00:14.340)

It's a major part of the portfolio.

Lex Fridman (2:00:17.480)

There's also a couple other projects

Russ Tedrake (2:00:19.100)

in advanced materials discovery,

Lex Fridman (2:00:21.300)

using AI and machine learning to discover new materials

Russ Tedrake (2:00:24.420)

for car batteries and the like, for instance, yeah.

Lex Fridman (2:00:28.540)

And that's been actually an incredibly successful team.

Russ Tedrake (2:00:31.500)

There's new projects starting up too, so.

Lex Fridman (2:00:33.540)

Do you see a future of where like robots are in our home

Lex Fridman (2:00:38.940)

and like robots that have like actuators

Lex Fridman (2:00:44.040)

that look like arms in our home

Lex Fridman (2:00:46.620)

or like, you know, more like humanoid type robots?

Lex Fridman (2:00:49.340)

Or is this, are we gonna do the same thing

Russ Tedrake (2:00:51.820)

that you just mentioned that, you know,

Lex Fridman (2:00:53.860)

the dishwasher is no longer a robot.

Russ Tedrake (2:00:55.980)

We're going to just not even see them as robots.

Lex Fridman (2:00:58.700)

But I mean, what's your vision of the home of the future

Lex Fridman (2:01:02.500)

10, 20 years from now, 50 years, if you get crazy?

Lex Fridman (2:01:06.220)

Yeah, I think we already have Roombas cruising around.

Russ Tedrake (2:01:10.720)

We have, you know, Alexis or Google Homes

Lex Fridman (2:01:13.700)

on our kitchen counter.

Russ Tedrake (2:01:16.240)

It's only a matter of time until they spring arms

Lex Fridman (2:01:18.060)

and start doing something useful like that.

Lex Fridman (2:01:21.860)

So I do think it's coming.

Lex Fridman (2:01:23.860)

I think lots of people have lots of motivations

Russ Tedrake (2:01:27.660)

for doing it.

Lex Fridman (2:01:29.380)

It's been super interesting actually learning

Russ Tedrake (2:01:31.520)

about Toyota's vision for it,

Lex Fridman (2:01:33.900)

which is about helping people age in place.

Russ Tedrake (2:01:38.700)

Cause I think that's not necessarily the first entry,

Lex Fridman (2:01:41.620)

the most lucrative entry point,

Lex Fridman (2:01:44.340)

but it's the problem maybe that we really need to solve

Lex Fridman (2:01:48.680)

no matter what.

Lex Fridman (2:01:50.020)

And so I think there's a real opportunity.

Lex Fridman (2:01:53.900)

It's a delicate problem.

Lex Fridman (2:01:55.740)

How do you work with people, help people,

Lex Fridman (2:01:59.340)

keep them active, engaged, you know,

Lex Fridman (2:02:03.300)

but improve their quality of life

Lex Fridman (2:02:05.060)

and help them age in place, for instance.

Russ Tedrake (2:02:08.340)

It's interesting because older folks are also,

Lex Fridman (2:02:12.440)

I mean, there's a contrast there

Russ Tedrake (2:02:13.700)

because they're not always the folks

Lex Fridman (2:02:18.080)

who are the most comfortable with technology, for example.

Lex Fridman (2:02:20.900)

So there's a division that's interesting.

Lex Fridman (2:02:24.860)

You can do so much good with a robot for older folks,

Lex Fridman (2:02:32.020)

but there's a gap to fill of understanding.

Lex Fridman (2:02:36.380)

I mean, it's actually kind of beautiful.

Russ Tedrake (2:02:39.360)

Robot is learning about the human

Lex Fridman (2:02:41.140)

and the human is kind of learning about this new robot thing.

Lex Fridman (2:02:44.820)

And it's also with, at least with,

Lex Fridman (2:02:49.660)

like when I talked to my parents about robots,

Russ Tedrake (2:02:51.460)

there's a little bit of a blank slate there too.

Lex Fridman (2:02:54.540)

Like you can, I mean, they don't know anything

Russ Tedrake (2:02:58.020)

about robotics, so it's completely like wide open.

Lex Fridman (2:03:02.640)

They don't have, they haven't,

Russ Tedrake (2:03:03.880)

my parents haven't seen Black Mirror.

Lex Fridman (2:03:06.780)

So like they, it's a blank slate.

Lex Fridman (2:03:09.460)

Here's a cool thing, like what can it do for me?

Lex Fridman (2:03:11.980)

Yeah, so it's an exciting space.

Russ Tedrake (2:03:14.380)

I think it's a really important space.

Lex Fridman (2:03:16.340)

I do feel like a few years ago,

Russ Tedrake (2:03:20.020)

drones were successful enough in academia.

Lex Fridman (2:03:22.740)

They kind of broke out and started an industry

Lex Fridman (2:03:25.980)

and autonomous cars have been happening.

Lex Fridman (2:03:29.100)

It does feel like manipulation in logistics, of course,

Russ Tedrake (2:03:32.900)

first, but in the home shortly after,

Lex Fridman (2:03:35.700)

seems like one of the next big things

Russ Tedrake (2:03:37.180)

that's gonna really pop.

Lex Fridman (2:03:40.060)

So I don't think we talked about it,

Lex Fridman (2:03:42.100)

but what's soft robotics?

Lex Fridman (2:03:44.540)

So we talked about like rigid bodies.

Russ Tedrake (2:03:49.300)

Like if we can just linger on this whole touch thing.

Lex Fridman (2:03:52.940)

Yeah, so what's soft robotics?

Lex Fridman (2:03:54.620)

So I told you that I really dislike the fact

Lex Fridman (2:04:00.780)

that robots are afraid of touching the world

Russ Tedrake (2:04:03.140)

all over their body.

Lex Fridman (2:04:04.860)

So there's a couple reasons for that.

Russ Tedrake (2:04:06.900)

If you look carefully at all the places

Lex Fridman (2:04:08.740)

that robots actually do touch the world,

Russ Tedrake (2:04:11.220)

they're almost always soft.

Lex Fridman (2:04:12.540)

They have some sort of pad on their fingers

Russ Tedrake (2:04:14.700)

or a rubber sole on their foot.

Lex Fridman (2:04:17.900)

But if you look up and down the arm,

Russ Tedrake (2:04:19.300)

we're just pure aluminum or something.

Lex Fridman (2:04:25.340)

So that makes it hard actually.

Russ Tedrake (2:04:26.660)

In fact, hitting the table with your rigid arm

Lex Fridman (2:04:30.460)

or nearly rigid arm has some of the problems

Russ Tedrake (2:04:34.580)

that we talked about in terms of simulation.

Lex Fridman (2:04:37.260)

I think it fundamentally changes the mechanics of contact

Lex Fridman (2:04:39.940)

when you're soft, right?

Lex Fridman (2:04:41.260)

You turn point contacts into patch contacts,

Russ Tedrake (2:04:45.020)

which can have torsional friction.

Lex Fridman (2:04:47.020)

You can have distributed load.

Lex Fridman (2:04:49.260)

If I wanna pick up an egg, right?

Lex Fridman (2:04:52.460)

If I pick it up with two points,

Russ Tedrake (2:04:54.300)

then in order to put enough force

Lex Fridman (2:04:56.220)

to sustain the weight of the egg,

Russ Tedrake (2:04:57.340)

I might have to put a lot of force to break the egg.

Lex Fridman (2:04:59.980)

If I envelop it with contact all around,

Russ Tedrake (2:05:04.460)

then I can distribute my force across the shell of the egg

Lex Fridman (2:05:07.540)

and have a better chance of not breaking it.

Lex Fridman (2:05:10.620)

So soft robotics is for me a lot about changing

Lex Fridman (2:05:12.860)

the mechanics of contact.

Lex Fridman (2:05:15.500)

Does it make the problem a lot harder?

Lex Fridman (2:05:19.380)

Quite the opposite.

Russ Tedrake (2:05:24.020)

It changes the computational problem.

Lex Fridman (2:05:26.740)

I think because of the, I think our world

Lex Fridman (2:05:30.460)

and our mathematics has biased us towards rigid.

Lex Fridman (2:05:34.180)

I see.

Lex Fridman (2:05:35.020)

But it really should make things better in some ways, right?

Lex Fridman (2:05:40.740)

I think the future is unwritten there.

Lex Fridman (2:05:44.620)

But the other thing it can do.

Lex Fridman (2:05:45.460)

I think ultimately, sorry to interrupt,

Lex Fridman (2:05:46.820)

but I think ultimately it will make things simpler

Lex Fridman (2:05:49.540)

if we embrace the softness of the world.

Lex Fridman (2:05:51.580)

It makes things smoother, right?

Lex Fridman (2:05:55.740)

So the result of small actions is less discontinuous,

Lex Fridman (2:06:00.740)

but it also means potentially less instantaneously bad.

Lex Fridman (2:06:05.980)

For instance, I won't necessarily contact something

Lex Fridman (2:06:09.060)

and send it flying off.

Lex Fridman (2:06:12.300)

The other aspect of it

Russ Tedrake (2:06:13.140)

that just happens to dovetail really well

Lex Fridman (2:06:14.860)

is that soft robotics tends to be a place

Russ Tedrake (2:06:17.260)

where we can embed a lot of sensors too.

Lex Fridman (2:06:19.100)

So if you change your hardware and make it more soft,

Russ Tedrake (2:06:23.540)

then you can potentially have a tactile sensor,

Lex Fridman (2:06:25.620)

which is measuring the deformation.

Lex Fridman (2:06:27.820)

So there's a team at TRI that's working on soft hands

Lex Fridman (2:06:32.180)

and you get so much more information.

Russ Tedrake (2:06:35.500)

You can put a camera behind the skin roughly

Lex Fridman (2:06:38.820)

and get fantastic tactile information,

Russ Tedrake (2:06:42.860)

which is, it's super important.

Lex Fridman (2:06:46.180)

Like in manipulation,

Russ Tedrake (2:06:47.020)

one of the things that really is frustrating

Lex Fridman (2:06:49.820)

is if you work super hard on your head mounted,

Russ Tedrake (2:06:52.140)

on your perception system for your head mounted cameras,

Lex Fridman (2:06:54.540)

and then you get a lot of information

Russ Tedrake (2:06:56.060)

for your head mounted cameras,

Lex Fridman (2:06:57.700)

and then you've identified an object,

Russ Tedrake (2:06:59.460)

you reach down to touch it,

Lex Fridman (2:07:00.380)

and the last thing that happens,

Russ Tedrake (2:07:01.900)

right before the most important time,

Lex Fridman (2:07:03.980)

you stick your hand

Lex Fridman (2:07:04.820)

and you're occluding your head mounted sensors.

Lex Fridman (2:07:07.380)

So in all the part that really matters,

Russ Tedrake (2:07:10.220)

all of your off board sensors are occluded.

Lex Fridman (2:07:13.580)

And really, if you don't have tactile information,

Russ Tedrake (2:07:15.900)

then you're blind in an important way.

Lex Fridman (2:07:19.300)

So it happens that soft robotics and tactile sensing

Russ Tedrake (2:07:23.140)

tend to go hand in hand.

Lex Fridman (2:07:25.100)

I think we've kind of talked about it,

Lex Fridman (2:07:26.820)

but you taught a course on underactuated robotics.

Lex Fridman (2:07:31.060)

I believe that was the name of it, actually.

Russ Tedrake (2:07:32.780)

That's right.

Lex Fridman (2:07:34.980)

Can you talk about it in that context?

Lex Fridman (2:07:37.340)

What is underactuated robotics?

Lex Fridman (2:07:40.380)

Right, so underactuated robotics is my graduate course.

Russ Tedrake (2:07:43.740)

It's online mostly now,

Lex Fridman (2:07:46.620)

in the sense that the lectures.

Russ Tedrake (2:07:47.460)

Several versions of it, I think.

Lex Fridman (2:07:49.060)

Right, the YouTube.

Russ Tedrake (2:07:49.900)

It's really great, I recommend it highly.

Lex Fridman (2:07:52.060)

Look on YouTube for the 2020 versions.

Russ Tedrake (2:07:55.060)

Until March, and then you have to go back to 2019,

Lex Fridman (2:07:57.460)

thanks to COVID.

Russ Tedrake (2:08:00.740)

No, I've poured my heart into that class.

Lex Fridman (2:08:04.820)

And lecture one is basically explaining

Lex Fridman (2:08:06.620)

what the word underactuated means.

Lex Fridman (2:08:07.940)

So people are very kind to show up

Lex Fridman (2:08:09.860)

and then maybe have to learn

Lex Fridman (2:08:12.220)

what the title of the course means

Russ Tedrake (2:08:13.460)

over the course of the first lecture.

Lex Fridman (2:08:15.420)

That first lecture is really good.

Russ Tedrake (2:08:17.500)

You should watch it.

Lex Fridman (2:08:18.780)

Thanks.

Russ Tedrake (2:08:19.860)

It's a strange name,

Lex Fridman (2:08:21.500)

but I thought it captured the essence

Russ Tedrake (2:08:25.860)

of what control was good at doing

Lex Fridman (2:08:27.940)

and what control was bad at doing.

Lex Fridman (2:08:29.980)

So what do I mean by underactuated?

Lex Fridman (2:08:31.940)

So a mechanical system

Russ Tedrake (2:08:36.340)

has many degrees of freedom, for instance.

Lex Fridman (2:08:39.500)

I think of a joint as a degree of freedom.

Lex Fridman (2:08:41.940)

And it has some number of actuators, motors.

Lex Fridman (2:08:46.180)

So if you have a robot that's bolted to the table

Russ Tedrake (2:08:49.220)

that has five degrees of freedom and five motors,

Lex Fridman (2:08:54.100)

then you have a fully actuated robot.

Russ Tedrake (2:08:57.140)

If you take away one of those motors,

Lex Fridman (2:09:00.540)

then you have an underactuated robot.

Lex Fridman (2:09:03.180)

Now, why on earth?

Lex Fridman (2:09:04.940)

I have a good friend who likes to tease me.

Russ Tedrake (2:09:07.460)

He said, Ross, if you had more research funding,

Lex Fridman (2:09:09.500)

would you work on fully actuated robots?

Russ Tedrake (2:09:11.740)

Yeah.

Lex Fridman (2:09:12.580)

And the answer is no.

Russ Tedrake (2:09:15.180)

The world gives us underactuated robots,

Lex Fridman (2:09:17.420)

whether we like it or not.

Russ Tedrake (2:09:18.460)

I'm a human.

Lex Fridman (2:09:19.860)

I'm an underactuated robot,

Russ Tedrake (2:09:21.500)

even though I have more muscles

Lex Fridman (2:09:23.540)

than my big degrees of freedom,

Russ Tedrake (2:09:25.220)

because I have in some places

Lex Fridman (2:09:27.740)

multiple muscles attached to the same joint.

Lex Fridman (2:09:30.820)

But still, there's a really important degree of freedom

Lex Fridman (2:09:33.900)

that I have, which is the location of my center of mass

Russ Tedrake (2:09:37.140)

in space, for instance.

Lex Fridman (2:09:39.580)

All right, I can jump into the air,

Lex Fridman (2:09:42.500)

and there's no motor that connects my center of mass

Lex Fridman (2:09:45.220)

to the ground in that case.

Lex Fridman (2:09:47.220)

So I have to think about the implications

Lex Fridman (2:09:49.420)

of not having control over everything.

Russ Tedrake (2:09:52.740)

The passive dynamic walkers are the extreme view of that,

Lex Fridman (2:09:56.540)

where you've taken away all the motors,

Lex Fridman (2:09:57.860)

and you have to let physics do the work.

Lex Fridman (2:09:59.980)

But it shows up in all of the walking robots,

Russ Tedrake (2:10:02.220)

where you have to use some of the actuators

Lex Fridman (2:10:04.540)

to push and pull even the degrees of freedom

Russ Tedrake (2:10:06.980)

that you don't have an actuator on.

Lex Fridman (2:10:09.980)

That's referring to walking if you're falling forward.

Lex Fridman (2:10:13.140)

Is there a way to walk that's fully actuated?

Lex Fridman (2:10:16.260)

So it's a subtle point.

Russ Tedrake (2:10:18.340)

When you're in contact and you have your feet on the ground,

Lex Fridman (2:10:23.940)

there are still limits to what you can do, right?

Russ Tedrake (2:10:26.540)

Unless I have suction cups on my feet,

Lex Fridman (2:10:29.140)

I cannot accelerate my center of mass towards the ground

Russ Tedrake (2:10:32.620)

faster than gravity,

Lex Fridman (2:10:33.780)

because I can't get a force pushing me down, right?

Lex Fridman (2:10:37.420)

But I can still do most of the things that I want to.

Lex Fridman (2:10:39.420)

So you can get away with basically thinking of the system

Russ Tedrake (2:10:42.460)

as fully actuated,

Lex Fridman (2:10:43.420)

unless you suddenly needed to accelerate down super fast.

Lex Fridman (2:10:47.460)

But as soon as I take a step,

Lex Fridman (2:10:49.260)

I get into the more nuanced territory,

Lex Fridman (2:10:52.980)

and to get to really dynamic robots,

Lex Fridman (2:10:55.780)

or airplanes or other things,

Russ Tedrake (2:10:59.220)

I think you have to embrace the underactuated dynamics.

Lex Fridman (2:11:02.620)

Manipulation, people think, is manipulation underactuated?

Russ Tedrake (2:11:06.940)

Even if my arm is fully actuated, I have a motor,

Lex Fridman (2:11:10.580)

if my goal is to control the position and orientation

Russ Tedrake (2:11:14.260)

of this cup, then I don't have an actuator

Lex Fridman (2:11:18.460)

for that directly.

Lex Fridman (2:11:19.300)

So I have to use my actuators over here

Lex Fridman (2:11:21.100)

to control this thing.

Russ Tedrake (2:11:23.380)

Now it gets even worse,

Lex Fridman (2:11:24.340)

like what if I have to button my shirt, okay?

Lex Fridman (2:11:29.300)

What are the degrees of freedom of my shirt, right?

Lex Fridman (2:11:31.340)

I suddenly, that's a hard question to think about.

Russ Tedrake (2:11:34.540)

It kind of makes me queasy

Lex Fridman (2:11:36.740)

thinking about my state space control ideas.

Lex Fridman (2:11:40.740)

But actually those are the problems

Lex Fridman (2:11:41.820)

that make me so excited about manipulation right now,

Russ Tedrake (2:11:44.540)

is that it breaks some of the,

Lex Fridman (2:11:48.020)

it breaks a lot of the foundational control stuff

Russ Tedrake (2:11:50.060)

that I've been thinking about.

Lex Fridman (2:11:51.420)

Is there, what are some interesting insights

Russ Tedrake (2:11:54.580)

you could say about trying to solve an underactuated,

Lex Fridman (2:11:58.060)

a control in an underactuated system?

Lex Fridman (2:12:02.380)

So I think the philosophy there

Lex Fridman (2:12:04.820)

is let physics do more of the work.

Russ Tedrake (2:12:08.460)

The technical approach has been optimization.

Lex Fridman (2:12:12.180)

So you typically formulate your decision making

Russ Tedrake (2:12:14.260)

for control as an optimization problem.

Lex Fridman (2:12:17.140)

And you use the language of optimal control

Lex Fridman (2:12:19.420)

and sometimes often numerical optimal control

Lex Fridman (2:12:22.780)

in order to make those decisions and balance,

Russ Tedrake (2:12:26.620)

these complicated equations of,

Lex Fridman (2:12:29.100)

and in order to control,

Russ Tedrake (2:12:30.900)

you don't have to use optimal control

Lex Fridman (2:12:33.140)

to do underactuated systems,

Lex Fridman (2:12:34.900)

but that has been the technical approach

Lex Fridman (2:12:36.340)

that has borne the most fruit in our,

Russ Tedrake (2:12:39.100)

at least in our line of work.

Lex Fridman (2:12:40.900)

And there's some, so in underactuated systems,

Russ Tedrake (2:12:44.060)

when you say let physics do some of the work,

Lex Fridman (2:12:46.820)

so there's a kind of feedback loop

Russ Tedrake (2:12:50.380)

that observes the state that the physics brought you to.

Lex Fridman (2:12:54.540)

So like you've, there's a perception there,

Russ Tedrake (2:12:57.780)

there's a feedback somehow.

Lex Fridman (2:13:00.420)

Do you ever loop in like complicated perception systems

Lex Fridman (2:13:05.420)

into this whole picture?

Lex Fridman (2:13:06.900)

Right, right around the time of the DARPA challenge,

Russ Tedrake (2:13:09.620)

we had a complicated perception system

Lex Fridman (2:13:11.340)

in the DARPA challenge.

Russ Tedrake (2:13:12.700)

We also started to embrace perception

Lex Fridman (2:13:15.580)

for our flying vehicles at the time.

Russ Tedrake (2:13:17.340)

We had a really good project

Lex Fridman (2:13:20.100)

on trying to make airplanes fly

Russ Tedrake (2:13:21.820)

at high speeds through forests.

Lex Fridman (2:13:24.780)

Sirtash Karaman was on that project

Lex Fridman (2:13:27.460)

and we had, it was a really fun team to work on.

Lex Fridman (2:13:30.700)

He's carried it farther, much farther forward since then.

Lex Fridman (2:13:34.220)

And that's using cameras for perception?

Lex Fridman (2:13:35.980)

So that was using cameras.

Russ Tedrake (2:13:37.580)

That was, at the time we felt like LIDAR

Lex Fridman (2:13:40.300)

was too heavy and too power heavy

Russ Tedrake (2:13:44.860)

to be carried on a light UAV,

Lex Fridman (2:13:47.740)

and we were using cameras.

Lex Fridman (2:13:49.220)

And that was a big part of it was just

Lex Fridman (2:13:50.660)

how do you do even stereo matching

Russ Tedrake (2:13:53.100)

at a fast enough rate with a small camera,

Lex Fridman (2:13:56.460)

small onboard compute.

Russ Tedrake (2:13:58.620)

Since then we have now,

Lex Fridman (2:14:00.700)

so the deep learning revolution

Russ Tedrake (2:14:02.140)

unquestionably changed what we can do

Lex Fridman (2:14:05.540)

with perception for robotics and control.

Lex Fridman (2:14:09.020)

So in manipulation, we can address,

Lex Fridman (2:14:11.020)

we can use perception in I think a much deeper way.

Lex Fridman (2:14:14.660)

And we get into not only,

Lex Fridman (2:14:17.340)

I think the first use of it naturally

Russ Tedrake (2:14:19.820)

would be to ask your deep learning system

Lex Fridman (2:14:22.940)

to look at the cameras and produce the state,

Russ Tedrake (2:14:25.980)

which is like the pose of my thing, for instance.

Lex Fridman (2:14:28.900)

But I think we've quickly found out

Russ Tedrake (2:14:30.460)

that that's not always the right thing to do.

Lex Fridman (2:14:34.460)

Why is that?

Lex Fridman (2:14:35.620)

Because what's the state of my shirt?

Lex Fridman (2:14:38.420)

Imagine, I've always,

Lex Fridman (2:14:39.740)

Very noisy, you mean, or?

Lex Fridman (2:14:41.300)

It's, if the first step of me trying to button my shirt

Russ Tedrake (2:14:46.140)

is estimate the full state of my shirt,

Lex Fridman (2:14:48.580)

including like what's happening in the back here,

Russ Tedrake (2:14:50.460)

whatever, whatever.

Lex Fridman (2:14:51.820)

That's just not the right specification.

Russ Tedrake (2:14:55.780)

There are aspects of the state

Lex Fridman (2:14:57.500)

that are very important to the task.

Russ Tedrake (2:15:00.260)

There are many that are unobservable

Lex Fridman (2:15:03.220)

and not important to the task.

Lex Fridman (2:15:05.860)

So you really need,

Lex Fridman (2:15:06.940)

it begs new questions about state representation.

Russ Tedrake (2:15:11.100)

Another example that we've been playing with in lab

Lex Fridman (2:15:13.100)

has been just the idea of chopping onions, okay?

Russ Tedrake (2:15:17.660)

Or carrots, turns out to be better.

Lex Fridman (2:15:20.540)

So onions stink up the lab.

Lex Fridman (2:15:22.500)

And they're hard to see in a camera.

Lex Fridman (2:15:26.220)

But so,

Russ Tedrake (2:15:27.900)

Details matter, yeah.

Lex Fridman (2:15:28.740)

Details matter, you know?

Lex Fridman (2:15:30.180)

So if I'm moving around a particular object, right?

Lex Fridman (2:15:35.220)

Then I think about,

Russ Tedrake (2:15:36.060)

oh, it's got a position or an orientation in space.

Lex Fridman (2:15:38.020)

That's the description I want.

Lex Fridman (2:15:39.780)

Now, when I'm chopping an onion, okay?

Lex Fridman (2:15:42.300)

Like the first chop comes down.

Russ Tedrake (2:15:44.260)

I have now a hundred pieces of onion.

Lex Fridman (2:15:48.420)

Does my control system really need to understand

Russ Tedrake (2:15:50.300)

the position and orientation and even the shape

Lex Fridman (2:15:52.660)

of the hundred pieces of onion in order to make a decision?

Lex Fridman (2:15:56.100)

Probably not, you know?

Lex Fridman (2:15:56.940)

And if I keep going, I'm just getting,

Lex Fridman (2:15:58.900)

more and more is my state space getting bigger as I cut?

Lex Fridman (2:16:04.740)

It's not right.

Lex Fridman (2:16:06.020)

So somehow there's a,

Lex Fridman (2:16:08.100)

I think there's a richer idea of state.

Russ Tedrake (2:16:13.100)

It's not the state that is given to us

Lex Fridman (2:16:15.740)

by Lagrangian mechanics.

Russ Tedrake (2:16:17.180)

There is a proper Lagrangian state of the system,

Lex Fridman (2:16:21.340)

but the relevant state for this is some latent state

Russ Tedrake (2:16:26.460)

is what we call it in machine learning.

Lex Fridman (2:16:28.540)

But, you know, there's some different state representation.

Russ Tedrake (2:16:32.180)

Some compressed representation, some.

Lex Fridman (2:16:35.020)

And that's what I worry about saying compressed

Russ Tedrake (2:16:37.260)

because it doesn't,

Lex Fridman (2:16:38.260)

I don't mind that it's low dimensional or not,

Lex Fridman (2:16:43.020)

but it has to be something that's easier to think about.

Lex Fridman (2:16:46.260)

By us humans.

Russ Tedrake (2:16:48.460)

Or my algorithms.

Lex Fridman (2:16:49.300)

Or the algorithms being like control, optimal.

Lex Fridman (2:16:53.860)

So for instance, if the contact mechanics

Lex Fridman (2:16:56.540)

of all of those onion pieces and all the permutations

Russ Tedrake (2:16:59.660)

of possible touches between those onion pieces,

Lex Fridman (2:17:02.540)

you know, you can give me

Russ Tedrake (2:17:03.620)

a high dimensional state representation,

Lex Fridman (2:17:05.100)

I'm okay if it's linear.

Lex Fridman (2:17:06.780)

But if I have to think about all the possible

Lex Fridman (2:17:08.660)

shattering combinatorics of that,

Russ Tedrake (2:17:11.700)

then my robot's gonna sit there thinking

Lex Fridman (2:17:13.860)

and the soup's gonna get cold or something.

Lex Fridman (2:17:17.380)

So since you taught the course,

Lex Fridman (2:17:20.100)

it kind of entered my mind,

Russ Tedrake (2:17:22.740)

the idea of underactuated as really compelling

Lex Fridman (2:17:25.980)

to see the world in this kind of way.

Lex Fridman (2:17:29.540)

Do you ever, you know, if we talk about onions

Lex Fridman (2:17:32.420)

or you talk about the world with people in it in general,

Lex Fridman (2:17:35.480)

do you see the world as basically an underactuated system?

Lex Fridman (2:17:39.980)

Do you like often look at the world in this way?

Lex Fridman (2:17:42.380)

Or is this overreach?

Lex Fridman (2:17:47.040)

Underactuated is a way of life, man.

Russ Tedrake (2:17:49.160)

Exactly, I guess that's what I'm asking.

Lex Fridman (2:17:53.560)

I do think it's everywhere.

Russ Tedrake (2:17:54.960)

I think in some places,

Lex Fridman (2:17:58.840)

we already have natural tools to deal with it.

Russ Tedrake (2:18:01.380)

You know, it rears its head.

Lex Fridman (2:18:02.480)

I mean, in linear systems, it's not a problem.

Russ Tedrake (2:18:04.280)

We just, like an underactuated linear system

Lex Fridman (2:18:07.340)

is really not sufficiently distinct

Russ Tedrake (2:18:09.000)

from a fully actuated linear system.

Lex Fridman (2:18:10.760)

It's a subtle point about when that becomes a bottleneck

Russ Tedrake (2:18:15.600)

in what we know how to do with control.

Lex Fridman (2:18:17.220)

It happens to be a bottleneck,

Russ Tedrake (2:18:19.800)

although we've gotten incredibly good solutions now,

Lex Fridman (2:18:22.500)

but for a long time that I felt

Russ Tedrake (2:18:24.200)

that that was the key bottleneck in legged robots.

Lex Fridman (2:18:27.100)

And roughly now the underactuated course

Russ Tedrake (2:18:29.200)

is me trying to tell people everything I can

Lex Fridman (2:18:33.840)

about how to make Atlas do a backflip, right?

Russ Tedrake (2:18:38.500)

I have a second course now

Lex Fridman (2:18:39.920)

that I teach in the other semesters,

Russ Tedrake (2:18:41.280)

which is on manipulation.

Lex Fridman (2:18:43.600)

And that's where we get into now more of the,

Russ Tedrake (2:18:45.840)

that's a newer class.

Lex Fridman (2:18:47.160)

I'm hoping to put it online this fall completely.

Lex Fridman (2:18:51.600)

And that's gonna have much more aspects

Lex Fridman (2:18:53.700)

about these perception problems

Lex Fridman (2:18:55.460)

and the state representation questions,

Lex Fridman (2:18:57.200)

and then how do you do control.

Lex Fridman (2:18:59.260)

And the thing that's a little bit sad is that,

Lex Fridman (2:19:04.040)

for me at least, is there's a lot of manipulation tasks

Russ Tedrake (2:19:07.480)

that people wanna do and should wanna do.

Lex Fridman (2:19:09.280)

They could start a company with it and be very successful

Russ Tedrake (2:19:12.740)

that don't actually require you to think that much

Lex Fridman (2:19:15.600)

about underact, or dynamics at all even,

Lex Fridman (2:19:18.040)

but certainly underactuated dynamics.

Lex Fridman (2:19:20.020)

Once I have, if I reach out and grab something,

Russ Tedrake (2:19:23.100)

if I can sort of assume it's rigidly attached to my hand,

Lex Fridman (2:19:25.720)

then I can do a lot of interesting,

Russ Tedrake (2:19:26.920)

meaningful things with it

Lex Fridman (2:19:28.800)

without really ever thinking about the dynamics

Russ Tedrake (2:19:30.960)

of that object.

Lex Fridman (2:19:32.860)

So we've built systems that kind of reduce the need for that.

Russ Tedrake (2:19:37.860)

Enveloping grasps and the like.

Lex Fridman (2:19:40.780)

But I think the really good problems in manipulation.

Lex Fridman (2:19:43.060)

So manipulation, by the way, is more than just pick and place.

Lex Fridman (2:19:48.540)

That's like a lot of people think of that, just grasping.

Russ Tedrake (2:19:51.780)

I don't mean that.

Lex Fridman (2:19:52.620)

I mean buttoning my shirt, I mean tying shoelaces.

Lex Fridman (2:19:56.500)

How do you program a robot to tie shoelaces?

Lex Fridman (2:19:59.060)

And not just one shoe, but every shoe, right?

Russ Tedrake (2:20:02.860)

That's a really good problem.

Lex Fridman (2:20:05.580)

It's tempting to write down like the infinite dimensional

Russ Tedrake (2:20:08.420)

state of the laces, that's probably not needed

Lex Fridman (2:20:13.180)

to write a good controller.

Russ Tedrake (2:20:15.100)

I know we could hand design a controller that would do it,

Lex Fridman (2:20:18.340)

but I don't want that.

Russ Tedrake (2:20:19.180)

I want to understand the principles that would allow me

Lex Fridman (2:20:22.460)

to solve another problem that's kind of like that.

Lex Fridman (2:20:25.380)

But I think if we can stay pure in our approach,

Lex Fridman (2:20:29.820)

then the challenge of tying anybody's shoes

Russ Tedrake (2:20:33.820)

is a great challenge.

Lex Fridman (2:20:36.300)

That's a great challenge.

Russ Tedrake (2:20:37.220)

I mean, and the soft touch comes into play there.

Lex Fridman (2:20:40.940)

That's really interesting.

Russ Tedrake (2:20:43.100)

Let me ask another ridiculous question on this topic.

Lex Fridman (2:20:47.500)

How important is touch?

Russ Tedrake (2:20:49.780)

We haven't talked much about humans,

Lex Fridman (2:20:52.300)

but I have this argument with my dad

Russ Tedrake (2:20:56.220)

where like I think you can fall in love with a robot

Lex Fridman (2:20:59.620)

based on language alone.

Lex Fridman (2:21:02.580)

And he believes that touch is essential.

Lex Fridman (2:21:06.460)

Touch and smell, he says.

Lex Fridman (2:21:07.660)

But so in terms of robots, connecting with humans,

Lex Fridman (2:21:17.380)

we can go philosophical in terms of like a deep,

Russ Tedrake (2:21:19.660)

meaningful connection, like love,

Lex Fridman (2:21:21.820)

but even just like collaborating in an interesting way,

Lex Fridman (2:21:25.580)

how important is touch like from an engineering perspective

Lex Fridman (2:21:30.580)

and a philosophical one?

Russ Tedrake (2:21:32.780)

I think it's super important.

Lex Fridman (2:21:35.700)

Even just in a practical sense,

Russ Tedrake (2:21:37.020)

if we forget about the emotional part of it.

Lex Fridman (2:21:40.700)

But for robots to interact safely

Russ Tedrake (2:21:43.300)

while they're doing meaningful mechanical work

Lex Fridman (2:21:47.220)

in the close contact with or vicinity of people

Russ Tedrake (2:21:52.420)

that need help, I think we have to have them,

Lex Fridman (2:21:55.220)

we have to build them differently.

Russ Tedrake (2:21:57.500)

They have to be afraid, not afraid of touching the world.

Lex Fridman (2:21:59.860)

So I think Baymax is just awesome.

Russ Tedrake (2:22:02.820)

That's just like the movie of Big Hero 6

Lex Fridman (2:22:06.260)

and the concept of Baymax, that's just awesome.

Russ Tedrake (2:22:08.700)

I think we should, and we have some folks at Toyota

Lex Fridman (2:22:13.060)

that are trying to, Toyota Research

Russ Tedrake (2:22:14.420)

that are trying to build Baymax roughly.

Lex Fridman (2:22:16.860)

And I think it's just a fantastically good project.

Russ Tedrake (2:22:21.900)

I think it will change the way people physically interact.

Lex Fridman (2:22:25.620)

The same way, I mean, you gave a couple examples earlier,

Lex Fridman (2:22:27.980)

but if the robot that was walking around my home

Lex Fridman (2:22:31.940)

looked more like a teddy bear

Lex Fridman (2:22:33.980)

and a little less like the Terminator,

Lex Fridman (2:22:35.980)

that could change completely the way people perceive it

Lex Fridman (2:22:38.900)

and interact with it.

Lex Fridman (2:22:39.820)

And maybe they'll even wanna teach it, like you said, right?

Russ Tedrake (2:22:44.340)

You could not quite gamify it,

Lex Fridman (2:22:47.660)

but somehow instead of people judging it

Lex Fridman (2:22:50.060)

and looking at it as if it's not doing as well as a human,

Lex Fridman (2:22:54.340)

they're gonna try to help out the cute teddy bear, right?

Russ Tedrake (2:22:57.060)

Who knows, but I think we're building robots wrong

Lex Fridman (2:23:01.260)

and being more soft and more contact is important, right?

Russ Tedrake (2:23:07.780)

Yeah, I mean, like all the magical moments

Lex Fridman (2:23:09.860)

I can remember with robots,

Russ Tedrake (2:23:12.380)

well, first of all, just visiting your lab and seeing Atlas,

Lex Fridman (2:23:16.900)

but also Spotmini, when I first saw Spotmini in person

Lex Fridman (2:23:21.660)

and hung out with him, her, it,

Lex Fridman (2:23:26.260)

I don't have trouble engendering robots.

Lex Fridman (2:23:28.380)

I feel the robotics people really say, oh, is it it?

Lex Fridman (2:23:31.500)

I kinda like the idea that it's a her or a him.

Russ Tedrake (2:23:35.780)

There's a magical moment, but there's no touching.

Lex Fridman (2:23:38.780)

I guess the question I have, have you ever been,

Russ Tedrake (2:23:41.620)

like, have you had a human robot experience

Lex Fridman (2:23:44.940)

where a robot touched you?

Lex Fridman (2:23:49.580)

And like, it was like, wait,

Lex Fridman (2:23:51.660)

like, was there a moment that you've forgotten

Russ Tedrake (2:23:53.980)

that a robot is a robot and like,

Lex Fridman (2:23:57.740)

the anthropomorphization stepped in

Lex Fridman (2:24:00.820)

and for a second you forgot that it's not human?

Lex Fridman (2:24:04.900)

I mean, I think when you're in on the details,

Russ Tedrake (2:24:07.820)

then we, of course, anthropomorphized our work with Atlas,

Lex Fridman (2:24:12.380)

but in verbal communication and the like,

Russ Tedrake (2:24:17.100)

I think we were pretty aware of it

Lex Fridman (2:24:18.980)

as a machine that needed to be respected.

Lex Fridman (2:24:21.740)

And I actually, I worry more about the smaller robots

Lex Fridman (2:24:26.260)

that could still move quickly if programmed wrong

Lex Fridman (2:24:29.540)

and we have to be careful actually

Lex Fridman (2:24:31.660)

about safety and the like right now.

Lex Fridman (2:24:33.740)

And that, if we build our robots correctly,

Lex Fridman (2:24:36.380)

I think then those, a lot of those concerns could go away.

Lex Fridman (2:24:40.300)

And we're seeing that trend.

Lex Fridman (2:24:41.260)

We're seeing the lower cost, lighter weight arms now

Russ Tedrake (2:24:44.100)

that could be fundamentally safe.

Lex Fridman (2:24:46.740)

I mean, I do think touch is so fundamental.

Russ Tedrake (2:24:49.060)

Ted Adelson is great.

Lex Fridman (2:24:51.100)

He's a perceptual scientist at MIT

Lex Fridman (2:24:55.740)

and he studied vision most of his life.

Lex Fridman (2:24:58.180)

And he said, when I had kids,

Russ Tedrake (2:25:01.220)

I expected to be fascinated by their perceptual development.

Lex Fridman (2:25:06.380)

But what really, what he noticed was,

Russ Tedrake (2:25:09.260)

felt more impressive, more dominant

Lex Fridman (2:25:10.780)

was the way that they would touch everything

Lex Fridman (2:25:13.060)

and lick everything.

Lex Fridman (2:25:13.900)

And pick things up, stick it on their tongue and whatever.

Lex Fridman (2:25:16.900)

And he said, watching his daughter convinced him

Lex Fridman (2:25:22.180)

that actually he needed to study tactile sensing more.

Lex Fridman (2:25:25.580)

So there's something very important.

Lex Fridman (2:25:30.580)

I think it's a little bit also of the passive

Lex Fridman (2:25:32.780)

versus active part of the world, right?

Lex Fridman (2:25:35.660)

You can passively perceive the world.

Lex Fridman (2:25:38.460)

But it's fundamentally different if you can do an experiment

Lex Fridman (2:25:41.460)

and if you can change the world

Lex Fridman (2:25:43.340)

and you can learn a lot more than a passive observer.

Lex Fridman (2:25:47.460)

So you can in dialogue, that was your initial example,

Russ Tedrake (2:25:51.500)

you could have an active experiment exchange.

Lex Fridman (2:25:54.580)

But I think if you're just a camera watching YouTube,

Russ Tedrake (2:25:57.460)

I think that's a very different problem

Lex Fridman (2:26:00.380)

than if you're a robot that can apply force.

Lex Fridman (2:26:03.700)

And I think that's a very different problem

Lex Fridman (2:26:05.900)

than if you're a robot that can apply force and touch.

Russ Tedrake (2:26:13.260)

I think it's important.

Lex Fridman (2:26:15.540)

Yeah, I think it's just an exciting area of research.

Russ Tedrake (2:26:18.020)

I think you're probably right

Lex Fridman (2:26:19.260)

that this hasn't been under researched.

Russ Tedrake (2:26:23.900)

To me as a person who's captivated

Lex Fridman (2:26:25.780)

by the idea of human robot interaction,

Russ Tedrake (2:26:27.820)

it feels like such a rich opportunity to explore touch.

Lex Fridman (2:26:34.140)

Not even from a safety perspective,

Lex Fridman (2:26:35.860)

but like you said, the emotional too.

Lex Fridman (2:26:38.060)

I mean, safety comes first,

Lex Fridman (2:26:41.220)

but the next step is like a real human connection.

Lex Fridman (2:26:48.300)

Even in the industrial setting,

Russ Tedrake (2:26:51.380)

it just feels like it's nice for the robot.

Lex Fridman (2:26:55.540)

I don't know, you might disagree with this,

Lex Fridman (2:26:58.060)

but because I think it's important

Lex Fridman (2:27:01.220)

to see robots as tools often,

Lex Fridman (2:27:04.340)

but I don't know,

Lex Fridman (2:27:06.060)

I think they're just always going to be more effective

Russ Tedrake (2:27:08.540)

once you humanize them.

Lex Fridman (2:27:11.700)

Like it's convenient now to think of them as tools

Russ Tedrake (2:27:14.340)

because we want to focus on the safety,

Lex Fridman (2:27:16.140)

but I think ultimately to create like a good experience

Russ Tedrake (2:27:22.300)

for the worker, for the person,

Lex Fridman (2:27:24.860)

there has to be a human element.

Russ Tedrake (2:27:27.980)

I don't know, for me,

Lex Fridman (2:27:30.140)

it feels like an industrial robotic arm

Russ Tedrake (2:27:33.140)

would be better if it has a human element.

Lex Fridman (2:27:34.860)

I think like Rethink Robotics had that idea

Russ Tedrake (2:27:37.060)

with the Baxter and having eyes and so on,

Lex Fridman (2:27:40.260)

having, I don't know, I'm a big believer in that.

Russ Tedrake (2:27:45.220)

It's not my area, but I am also a big believer.

Lex Fridman (2:27:49.300)

Do you have an emotional connection to Atlas?

Lex Fridman (2:27:51.900)

Like do you miss him?

Lex Fridman (2:27:54.940)

I mean, yes, I don't know if I more so

Russ Tedrake (2:27:59.940)

than if I had a different science project

Lex Fridman (2:28:01.620)

that I'd worked on super hard, right?

Lex Fridman (2:28:03.420)

But yeah, I mean, the robot,

Lex Fridman (2:28:09.900)

we basically had to do heart surgery on the robot

Russ Tedrake (2:28:11.780)

in the final competition because we melted the core.

Lex Fridman (2:28:18.380)

Yeah, there was something about watching that robot

Russ Tedrake (2:28:20.140)

hanging there.

Lex Fridman (2:28:20.980)

We know we had to compete with it in an hour

Lex Fridman (2:28:22.540)

and it was getting its guts ripped out.

Lex Fridman (2:28:25.260)

Those are all historic moments.

Russ Tedrake (2:28:27.460)

I think if you look back like a hundred years from now,

Lex Fridman (2:28:32.140)

yeah, I think those are important moments in robotics.

Russ Tedrake (2:28:35.140)

I mean, these are the early days.

Lex Fridman (2:28:36.660)

You look at like the early days

Russ Tedrake (2:28:37.980)

of a lot of scientific disciplines.

Lex Fridman (2:28:39.500)

They look ridiculous, they're full of failure,

Lex Fridman (2:28:42.020)

but it feels like robotics will be important

Lex Fridman (2:28:45.060)

in the coming a hundred years.

Lex Fridman (2:28:48.940)

And these are the early days.

Lex Fridman (2:28:50.860)

So I think a lot of people are,

Russ Tedrake (2:28:54.420)

look at a brilliant person such as yourself

Lex Fridman (2:28:57.900)

and are curious about the intellectual journey they've took.

Russ Tedrake (2:29:01.740)

Is there maybe three books, technical, fiction,

Lex Fridman (2:29:06.260)

philosophical that had a big impact on your life

Lex Fridman (2:29:10.540)

that you would recommend perhaps others reading?

Lex Fridman (2:29:15.260)

Yeah, so I actually didn't read that much as a kid,

Lex Fridman (2:29:18.460)

but I read fairly voraciously now.

Lex Fridman (2:29:21.260)

There are some recent books that if you're interested

Russ Tedrake (2:29:24.940)

in this kind of topic, like AI Superpowers by Kai Fu Lee

Lex Fridman (2:29:29.940)

is just a fantastic read.

Russ Tedrake (2:29:31.660)

You must read that.

Lex Fridman (2:29:35.100)

Yuval Harari is just, I think that can open your mind.

Russ Tedrake (2:29:40.500)

Sapiens.

Lex Fridman (2:29:41.580)

Sapiens is the first one, Homo Deus is the second, yeah.

Russ Tedrake (2:29:46.980)

We mentioned it in the book,

Lex Fridman (2:29:48.340)

Homo Deus is the second, yeah.

Russ Tedrake (2:29:51.060)

We mentioned The Black Swan by Taleb.

Lex Fridman (2:29:53.500)

I think that's a good sort of mind opener.

Russ Tedrake (2:29:57.220)

I actually, so there's maybe a more controversial

Lex Fridman (2:30:04.420)

recommendation I could give.

Russ Tedrake (2:30:06.220)

Great, we love controversy.

Lex Fridman (2:30:08.740)

In some sense, it's so classical it might surprise you,

Lex Fridman (2:30:11.580)

but I actually recently read Mortimer Adler's

Lex Fridman (2:30:16.020)

How to Read a Book, not so long, it was a while ago,

Lex Fridman (2:30:19.020)

but some people hate that book.

Lex Fridman (2:30:23.220)

I loved it.

Russ Tedrake (2:30:24.820)

I think we're in this time right now where,

Lex Fridman (2:30:30.860)

boy, we're just inundated with research papers

Russ Tedrake (2:30:33.780)

that you could read on archive with limited peer review

Lex Fridman (2:30:38.580)

and just this wealth of information.

Russ Tedrake (2:30:40.980)

I don't know, I think the passion of what you can get

Lex Fridman (2:30:46.460)

out of a book, a really good book or a really good paper

Russ Tedrake (2:30:49.460)

if you find it, the attitude, the realization

Lex Fridman (2:30:52.220)

that you're only gonna find a few that really

Russ Tedrake (2:30:54.220)

are worth all your time, but then once you find them,

Lex Fridman (2:30:58.300)

you should just dig in and understand it very deeply

Lex Fridman (2:31:02.660)

and it's worth marking it up and having the hard copy

Lex Fridman (2:31:07.660)

writing in the side notes, side margins.

Russ Tedrake (2:31:11.340)

I think that was really, I read it at the right time

Lex Fridman (2:31:16.340)

where I was just feeling just overwhelmed

Russ Tedrake (2:31:19.260)

with really low quality stuff, I guess.

Lex Fridman (2:31:23.780)

And similarly, I'm just giving more than three now,

Russ Tedrake (2:31:28.780)

I'm sorry if I've exceeded my quota.

Lex Fridman (2:31:31.460)

But on that topic just real quick is,

Lex Fridman (2:31:34.140)

so basically finding a few companions to keep

Lex Fridman (2:31:38.140)

for the rest of your life in terms of papers and books

Lex Fridman (2:31:41.340)

and so on and those are the ones,

Lex Fridman (2:31:44.140)

like not doing, what is it, FOMO, fear of missing out,

Russ Tedrake (2:31:48.900)

constantly trying to update yourself,

Lex Fridman (2:31:50.820)

but really deeply making a life journey

Russ Tedrake (2:31:53.700)

of studying a particular paper, essentially, set of papers.

Lex Fridman (2:31:57.500)

Yeah, I think when you really start to understand

Russ Tedrake (2:32:02.500)

when you really find something,

Lex Fridman (2:32:06.100)

which a book that resonates with you

Russ Tedrake (2:32:07.780)

might not be the same book that resonates with me,

Lex Fridman (2:32:10.420)

but when you really find one that resonates with you,

Russ Tedrake (2:32:13.180)

I think the dialogue that happens and that's what,

Lex Fridman (2:32:16.260)

I loved that Adler was saying, I think Socrates and Plato

Russ Tedrake (2:32:20.140)

say the written word is never gonna capture

Lex Fridman (2:32:25.740)

the beauty of dialogue, right?

Lex Fridman (2:32:28.020)

But Adler says, no, no, a really good book

Lex Fridman (2:32:33.100)

is a dialogue between you and the author

Lex Fridman (2:32:35.380)

and it crosses time and space and I don't know,

Lex Fridman (2:32:39.180)

I think it's a very romantic,

Russ Tedrake (2:32:40.740)

there's a bunch of like specific advice,

Lex Fridman (2:32:42.740)

which you can just gloss over,

Lex Fridman (2:32:44.380)

but the romantic view of how to read

Lex Fridman (2:32:47.260)

and really appreciate it is so good.

Lex Fridman (2:32:52.140)

And similarly, teaching,

Lex Fridman (2:32:53.900)

yeah, I thought a lot about teaching

Lex Fridman (2:32:58.820)

and so Isaac Asimov, great science fiction writer,

Lex Fridman (2:33:03.300)

has also actually spent a lot of his career

Lex Fridman (2:33:05.340)

writing nonfiction, right?

Lex Fridman (2:33:07.260)

His memoir is fantastic.

Lex Fridman (2:33:09.940)

He was passionate about explaining things, right?

Lex Fridman (2:33:12.740)

He wrote all kinds of books

Russ Tedrake (2:33:13.700)

on all kinds of topics in science.

Lex Fridman (2:33:16.100)

He was known as the great explainer

Lex Fridman (2:33:17.740)

and I do really resonate with his style

Lex Fridman (2:33:22.340)

and just his way of talking about,

Russ Tedrake (2:33:28.420)

by communicating and explaining to something

Lex Fridman (2:33:30.540)

is really the way that you learn something.

Russ Tedrake (2:33:32.540)

I think about problems very differently

Lex Fridman (2:33:36.260)

because of the way I've been given the opportunity

Russ Tedrake (2:33:39.220)

to teach them at MIT.

Lex Fridman (2:33:42.140)

We have questions asked, the fear of the lecture,

Russ Tedrake (2:33:45.500)

the experience of the lecture

Lex Fridman (2:33:47.700)

and the questions I get and the interactions

Russ Tedrake (2:33:50.220)

just forces me to be rock solid on these ideas

Lex Fridman (2:33:53.140)

in a way that if I didn't have that,

Russ Tedrake (2:33:55.060)

I don't know, I would be in a different intellectual space.

Lex Fridman (2:33:58.260)

Also, video, does that scare you

Russ Tedrake (2:34:00.420)

that your lectures are online

Lex Fridman (2:34:02.140)

and people like me in sweatpants can sit sipping coffee

Lex Fridman (2:34:05.460)

and watch you give lectures?

Lex Fridman (2:34:08.260)

I think it's great.

Russ Tedrake (2:34:09.980)

I do think that something's changed right now,

Lex Fridman (2:34:12.820)

which is, right now we're giving lectures over Zoom.

Russ Tedrake (2:34:16.900)

I mean, giving seminars over Zoom and everything.

Lex Fridman (2:34:21.260)

I'm trying to figure out, I think it's a new medium.

Russ Tedrake (2:34:24.380)

I'm trying to figure out how to exploit it.

Lex Fridman (2:34:28.020)

Yeah, I've been quite cynical

Russ Tedrake (2:34:34.500)

about human to human connection over that medium,

Lex Fridman (2:34:39.820)

but I think that's because it hasn't been explored fully

Lex Fridman (2:34:43.420)

and teaching is a different thing.

Lex Fridman (2:34:45.780)

Every lecture is a, I'm sorry, every seminar even,

Russ Tedrake (2:34:49.100)

I think every talk I give is an opportunity

Lex Fridman (2:34:53.460)

to give that differently.

Russ Tedrake (2:34:54.980)

I can deliver content directly into your browser.

Lex Fridman (2:34:57.940)

You have a WebGL engine right there.

Russ Tedrake (2:35:00.020)

I can throw 3D content into your browser

Lex Fridman (2:35:04.900)

while you're listening to me, right?

Lex Fridman (2:35:06.900)

And I can assume that you have at least

Lex Fridman (2:35:10.020)

a powerful enough laptop or something to watch Zoom

Russ Tedrake (2:35:13.020)

while I'm doing that, while I'm giving a lecture.

Lex Fridman (2:35:15.460)

That's a new communication tool

Lex Fridman (2:35:18.060)

that I didn't have last year, right?

Lex Fridman (2:35:19.980)

And I think robotics can potentially benefit a lot

Russ Tedrake (2:35:24.180)

from teaching that way.

Lex Fridman (2:35:26.420)

We'll see, it's gonna be an experiment this fall.

Russ Tedrake (2:35:28.180)

It's interesting.

Lex Fridman (2:35:29.020)

I'm thinking a lot about it.

Russ Tedrake (2:35:30.340)

Yeah, and also like the length of lectures

Lex Fridman (2:35:35.580)

or the length of like, there's something,

Lex Fridman (2:35:38.820)

so like I guarantee you, it's like 80% of people

Lex Fridman (2:35:42.900)

who started listening to our conversation

Russ Tedrake (2:35:44.900)

are still listening to now, which is crazy to me.

Lex Fridman (2:35:48.180)

But so there's a patience and interest

Russ Tedrake (2:35:51.140)

in long form content, but at the same time,

Lex Fridman (2:35:53.540)

there's a magic to forcing yourself to condense

Russ Tedrake (2:35:57.940)

an idea to as short as possible.

Lex Fridman (2:36:02.740)

As short as possible, like clip,

Russ Tedrake (2:36:04.660)

it can be a part of a longer thing,

Lex Fridman (2:36:06.180)

but like just like really beautifully condense an idea.

Russ Tedrake (2:36:09.620)

There's a lot of opportunity there

Lex Fridman (2:36:11.900)

that's easier to do in remote with, I don't know,

Russ Tedrake (2:36:17.500)

with editing too.

Lex Fridman (2:36:19.020)

Editing is an interesting thing.

Russ Tedrake (2:36:20.980)

Like what, you know, most professors don't get,

Lex Fridman (2:36:25.020)

when they give a lecture,

Russ Tedrake (2:36:25.860)

they don't get to go back and edit out parts,

Lex Fridman (2:36:28.220)

like crisp it up a little bit.

Russ Tedrake (2:36:31.580)

That's also, it can do magic.

Lex Fridman (2:36:34.180)

Like if you remove like five to 10 minutes

Russ Tedrake (2:36:37.620)

from an hour lecture, it can actually,

Lex Fridman (2:36:41.140)

it can make something special of a lecture.

Russ Tedrake (2:36:43.220)

I've seen that in myself and in others too,

Lex Fridman (2:36:47.860)

because I edit other people's lectures to extract clips.

Russ Tedrake (2:36:50.580)

It's like, there's certain tangents that are like,

Lex Fridman (2:36:52.740)

that lose, they're not interesting.

Russ Tedrake (2:36:54.420)

They're mumbling, they're just not,

Lex Fridman (2:36:57.180)

they're not clarifying, they're not helpful at all.

Lex Fridman (2:36:59.780)

And once you remove them, it's just, I don't know.

Lex Fridman (2:37:02.820)

Editing can be magic.

Russ Tedrake (2:37:04.580)

It takes a lot of time.

Lex Fridman (2:37:05.900)

Yeah, it takes, it depends like what is teaching,

Russ Tedrake (2:37:08.940)

you have to ask.

Lex Fridman (2:37:09.780)

Yeah, yeah.

Russ Tedrake (2:37:13.100)

Cause I find the editing process is also beneficial

Lex Fridman (2:37:18.020)

as for teaching, but also for your own learning.

Lex Fridman (2:37:21.620)

I don't know if, have you watched yourself?

Lex Fridman (2:37:23.740)

Yeah, sure.

Lex Fridman (2:37:24.780)

Have you watched those videos?

Lex Fridman (2:37:26.180)

I mean, not all of them.

Russ Tedrake (2:37:27.900)

It could be painful to see like how to improve.

Lex Fridman (2:37:33.340)

So do you find that, I know you segment your podcast.

Lex Fridman (2:37:37.180)

Do you think that helps people with the,

Lex Fridman (2:37:40.740)

the attention span aspect of it?

Russ Tedrake (2:37:42.220)

Or is it the segment like sections like,

Lex Fridman (2:37:44.220)

yeah, we're talking about this topic, whatever.

Russ Tedrake (2:37:46.380)

Nope, nope, that just helps me.

Lex Fridman (2:37:48.260)

It's actually bad.

Russ Tedrake (2:37:49.420)

So, and you've been incredible.

Lex Fridman (2:37:53.820)

So I'm learning, like I'm afraid of conversation.

Russ Tedrake (2:37:56.420)

This is even today, I'm terrified of talking to you.

Lex Fridman (2:37:59.180)

I mean, it's something I'm trying to remove for myself.

Russ Tedrake (2:38:04.180)

There's a guy, I mean, I've learned from a lot of people,

Lex Fridman (2:38:07.420)

but really there's been a few people

Russ Tedrake (2:38:10.740)

who's been inspirational to me in terms of conversation.

Lex Fridman (2:38:14.100)

Whatever people think of him,

Russ Tedrake (2:38:15.700)

Joe Rogan has been inspirational to me

Lex Fridman (2:38:17.500)

because comedians have been too.

Russ Tedrake (2:38:20.500)

Being able to just have fun and enjoy themselves

Lex Fridman (2:38:23.300)

and lose themselves in conversation

Russ Tedrake (2:38:25.580)

that requires you to be a great storyteller,

Lex Fridman (2:38:28.820)

to be able to pull a lot of different pieces

Russ Tedrake (2:38:31.500)

of information together.

Lex Fridman (2:38:32.820)

But mostly just to enjoy yourself in conversations.

Lex Fridman (2:38:36.500)

And I'm trying to learn that.

Lex Fridman (2:38:38.060)

These notes are, you see me looking down.

Russ Tedrake (2:38:41.660)

That's like a safety blanket

Lex Fridman (2:38:43.020)

that I'm trying to let go of more and more.

Russ Tedrake (2:38:45.260)

Cool.

Lex Fridman (2:38:46.260)

So that's, people love just regular conversation.

Russ Tedrake (2:38:49.420)

That's what they, the structure is like, whatever.

Lex Fridman (2:38:52.660)

I would say, I would say maybe like 10 to like,

Lex Fridman (2:38:57.620)

so there's a bunch of, you know,

Lex Fridman (2:38:59.820)

there's probably a couple of thousand PhD students

Lex Fridman (2:39:03.820)

listening to this right now, right?

Lex Fridman (2:39:06.980)

And they might know what we're talking about.

Lex Fridman (2:39:09.540)

But there is somebody, I guarantee you right now,

Lex Fridman (2:39:13.460)

in Russia, some kid who's just like,

Russ Tedrake (2:39:16.580)

who's just smoked some weed, is sitting back

Lex Fridman (2:39:19.380)

and just enjoying the hell out of this conversation.

Russ Tedrake (2:39:22.580)

Not really understanding.

Lex Fridman (2:39:23.860)

He kind of watched some Boston Dynamics videos.

Russ Tedrake (2:39:25.980)

He's just enjoying it.

Lex Fridman (2:39:27.300)

And I salute you, sir.

Russ Tedrake (2:39:29.300)

No, but just like, there's so much variety of people

Lex Fridman (2:39:32.780)

that just have curiosity about engineering,

Russ Tedrake (2:39:35.260)

about sciences, about mathematics.

Lex Fridman (2:39:37.980)

And also like, I should, I mean,

Russ Tedrake (2:39:43.940)

enjoying it is one thing,

Lex Fridman (2:39:44.980)

but also often notice it inspires people to,

Russ Tedrake (2:39:49.180)

there's a lot of people who are like

Lex Fridman (2:39:50.860)

in their undergraduate studies trying to figure out what,

Russ Tedrake (2:39:54.700)

trying to figure out what to pursue.

Lex Fridman (2:39:56.140)

And these conversations can really spark

Russ Tedrake (2:39:59.220)

the direction of their life.

Lex Fridman (2:40:01.820)

And in terms of robotics, I hope it does,

Russ Tedrake (2:40:03.580)

because I'm excited about the possibilities

Lex Fridman (2:40:06.540)

of what robotics brings.

Lex Fridman (2:40:07.580)

On that topic, do you have advice?

Lex Fridman (2:40:12.580)

Like what advice would you give

Lex Fridman (2:40:14.060)

to a young person about life?

Lex Fridman (2:40:18.260)

A young person about life

Lex Fridman (2:40:19.380)

or a young person about life in robotics?

Lex Fridman (2:40:23.060)

It could be in robotics.

Russ Tedrake (2:40:24.380)

Robotics, it could be in life in general.

Lex Fridman (2:40:26.660)

It could be career.

Russ Tedrake (2:40:28.460)

It could be a relationship advice.

Lex Fridman (2:40:31.300)

It could be running advice.

Russ Tedrake (2:40:32.900)

Just like they're, that's one of the things I see,

Lex Fridman (2:40:36.620)

like we talked to like 20 year olds.

Lex Fridman (2:40:38.620)

They're like, how do I do this thing?

Lex Fridman (2:40:42.500)

What do I do?

Lex Fridman (2:40:45.620)

If they come up to you, what would you tell them?

Lex Fridman (2:40:48.020)

I think it's an interesting time to be a kid these days.

Russ Tedrake (2:40:53.980)

Everything points to this being sort of a winner,

Lex Fridman (2:40:57.860)

take all economy and the like.

Russ Tedrake (2:40:59.300)

I think the people that will really excel in my opinion

Lex Fridman (2:41:04.500)

are going to be the ones that can think deeply

Russ Tedrake (2:41:06.820)

about problems.

Lex Fridman (2:41:11.180)

You have to be able to ask questions agilely

Lex Fridman (2:41:13.940)

and use the internet for everything it's good for

Lex Fridman (2:41:15.820)

and stuff like this.

Lex Fridman (2:41:16.660)

And I think a lot of people will develop those skills.

Lex Fridman (2:41:19.460)

I think the leaders, thought leaders,

Russ Tedrake (2:41:24.820)

robotics leaders, whatever,

Lex Fridman (2:41:26.860)

are gonna be the ones that can do more

Lex Fridman (2:41:29.100)

and they can think very deeply and critically.

Lex Fridman (2:41:32.420)

And that's a harder thing to learn.

Russ Tedrake (2:41:35.020)

I think one path to learning that is through mathematics,

Lex Fridman (2:41:38.140)

through engineering.

Russ Tedrake (2:41:41.660)

I would encourage people to start math early.

Lex Fridman (2:41:44.180)

I mean, I didn't really start.

Russ Tedrake (2:41:46.900)

I mean, I was always in the better math classes

Lex Fridman (2:41:50.460)

that I could take,

Lex Fridman (2:41:51.300)

but I wasn't pursuing super advanced mathematics

Lex Fridman (2:41:54.700)

or anything like that until I got to MIT.

Russ Tedrake (2:41:56.700)

I think MIT lit me up

Lex Fridman (2:41:59.020)

and really started the life that I'm living now.

Lex Fridman (2:42:05.580)

But yeah, I really want kids to dig deep,

Lex Fridman (2:42:10.740)

really understand things, building things too.

Russ Tedrake (2:42:12.460)

I mean, pull things apart, put them back together.

Lex Fridman (2:42:15.180)

Like that's just such a good way

Russ Tedrake (2:42:17.180)

to really understand things

Lex Fridman (2:42:19.980)

and expect it to be a long journey, right?

Russ Tedrake (2:42:23.660)

It's, you don't have to know everything.

Lex Fridman (2:42:27.260)

You're never gonna know everything.

Lex Fridman (2:42:29.500)

So think deeply and stick with it.

Lex Fridman (2:42:32.860)

Enjoy the ride, but just make sure you're not,

Russ Tedrake (2:42:37.580)

yeah, just make sure you're stopping

Lex Fridman (2:42:40.580)

to think about why things work.

Lex Fridman (2:42:43.180)

And it's true, it's easy to lose yourself

Lex Fridman (2:42:45.420)

in the distractions of the world.

Russ Tedrake (2:42:51.180)

We're overwhelmed with content right now,

Lex Fridman (2:42:52.740)

but you have to stop and pick some of it

Lex Fridman (2:42:56.260)

and really understand it.

Lex Fridman (2:42:58.780)

Yeah, on the book point,

Russ Tedrake (2:43:00.380)

I've read Animal Farm by George Orwell

Lex Fridman (2:43:04.940)

a ridiculous number of times.

Lex Fridman (2:43:06.100)

So for me, like that book,

Lex Fridman (2:43:07.860)

I don't know if it's a good book in general,

Lex Fridman (2:43:09.780)

but for me it connects deeply somehow.

Lex Fridman (2:43:13.340)

It somehow connects, so I was born in the Soviet Union.

Lex Fridman (2:43:18.260)

So it connects to me into the entirety of the history

Lex Fridman (2:43:20.460)

of the Soviet Union and to World War II

Lex Fridman (2:43:23.180)

and to the love and hatred and suffering

Lex Fridman (2:43:26.500)

that went on there and the corrupting nature of power

Lex Fridman (2:43:33.140)

and greed and just somehow I just,

Lex Fridman (2:43:36.340)

that book has taught me more about life

Russ Tedrake (2:43:38.100)

than like anything else.

Lex Fridman (2:43:39.380)

Even though it's just like a silly childlike book

Russ Tedrake (2:43:42.860)

about pigs, I don't know why,

Lex Fridman (2:43:46.980)

it just connects and inspires.

Russ Tedrake (2:43:49.300)

The same, there's a few technical books too

Lex Fridman (2:43:53.780)

and algorithms that just, yeah, you return to often.

Russ Tedrake (2:43:58.020)

I'm with you.

Lex Fridman (2:44:01.900)

Yeah, there's, and I've been losing that

Russ Tedrake (2:44:04.100)

because of the internet.

Lex Fridman (2:44:05.380)

I've been like going on, I've been going on archive

Lex Fridman (2:44:09.700)

and blog posts and GitHub and the new thing

Lex Fridman (2:44:12.420)

and you lose your ability to really master an idea.

Russ Tedrake (2:44:18.100)

Right.

Lex Fridman (2:44:18.940)

Wow.

Russ Tedrake (2:44:19.780)

Exactly right.

Lex Fridman (2:44:21.100)

What's a fond memory from childhood?

Russ Tedrake (2:44:24.940)

When baby Russ Tedrick.

Lex Fridman (2:44:29.540)

Well, I guess I just said that at least my current life

Russ Tedrake (2:44:33.940)

began when I got to MIT.

Lex Fridman (2:44:36.780)

If I have to go farther than that.

Lex Fridman (2:44:38.900)

Yeah, what was, was there a life before MIT?

Lex Fridman (2:44:42.260)

Oh, absolutely, but let me actually tell you

Lex Fridman (2:44:47.380)

what happened when I first got to MIT

Lex Fridman (2:44:48.900)

because that I think might be relevant here,

Lex Fridman (2:44:52.220)

but I had taken a computer engineering degree at Michigan.

Lex Fridman (2:44:57.540)

I enjoyed it immensely, learned a bunch of stuff.

Russ Tedrake (2:45:00.420)

I liked computers, I liked programming,

Lex Fridman (2:45:04.580)

but when I did get to MIT and started working

Russ Tedrake (2:45:07.340)

with Sebastian Sung, theoretical physicist,

Lex Fridman (2:45:10.300)

computational neuroscientist, the culture here

Russ Tedrake (2:45:15.180)

was just different.

Lex Fridman (2:45:17.220)

It demanded more of me, certainly mathematically

Lex Fridman (2:45:20.260)

and in the critical thinking.

Lex Fridman (2:45:22.660)

And I remember the day that I borrowed one of the books

Russ Tedrake (2:45:27.700)

from my advisor's office and walked down

Lex Fridman (2:45:29.780)

to the Charles River and was like,

Russ Tedrake (2:45:32.140)

I'm getting my butt kicked.

Lex Fridman (2:45:36.620)

And I think that's gonna happen to everybody

Russ Tedrake (2:45:38.180)

who's doing this kind of stuff.

Lex Fridman (2:45:40.220)

I think I expected you to ask me the meaning of life.

Russ Tedrake (2:45:46.020)

I think that somehow I think that's gotta be part of it.

Lex Fridman (2:45:52.780)

Doing hard things?

Russ Tedrake (2:45:55.140)

Yeah.

Lex Fridman (2:45:56.460)

Did you consider quitting at any point?

Lex Fridman (2:45:58.220)

Did you consider this isn't for me?

Lex Fridman (2:45:59.780)

No, never that.

Russ Tedrake (2:46:01.740)

I was working hard, but I was loving it.

Lex Fridman (2:46:07.180)

I think there's this magical thing

Russ Tedrake (2:46:08.860)

where I'm lucky to surround myself with people

Lex Fridman (2:46:11.900)

that basically almost every day I'll see something,

Russ Tedrake (2:46:17.900)

I'll be told something or something that I realize,

Lex Fridman (2:46:20.340)

wow, I don't understand that.

Lex Fridman (2:46:22.020)

And if I could just understand that,

Lex Fridman (2:46:24.180)

there's something else to learn.

Russ Tedrake (2:46:26.020)

That if I could just learn that thing,

Lex Fridman (2:46:28.140)

I would connect another piece of the puzzle.

Lex Fridman (2:46:30.220)

And I think that is just such an important aspect

Lex Fridman (2:46:36.220)

and being willing to understand what you can and can't do

Lex Fridman (2:46:40.260)

and loving the journey of going

Lex Fridman (2:46:43.580)

and learning those other things.

Russ Tedrake (2:46:44.820)

I think that's the best part.

Lex Fridman (2:46:47.340)

I don't think there's a better way to end it, Russ.

Russ Tedrake (2:46:51.500)

You've been an inspiration to me since I showed up at MIT.

Lex Fridman (2:46:55.580)

Your work has been an inspiration to the world.

Russ Tedrake (2:46:57.700)

This conversation was amazing.

Lex Fridman (2:46:59.740)

I can't wait to see what you do next

Russ Tedrake (2:47:01.700)

with robotics, home robots.

Lex Fridman (2:47:03.220)

I hope to see you work in my home one day.

Lex Fridman (2:47:05.780)

So thanks so much for talking today, it's been awesome.

Lex Fridman (2:47:08.100)

Cheers.

Russ Tedrake (2:47:09.480)

Thanks for listening to this conversation

Lex Fridman (2:47:11.060)

with Russ Tedrick and thank you to our sponsors,

Russ Tedrake (2:47:14.180)

Magic Spoon Cereal, BetterHelp and ExpressVPN.

Lex Fridman (2:47:18.220)

Please consider supporting this podcast

Russ Tedrake (2:47:20.180)

by going to magicspoon.com slash Lex

Lex Fridman (2:47:23.420)

and using code Lex at checkout.

Russ Tedrake (2:47:25.500)

Go into betterhelp.com slash Lex

Lex Fridman (2:47:27.780)

and signing up at expressvpn.com slash Lex pod.

Russ Tedrake (2:47:32.820)

Click the links, buy the stuff, get the discount.

Lex Fridman (2:47:36.180)

It really is the best way to support this podcast.

Russ Tedrake (2:47:39.380)

If you enjoy this thing, subscribe on YouTube,

Lex Fridman (2:47:41.520)

review it with five stars and up a podcast,

Russ Tedrake (2:47:43.700)

support on Patreon or connect with me on Twitter

Lex Fridman (2:47:46.540)

at Lex Friedman spelled somehow without the E

Russ Tedrake (2:47:50.620)

just F R I D M A N.

Lex Fridman (2:47:53.460)

And now let me leave you with some words

Russ Tedrake (2:47:55.100)

from Neil deGrasse Tyson talking about robots in space

Lex Fridman (2:47:58.540)

and the emphasis we humans put

Russ Tedrake (2:48:00.680)

on human based space exploration.

Lex Fridman (2:48:03.640)

Robots are important.

Russ Tedrake (2:48:05.680)

If I don my pure scientist hat,

Lex Fridman (2:48:07.980)

I would say just send robots.

Russ Tedrake (2:48:10.020)

I'll stay down here and get the data.

Lex Fridman (2:48:12.340)

But nobody's ever given a parade for a robot.

Russ Tedrake (2:48:15.080)

Nobody's ever named a high school after a robot.

Lex Fridman (2:48:17.940)

So when I don my public educator hat,

Russ Tedrake (2:48:20.180)

I have to recognize the elements of exploration

Lex Fridman (2:48:22.780)

that excite people.

Russ Tedrake (2:48:24.180)

It's not only the discoveries and the beautiful photos

Lex Fridman (2:48:26.980)

that come down from the heavens.

Russ Tedrake (2:48:29.020)

It's the vicarious participation in discovery itself.

Lex Fridman (2:48:33.020)

Thank you for listening and hope to see you next time.

Lex Fridman (30:02.200)

what do you do with that tension

Lex Fridman (30:03.580)

between the real world saying urgently,

Russ Tedrake (30:05.620)

you need to be there, this is important,

Lex Fridman (30:08.220)

everything is melting down,

Lex Fridman (30:10.060)

how are we gonna fix this robot?

Lex Fridman (30:11.820)

There's this critical meeting,

Lex Fridman (30:14.660)

and then there's this, the zen beauty of just running,

Lex Fridman (30:18.020)

the simplicity of it, you along with nature.

Lex Fridman (30:21.380)

What do you do with that?

Lex Fridman (30:22.700)

I would say I'm not a fast runner, particularly.

Russ Tedrake (30:25.540)

Probably my fastest splits ever was when

Lex Fridman (30:27.940)

I had to get to daycare on time

Russ Tedrake (30:29.220)

because they were gonna charge me, you know,

Lex Fridman (30:30.700)

some dollar per minute that I was late.

Russ Tedrake (30:33.540)

I've run some fast splits to daycare.

Lex Fridman (30:36.980)

But those times are past now.

Russ Tedrake (30:41.700)

I think work, you can find a work life balance in that way.

Lex Fridman (30:44.900)

I think you just have to.

Russ Tedrake (30:47.260)

I think I am better at work

Lex Fridman (30:48.620)

because I take time to think on the way in.

Lex Fridman (30:52.180)

So I plan my day around it,

Lex Fridman (30:55.300)

and I rarely feel that those are really at odds.

Lex Fridman (31:00.300)

So what, the bucket list item.

Lex Fridman (31:03.380)

If we're talking 12 times two, or approaching a marathon,

Lex Fridman (31:10.620)

what, have you run an ultra marathon before?

Lex Fridman (31:15.060)

Do you do races?

Russ Tedrake (31:16.740)

Is there, what's a...

Lex Fridman (31:17.580)

Not to win.

Russ Tedrake (31:21.620)

I'm not gonna like take a dinghy across the Atlantic

Lex Fridman (31:23.720)

or something if that's what you want.

Lex Fridman (31:24.780)

But if someone does and wants to write a book,

Lex Fridman (31:27.920)

I would totally read it

Russ Tedrake (31:28.760)

because I'm a sucker for that kind of thing.

Lex Fridman (31:31.140)

No, I do have some fun things that I will try.

Russ Tedrake (31:33.420)

You know, I like to, when I travel,

Lex Fridman (31:35.300)

I almost always bike to Logan Airport

Lex Fridman (31:37.020)

and fold up a little folding bike

Lex Fridman (31:38.740)

and then take it with me and bike to wherever I'm going.

Lex Fridman (31:41.040)

And it's taken me,

Lex Fridman (31:42.420)

or I'll take a stand up paddle board these days

Russ Tedrake (31:44.580)

on the airplane,

Lex Fridman (31:45.500)

and then I'll try to paddle around where I'm going

Russ Tedrake (31:47.100)

or whatever.

Lex Fridman (31:47.940)

And I've done some crazy things, but...

Lex Fridman (31:50.720)

But not for the, you know, I now talk,

Lex Fridman (31:55.140)

I don't know if you know who David Goggins is by any chance.

Russ Tedrake (31:57.500)

Not well, but yeah.

Lex Fridman (31:58.460)

But I talk to him now every day.

Lex Fridman (32:00.140)

So he's the person who made me do this stupid challenge.

Lex Fridman (32:05.940)

So he's insane and he does things for the purpose

Russ Tedrake (32:10.160)

in the best kind of way.

Lex Fridman (32:11.380)

He does things like for the explicit purpose of suffering.

Russ Tedrake (32:16.980)

Like he picks the thing that,

Lex Fridman (32:18.420)

like whatever he thinks he can do, he does more.

Lex Fridman (32:22.940)

So is that, do you have that thing in you or are you...

Lex Fridman (32:27.300)

I think it's become the opposite.

Russ Tedrake (32:29.820)

It's a...

Lex Fridman (32:30.660)

So you're like that dynamical system

Russ Tedrake (32:32.300)

that the walker, the efficient...

Lex Fridman (32:34.420)

Yeah, it's leave no pain, right?

Russ Tedrake (32:38.860)

You should end feeling better than you started.

Lex Fridman (32:40.900)

Okay.

Lex Fridman (32:41.720)

But it's mostly, I think, and COVID has tested this

Lex Fridman (32:45.940)

because I've lost my commute.

Russ Tedrake (32:47.740)

I think I'm perfectly happy walking around town

Lex Fridman (32:51.980)

with my wife and kids if they could get them to go.

Lex Fridman (32:55.220)

And it's more about just getting outside

Lex Fridman (32:57.780)

and getting away from the keyboard for some time

Russ Tedrake (32:59.980)

just to let things compress.

Lex Fridman (33:02.580)

Let's go into robotics a little bit.

Lex Fridman (33:04.100)

What to use the most beautiful idea in robotics?

Lex Fridman (33:07.800)

Whether we're talking about control

Russ Tedrake (33:10.780)

or whether we're talking about optimization

Lex Fridman (33:12.740)

and the math side of things or the engineering side of things

Russ Tedrake (33:16.180)

or the philosophical side of things.

Lex Fridman (33:20.380)

I think I've been lucky to experience something

Russ Tedrake (33:23.540)

that not so many roboticists have experienced,

Lex Fridman (33:27.700)

which is to hang out

Russ Tedrake (33:30.220)

with some really amazing control theorists.

Lex Fridman (33:34.420)

And the clarity of thought

Russ Tedrake (33:40.700)

that some of the more mathematical control theory

Lex Fridman (33:43.140)

can bring to even very complex, messy looking problems

Russ Tedrake (33:49.480)

is really, it really had a big impact on me

Lex Fridman (33:53.140)

and I had a day even just a couple of weeks ago

Russ Tedrake (33:57.900)

where I had spent the day on a Zoom robotics conference

Lex Fridman (34:01.020)

having great conversations with lots of people.

Russ Tedrake (34:04.020)

Felt really good about the ideas

Lex Fridman (34:06.780)

that were flowing and the like.

Lex Fridman (34:09.500)

And then I had a late afternoon meeting

Lex Fridman (34:12.940)

with one of my favorite control theorists

Lex Fridman (34:15.540)

and we went from these abstract discussions

Lex Fridman (34:20.540)

about maybes and what ifs and what a great idea

Russ Tedrake (34:25.540)

to these super precise statements

Lex Fridman (34:30.100)

about systems that aren't that much more simple

Russ Tedrake (34:33.660)

or abstract than the ones I care about deeply.

Lex Fridman (34:38.260)

And the contrast of that is,

Russ Tedrake (34:42.540)

I don't know, it really gets me.

Lex Fridman (34:43.780)

I think people underestimate

Russ Tedrake (34:47.580)

maybe the power of clear thinking.

Lex Fridman (34:51.580)

And so for instance, deep learning is amazing.

Russ Tedrake (34:58.580)

I use it heavily in our work.

Lex Fridman (35:00.380)

I think it's changed the world, unquestionable.

Russ Tedrake (35:04.700)

It makes it easy to get things to work

Lex Fridman (35:07.020)

without thinking as critically about it.

Lex Fridman (35:08.580)

So I think one of the challenges as an educator

Lex Fridman (35:11.300)

is to think about how do we make sure people get a taste

Russ Tedrake (35:14.940)

of the more rigorous thinking

Lex Fridman (35:17.860)

that I think goes along with some different approaches.

Russ Tedrake (35:22.620)

Yeah, so that's really interesting.

Lex Fridman (35:24.020)

So understanding like the fundamentals,

Russ Tedrake (35:26.900)

the first principles of the problem,

Lex Fridman (35:31.900)

where in this case it's mechanics,

Russ Tedrake (35:33.780)

like how a thing moves, how a thing behaves,

Lex Fridman (35:38.780)

like all the forces involved,

Russ Tedrake (35:40.420)

like really getting a deep understanding of that.

Lex Fridman (35:42.740)

I mean, from physics, the first principle thing

Russ Tedrake (35:45.340)

come from physics, and here it's literally physics.

Lex Fridman (35:50.100)

Yeah, and this applies, in deep learning,

Russ Tedrake (35:51.940)

this applies to not just, I mean,

Lex Fridman (35:54.980)

it applies so cleanly in robotics,

Lex Fridman (35:57.300)

but it also applies to just in any data set.

Lex Fridman (36:01.500)

I find this true, I mean, driving as well.

Russ Tedrake (36:05.100)

There's a lot of folks in that work on autonomous vehicles

Lex Fridman (36:09.100)

that work on autonomous vehicles that don't study driving,

Russ Tedrake (36:17.900)

like deeply.

Lex Fridman (36:20.300)

I might be coming a little bit from the psychology side,

Lex Fridman (36:23.100)

but I remember I spent a ridiculous number of hours

Lex Fridman (36:28.380)

at lunch, at this like lawn chair,

Lex Fridman (36:31.940)

and I would sit somewhere in MIT's campus,

Lex Fridman (36:35.740)

there's a few interesting intersections,

Lex Fridman (36:37.260)

and we'd just watch people cross.

Lex Fridman (36:39.380)

So we were studying pedestrian behavior,

Lex Fridman (36:43.220)

and I felt like, as we record a lot of video,

Lex Fridman (36:46.220)

to try, and then there's the computer vision

Russ Tedrake (36:47.820)

extracts their movement, how they move their head, and so on,

Lex Fridman (36:50.860)

but like every time, I felt like I didn't understand enough.

Russ Tedrake (36:55.340)

I just, I felt like I wasn't understanding

Lex Fridman (36:58.620)

what, how are people signaling to each other,

Lex Fridman (37:01.620)

what are they thinking,

Lex Fridman (37:03.580)

how cognizant are they of their fear of death?

Lex Fridman (37:07.820)

Like, what's the underlying game theory here?

Lex Fridman (37:11.900)

What are the incentives?

Lex Fridman (37:14.140)

And then I finally found a live stream of an intersection

Lex Fridman (37:17.860)

that's like high def that I just, I would watch

Lex Fridman (37:20.300)

so I wouldn't have to sit out there.

Lex Fridman (37:21.780)

But it's interesting, so like, I feel.

Lex Fridman (37:23.580)

But that's tough, that's a tough example,

Lex Fridman (37:25.180)

because I mean, the learning.

Russ Tedrake (37:27.100)

Humans are involved.

Lex Fridman (37:28.780)

Not just because human, but I think the learning mantra

Russ Tedrake (37:33.460)

is that basically the statistics of the data

Lex Fridman (37:35.500)

will tell me things I need to know, right?

Russ Tedrake (37:37.940)

And, you know, for the example you gave

Lex Fridman (37:41.860)

of all the nuances of, you know, eye contact,

Russ Tedrake (37:45.420)

or hand gestures, or whatever that are happening

Lex Fridman (37:47.620)

for these subtle interactions

Lex Fridman (37:48.900)

between pedestrians and traffic, right?

Lex Fridman (37:51.140)

Maybe the data will tell that story.

Russ Tedrake (37:54.460)

I maybe even, one level more meta than what you're saying.

Lex Fridman (38:01.300)

For a particular problem,

Russ Tedrake (38:02.660)

I think it might be the case

Lex Fridman (38:03.820)

that data should tell us the story.

Lex Fridman (38:07.220)

But I think there's a rigorous thinking

Lex Fridman (38:09.420)

that is just an essential skill

Russ Tedrake (38:11.700)

for a mathematician or an engineer

Lex Fridman (38:14.580)

that I just don't wanna lose it.

Russ Tedrake (38:18.380)

There are certainly super rigorous control,

Lex Fridman (38:22.460)

or sorry, machine learning people.

Russ Tedrake (38:24.940)

I just think deep learning makes it so easy

Lex Fridman (38:28.020)

to do some things that our next generation,

Russ Tedrake (38:31.580)

are not immediately rewarded

Lex Fridman (38:35.860)

for going through some of the more rigorous approaches.

Lex Fridman (38:38.540)

And then I wonder where that takes us.

Lex Fridman (38:40.740)

Well, I'm actually optimistic about it.

Russ Tedrake (38:42.260)

I just want to do my part

Lex Fridman (38:44.860)

to try to steer that rigorous thinking.

Lex Fridman (38:48.020)

So there's like two questions I wanna ask.

Lex Fridman (38:50.940)

Do you have sort of a good example of rigorous thinking

Lex Fridman (38:56.860)

where it's easy to get lazy and not do the rigorous thinking?

Lex Fridman (39:00.860)

And the other question I have is like,

Lex Fridman (39:02.500)

do you have advice of how to practice rigorous thinking

Lex Fridman (39:09.140)

in all the computer science disciplines that we've mentioned?

Russ Tedrake (39:16.380)

Yeah, I mean, there are times where problems

Lex Fridman (39:21.500)

that can be solved with well known mature methods

Russ Tedrake (39:25.860)

could also be solved with a deep learning approach.

Lex Fridman (39:30.300)

And there's an argument that you must use learning

Russ Tedrake (39:36.740)

even for the parts we already think we know,

Lex Fridman (39:38.380)

because if the human has touched it,

Russ Tedrake (39:39.780)

then you've biased the system

Lex Fridman (39:42.460)

and you've suddenly put a bottleneck in there

Russ Tedrake (39:44.340)

that is your own mental model.

Lex Fridman (39:46.300)

But something like converting a matrix,

Russ Tedrake (39:49.100)

I think we know how to do that pretty well,

Lex Fridman (39:50.780)

even if it's a pretty big matrix,

Lex Fridman (39:52.020)

and we understand that pretty well.

Lex Fridman (39:53.140)

And you could train a deep network to do it,

Lex Fridman (39:55.060)

but you shouldn't probably.

Lex Fridman (39:57.340)

So in that sense, rigorous thinking is understanding

Russ Tedrake (40:02.220)

the scope and the limitations of the methods that we have,

Lex Fridman (40:07.340)

like how to use the tools of mathematics properly.

Russ Tedrake (40:10.180)

Yeah, I think taking a class on analysis

Lex Fridman (40:15.100)

is all I'm sort of arguing is to take a chance to stop

Lex Fridman (40:18.620)

and force yourself to think rigorously

Lex Fridman (40:20.900)

about even the rational numbers or something.

Russ Tedrake (40:25.140)

It doesn't have to be the end all problem.

Lex Fridman (40:27.740)

But that exercise of clear thinking,

Russ Tedrake (40:31.100)

I think goes a long way,

Lex Fridman (40:33.420)

and I just wanna make sure we keep preaching it.

Russ Tedrake (40:35.260)

We don't lose it.

Lex Fridman (40:36.380)

But do you think when you're doing rigorous thinking

Russ Tedrake (40:39.540)

or maybe trying to write down equations

Lex Fridman (40:43.220)

or sort of explicitly formally describe a system,

Lex Fridman (40:47.980)

do you think we naturally simplify things too much?

Lex Fridman (40:51.580)

Is that a danger you run into?

Russ Tedrake (40:53.500)

Like in order to be able to understand something

Lex Fridman (40:56.180)

about the system mathematically,

Russ Tedrake (40:58.180)

we make it too much of a toy example.

Lex Fridman (41:01.700)

But I think that's the good stuff, right?

Lex Fridman (41:04.460)

That's how you understand the fundamentals?

Lex Fridman (41:07.060)

I think so.

Russ Tedrake (41:07.900)

I think maybe even that's a key to intelligence

Lex Fridman (41:10.380)

or something, but I mean, okay,

Lex Fridman (41:12.460)

what if Newton and Galileo had deep learning?

Lex Fridman (41:15.100)

And they had done a bunch of experiments

Lex Fridman (41:18.340)

and they told the world,

Lex Fridman (41:20.360)

here's your weights of your neural network.

Russ Tedrake (41:22.460)

We've solved the problem.

Lex Fridman (41:24.260)

Where would we be today?

Russ Tedrake (41:25.380)

I don't think we'd be as far as we are.

Lex Fridman (41:28.420)

There's something to be said

Russ Tedrake (41:29.260)

about having the simplest explanation for a phenomenon.

Lex Fridman (41:32.540)

So I don't doubt that we can train neural networks

Russ Tedrake (41:37.180)

to predict even physical F equals MA type equations.

Lex Fridman (41:46.300)

But I maybe, I want another Newton to come along

Russ Tedrake (41:51.300)

because I think there's more to do

Lex Fridman (41:52.940)

in terms of coming up with the simple models

Russ Tedrake (41:56.020)

for more complicated tasks.

Lex Fridman (41:59.860)

Yeah, let's not offend AI systems from 50 years

Russ Tedrake (42:04.240)

from now that are listening to this

Lex Fridman (42:06.340)

that are probably better at,

Russ Tedrake (42:08.260)

might be better coming up

Lex Fridman (42:10.180)

with F equals MA equations themselves.

Lex Fridman (42:13.080)

So sorry, I actually think learning is probably a route

Lex Fridman (42:16.940)

to achieving this, but the representation matters, right?

Lex Fridman (42:21.180)

And I think having a function that takes my inputs

Lex Fridman (42:26.200)

to outputs that is arbitrarily complex

Russ Tedrake (42:29.060)

may not be the end goal.

Lex Fridman (42:30.780)

I think there's still the most simple

Russ Tedrake (42:34.140)

or parsimonious explanation for the data.

Lex Fridman (42:37.620)

Simple doesn't mean low dimensional.

Russ Tedrake (42:39.000)

That's one thing I think that we've,

Lex Fridman (42:41.020)

a lesson that we've learned.

Lex Fridman (42:41.960)

So a standard way to do model reduction

Lex Fridman (42:46.080)

or system identification and controls

Russ Tedrake (42:47.860)

is the typical formulation is that you try to find

Lex Fridman (42:50.460)

the minimal state dimension realization of a system

Russ Tedrake (42:54.220)

that hits some error bounds or something like that.

Lex Fridman (42:57.760)

And that's maybe not, I think we're learning

Russ Tedrake (43:00.340)

that state dimension is not the right metric.

Lex Fridman (43:05.980)

Of complexity.

Russ Tedrake (43:06.820)

Of complexity.

Lex Fridman (43:07.640)

But for me, I think a lot about contact,

Russ Tedrake (43:09.460)

the mechanics of contact,

Lex Fridman (43:10.820)

if a robot hand is picking up an object or something.

Lex Fridman (43:14.520)

And when I write down the equations of motion for that,

Lex Fridman (43:17.220)

they look incredibly complex,

Russ Tedrake (43:19.100)

not because, actually not so much

Lex Fridman (43:23.420)

because of the dynamics of the hand when it's moving,

Lex Fridman (43:26.660)

but it's just the interactions

Lex Fridman (43:28.500)

and when they turn on and off, right?

Lex Fridman (43:30.860)

So having a high dimensional,

Lex Fridman (43:33.300)

but simple description of what's happening out here is fine.

Lex Fridman (43:36.420)

But if when I actually start touching,

Lex Fridman (43:38.480)

if I write down a different dynamical system

Russ Tedrake (43:41.860)

for every polygon on my robot hand

Lex Fridman (43:45.420)

and every polygon on the object,

Russ Tedrake (43:47.300)

whether it's in contact or not,

Lex Fridman (43:49.000)

with all the combinatorics that explodes there,

Russ Tedrake (43:51.700)

then that's too complex.

Lex Fridman (43:54.460)

So I need to somehow summarize that

Russ Tedrake (43:55.800)

with a more intuitive physics way of thinking.

Lex Fridman (44:01.460)

And yeah, I'm very optimistic

Russ Tedrake (44:03.500)

that machine learning will get us there.

Lex Fridman (44:05.700)

First of all, I mean, I'll probably do it

Russ Tedrake (44:08.220)

in the introduction,

Lex Fridman (44:09.140)

but you're one of the great robotics people at MIT.

Russ Tedrake (44:12.900)

You're a professor at MIT.

Lex Fridman (44:14.300)

You've teach him a lot of amazing courses.

Russ Tedrake (44:16.480)

You run a large group

Lex Fridman (44:19.180)

and you have a important history for MIT, I think,

Russ Tedrake (44:22.780)

as being a part of the DARPA Robotics Challenge.

Lex Fridman (44:26.340)

Can you maybe first say,

Lex Fridman (44:28.340)

what is the DARPA Robotics Challenge

Lex Fridman (44:30.000)

and then tell your story around it, your journey with it?

Russ Tedrake (44:36.380)

Yeah, sure.

Lex Fridman (44:39.260)

So the DARPA Robotics Challenge,

Russ Tedrake (44:41.060)

it came on the tails of the DARPA Grand Challenge

Lex Fridman (44:44.720)

and DARPA Urban Challenge,

Russ Tedrake (44:45.940)

which were the challenges that brought us,

Lex Fridman (44:49.660)

put a spotlight on self driving cars.

Russ Tedrake (44:55.400)

Gil Pratt was at DARPA and pitched a new challenge

Lex Fridman (45:01.360)

that involved disaster response.

Russ Tedrake (45:04.980)

It didn't explicitly require humanoids,

Lex Fridman (45:07.140)

although humanoids came into the picture.

Russ Tedrake (45:10.220)

This happened shortly after the Fukushima disaster in Japan

Lex Fridman (45:14.740)

and our challenge was motivated roughly by that

Russ Tedrake (45:17.660)

because that was a case where if we had had robots

Lex Fridman (45:21.060)

that were ready to be sent in,

Russ Tedrake (45:22.700)

there's a chance that we could have averted disaster.

Lex Fridman (45:26.580)

And certainly after the, in the disaster response,

Russ Tedrake (45:30.620)

there were times we would have loved

Lex Fridman (45:32.380)

to have sent robots in.

Lex Fridman (45:34.740)

So in practice, what we ended up with was a grand challenge,

Lex Fridman (45:39.220)

a DARPA Robotics Challenge,

Russ Tedrake (45:41.180)

where Boston Dynamics was to make humanoid robots.

Lex Fridman (45:48.660)

People like me and the amazing team at MIT

Russ Tedrake (45:53.660)

were competing first in a simulation challenge

Lex Fridman (45:56.780)

to try to be one of the ones that wins the right

Russ Tedrake (45:59.460)

to work on one of the Boston Dynamics humanoids

Lex Fridman (46:03.340)

in order to compete in the final challenge,

Russ Tedrake (46:06.620)

which was a physical challenge.

Lex Fridman (46:08.580)

And at that point, it was already, so it was decided

Russ Tedrake (46:11.260)

as humanoid robots early on.

Lex Fridman (46:13.420)

There were two tracks.

Russ Tedrake (46:15.140)

You could enter as a hardware team

Lex Fridman (46:16.900)

where you brought your own robot,

Russ Tedrake (46:18.480)

or you could enter through the virtual robotics challenge

Lex Fridman (46:21.380)

as a software team that would try to win the right

Russ Tedrake (46:24.300)

to use one of the Boston Dynamics robots.

Lex Fridman (46:25.940)

Sure, called Atlas.

Russ Tedrake (46:27.420)

Atlas.

Lex Fridman (46:28.260)

Humanoid robots.

Russ Tedrake (46:29.080)

Yeah, it was a 400 pound Marvel,

Lex Fridman (46:31.500)

but a pretty big, scary looking robot.

Russ Tedrake (46:35.620)

Expensive too.

Lex Fridman (46:36.700)

Expensive, yeah.

Russ Tedrake (46:38.260)

Okay, so I mean, how did you feel

Lex Fridman (46:42.300)

at the prospect of this kind of challenge?

Russ Tedrake (46:44.780)

I mean, it seems autonomous vehicles,

Lex Fridman (46:48.820)

yeah, I guess that sounds hard,

Lex Fridman (46:51.060)

but not really from a robotics perspective.

Lex Fridman (46:53.980)

It's like, didn't they do it in the 80s

Russ Tedrake (46:56.020)

is the kind of feeling I would have,

Lex Fridman (46:58.760)

like when you first look at the problem,

Russ Tedrake (47:00.820)

it's on wheels, but like humanoid robots,

Lex Fridman (47:04.900)

that sounds really hard.

Lex Fridman (47:07.060)

So what are your, psychologically speaking,

Lex Fridman (47:12.860)

what were you feeling, excited, scared?

Lex Fridman (47:15.780)

Why the heck did you get yourself involved

Lex Fridman (47:18.020)

in this kind of messy challenge?

Russ Tedrake (47:19.660)

We didn't really know for sure what we were signing up for

Lex Fridman (47:24.540)

in the sense that you could have something that,

Russ Tedrake (47:26.820)

as it was described in the call for participation,

Lex Fridman (47:30.780)

that could have put a huge emphasis on the dynamics

Russ Tedrake (47:33.900)

of walking and not falling down

Lex Fridman (47:35.700)

and walking over rough terrain,

Russ Tedrake (47:37.380)

or the same description,

Lex Fridman (47:38.580)

because the robot had to go into this disaster area

Lex Fridman (47:40.780)

and turn valves and pick up a drill,

Lex Fridman (47:44.580)

it cut the hole through a wall,

Russ Tedrake (47:45.780)

it had to do some interesting things.

Lex Fridman (47:48.420)

The challenge could have really highlighted perception

Lex Fridman (47:51.860)

and autonomous planning,

Lex Fridman (47:54.820)

or it ended up that locomoting over complex terrain

Russ Tedrake (48:01.060)

played a pretty big role in the competition.

Lex Fridman (48:03.600)

So...

Lex Fridman (48:05.520)

And the degree of autonomy wasn't clear.

Lex Fridman (48:08.360)

The degree of autonomy

Russ Tedrake (48:09.560)

was always a central part of the discussion.

Lex Fridman (48:11.920)

So what wasn't clear was how we would be able,

Lex Fridman (48:15.560)

how far we'd be able to get with it.

Lex Fridman (48:17.520)

So the idea was always that you want semi autonomy,

Russ Tedrake (48:21.640)

that you want the robot to have enough compute

Lex Fridman (48:24.280)

that you can have a degraded network link to a human.

Lex Fridman (48:27.640)

And so the same way we had degraded networks

Lex Fridman (48:30.640)

at many natural disasters,

Russ Tedrake (48:33.160)

you'd send your robot in,

Lex Fridman (48:34.960)

you'd be able to get a few bits back and forth,

Lex Fridman (48:37.540)

but you don't get to have enough

Lex Fridman (48:38.920)

potentially to fully operate the robot

Russ Tedrake (48:42.080)

in every joint of the robot.

Lex Fridman (48:44.600)

So, and then the question was,

Lex Fridman (48:46.160)

and the gamesmanship of the organizers

Lex Fridman (48:48.880)

was to figure out what we're capable of,

Russ Tedrake (48:50.680)

push us as far as we could,

Lex Fridman (48:52.600)

so that it would differentiate the teams

Russ Tedrake (48:55.300)

that put more autonomy on the robot

Lex Fridman (48:57.540)

and had a few clicks and just said,

Russ Tedrake (48:59.400)

go there, do this, go there, do this,

Lex Fridman (49:00.920)

versus someone who's picking every footstep

Russ Tedrake (49:03.400)

or something like that.

Lex Fridman (49:05.280)

So what were some memories,

Lex Fridman (49:10.760)

painful, triumphant from the experience?

Lex Fridman (49:13.620)

Like what was that journey?

Russ Tedrake (49:15.040)

Maybe if you can dig in a little deeper,

Lex Fridman (49:17.680)

maybe even on the technical side, on the team side,

Russ Tedrake (49:21.120)

that whole process of,

Lex Fridman (49:24.120)

from the early idea stages to actually competing.

Russ Tedrake (49:28.200)

I mean, this was a defining experience for me.

Lex Fridman (49:31.680)

It came at the right time for me in my career.

Russ Tedrake (49:33.940)

I had gotten tenure before I was due a sabbatical,

Lex Fridman (49:37.480)

and most people do something relaxing

Lex Fridman (49:39.840)

and restorative for a sabbatical.

Lex Fridman (49:41.920)

So you got tenure before this?

Russ Tedrake (49:44.520)

Yeah, yeah, yeah.

Lex Fridman (49:46.200)

It was a good time for me.

Russ Tedrake (49:48.120)

We had a bunch of algorithms that we were very happy with.

Lex Fridman (49:50.960)

We wanted to see how far we could push them,

Lex Fridman (49:52.560)

and this was a chance to really test our mettle

Lex Fridman (49:54.920)

to do more proper software engineering.

Lex Fridman (49:56.880)

So the team, we all just worked our butts off.

Lex Fridman (50:01.420)

We were in that lab almost all the time.

Russ Tedrake (50:07.680)

Okay, so there were some, of course,

Lex Fridman (50:09.600)

high highs and low lows throughout that.

Russ Tedrake (50:12.080)

Anytime you're not sleeping

Lex Fridman (50:13.720)

and devoting your life to a 400 pound humanoid.

Russ Tedrake (50:18.320)

I remember actually one funny moment

Lex Fridman (50:20.720)

where we're all super tired,

Lex Fridman (50:21.940)

and so Atlas had to walk across cinder blocks.

Lex Fridman (50:24.760)

That was one of the obstacles.

Lex Fridman (50:26.520)

And I remember Atlas was powered down

Lex Fridman (50:28.240)

and hanging limp on its harness,

Lex Fridman (50:31.280)

and the humans were there picking up

Lex Fridman (50:34.000)

and laying the brick down

Lex Fridman (50:35.200)

so that the robot could walk over it.

Lex Fridman (50:36.440)

And I thought, what is wrong with this?

Russ Tedrake (50:38.240)

We've got a robot just watching us

Lex Fridman (50:41.560)

do all the manual labor

Lex Fridman (50:42.500)

so that it can take its little stroll across the train.

Lex Fridman (50:47.040)

But I mean, even the virtual robotics challenge

Russ Tedrake (50:52.120)

was super nerve wracking and dramatic.

Lex Fridman (50:54.640)

I remember, so we were using Gazebo as a simulator

Russ Tedrake (51:01.520)

on the cloud,

Lex Fridman (51:02.360)

and there was all these interesting challenges.

Russ Tedrake (51:03.920)

I think the investment that OSR FC,

Lex Fridman (51:08.560)

whatever they were called at that time,

Russ Tedrake (51:10.020)

Brian Gerkey's team at Open Source Robotics,

Lex Fridman (51:14.160)

they were pushing on the capabilities of Gazebo

Russ Tedrake (51:16.000)

in order to scale it to the complexity of these challenges.

Lex Fridman (51:20.380)

So, you know, up to the virtual competition.

Lex Fridman (51:23.900)

So the virtual competition was,

Lex Fridman (51:26.220)

you will sign on at a certain time

Lex Fridman (51:28.480)

and we'll have a network connection

Lex Fridman (51:29.840)

to another machine on the cloud

Russ Tedrake (51:32.080)

that is running the simulator of your robot.

Lex Fridman (51:34.880)

And your controller will run on this computer

Lex Fridman (51:38.160)

and the physics will run on the other

Lex Fridman (51:40.920)

and you have to connect.

Russ Tedrake (51:43.060)

Now, the physics, they wanted it to run at real time rates

Lex Fridman (51:48.140)

because there was an element of human interaction.

Lex Fridman (51:50.740)

And humans, if you do want to teleop,

Lex Fridman (51:53.280)

it works way better if it's at frame rate.

Russ Tedrake (51:56.120)

Oh, cool.

Lex Fridman (51:57.120)

But it was very hard to simulate

Russ Tedrake (51:58.720)

these complex scenes at real time rate.

Lex Fridman (52:03.240)

So right up to like days before the competition,

Russ Tedrake (52:06.520)

the simulator wasn't quite at real time rate.

Lex Fridman (52:11.040)

And that was great for me because my controller

Russ Tedrake (52:13.280)

was solving a pretty big optimization problem

Lex Fridman (52:16.280)

and it wasn't quite at real time rate.

Lex Fridman (52:17.760)

So I was fine.

Lex Fridman (52:18.880)

I was keeping up with the simulator.

Russ Tedrake (52:20.480)

We were both running at about 0.7.

Lex Fridman (52:22.880)

And I remember getting this email.

Lex Fridman (52:24.960)

And by the way, the perception folks on our team hated

Lex Fridman (52:28.440)

that they knew that if my controller was too slow,

Russ Tedrake (52:31.440)

the robot was gonna fall down.

Lex Fridman (52:32.520)

And no matter how good their perception system was,

Russ Tedrake (52:34.920)

if I can't make my controller fast.

Lex Fridman (52:36.940)

Anyways, we get this email

Russ Tedrake (52:37.920)

like three days before the virtual competition.

Lex Fridman (52:40.480)

It's for all the marbles.

Russ Tedrake (52:41.480)

We're gonna either get a humanoid robot or we're not.

Lex Fridman (52:44.920)

And we get an email saying,

Russ Tedrake (52:45.740)

good news, we made the robot, the simulator faster.

Lex Fridman (52:48.680)

It's now at one point.

Lex Fridman (52:50.560)

And I was just like, oh man, what are we gonna do here?

Lex Fridman (52:54.800)

So that came in late at night for me.

Russ Tedrake (52:59.520)

A few days ahead.

Lex Fridman (53:00.560)

A few days ahead.

Russ Tedrake (53:01.440)

I went over, it happened at Frank Permenter,

Lex Fridman (53:04.000)

who's a very, very sharp.

Russ Tedrake (53:06.800)

He was a student at the time working on optimization.

Lex Fridman (53:11.160)

He was still in lab.

Russ Tedrake (53:13.640)

Frank, we need to make the quadratic programming solver

Lex Fridman (53:16.680)

faster, not like a little faster.

Russ Tedrake (53:18.360)

It's actually, you know, and we wrote a new solver

Lex Fridman (53:22.600)

for that QP together that night.

Russ Tedrake (53:28.160)

It was terrifying.

Lex Fridman (53:29.400)

So there's a really hard optimization problem

Russ Tedrake (53:31.920)

that you're constantly solving.

Lex Fridman (53:34.480)

You didn't make the optimization problem simpler?

Lex Fridman (53:36.820)

You wrote a new solver?

Lex Fridman (53:38.480)

So, I mean, your observation is almost spot on.

Lex Fridman (53:42.840)

What we did was what everybody,

Lex Fridman (53:44.520)

I mean, people know how to do this,

Lex Fridman (53:45.800)

but we had not yet done this idea of warm starting.

Lex Fridman (53:49.240)

So we are solving a big optimization problem

Russ Tedrake (53:51.320)

at every time step.

Lex Fridman (53:52.680)

But if you're running fast enough,

Russ Tedrake (53:54.280)

the optimization problem you're solving

Lex Fridman (53:55.680)

on the last time step is pretty similar

Russ Tedrake (53:57.920)

to the optimization you're gonna solve with the next.

Lex Fridman (54:00.040)

We had course had told our commercial solver

Russ Tedrake (54:02.240)

to use warm starting, but even the interface

Lex Fridman (54:05.520)

to that commercial solver was causing us these delays.

Lex Fridman (54:09.840)

So what we did was we basically wrote,

Lex Fridman (54:12.740)

we called it fast QP at the time.

Russ Tedrake (54:15.360)

We wrote a very lightweight, very fast layer,

Lex Fridman (54:18.480)

which would basically check if nearby solutions

Russ Tedrake (54:22.120)

to the quadratic program were,

Lex Fridman (54:24.240)

which were very easily checked,

Russ Tedrake (54:26.560)

could stabilize the robot.

Lex Fridman (54:28.000)

And if they couldn't, we would fall back to the solver.

Lex Fridman (54:30.720)

You couldn't really test this well, right?

Lex Fridman (54:33.120)

Or like?

Russ Tedrake (54:33.960)

I mean, so we always knew that if we fell back to,

Lex Fridman (54:37.360)

if we, it got to the point where if for some reason

Russ Tedrake (54:40.440)

things slowed down and we fell back to the original solver,

Lex Fridman (54:42.840)

the robot would actually literally fall down.

Lex Fridman (54:46.040)

So it was a harrowing sort of edge we were,

Lex Fridman (54:49.360)

ledge we were sort of on.

Lex Fridman (54:51.200)

But I mean, it actually,

Lex Fridman (54:53.200)

like the 400 pound human could come crashing to the ground

Russ Tedrake (54:55.840)

if your solver's not fast enough.

Lex Fridman (54:58.880)

But you know, we had lots of good experiences.

Lex Fridman (55:01.900)

So can I ask you a weird question I get

Lex Fridman (55:06.640)

about idea of hard work?

Lex Fridman (55:09.440)

So actually people, like students of yours

Lex Fridman (55:14.320)

that I've interacted with and just,

Lex Fridman (55:17.040)

and robotics people in general,

Lex Fridman (55:19.400)

but they have moments,

Russ Tedrake (55:23.400)

at moments have worked harder than most people I know

Lex Fridman (55:28.360)

in terms of, if you look at different disciplines

Russ Tedrake (55:30.600)

of how hard people work.

Lex Fridman (55:32.360)

But they're also like the happiest.

Russ Tedrake (55:34.560)

Like, just like, I don't know.

Lex Fridman (55:37.000)

It's the same thing with like running.

Russ Tedrake (55:39.200)

People that push themselves to like the limit,

Lex Fridman (55:41.380)

they also seem to be like the most like full of life

Russ Tedrake (55:44.760)

somehow.

Lex Fridman (55:46.720)

And I get often criticized like,

Russ Tedrake (55:48.680)

you're not getting enough sleep.

Lex Fridman (55:50.420)

What are you doing to your body?

Russ Tedrake (55:52.000)

Blah, blah, blah, like this kind of stuff.

Lex Fridman (55:54.680)

And I usually just kind of respond like,

Russ Tedrake (55:58.040)

I'm doing what I love.

Lex Fridman (55:59.720)

I'm passionate about it.

Russ Tedrake (56:00.920)

I love it.

Lex Fridman (56:01.760)

I feel like it's, it's invigorating.

Russ Tedrake (56:04.800)

I actually think, I don't think the lack of sleep

Lex Fridman (56:07.640)

is what hurts you.

Russ Tedrake (56:08.860)

I think what hurts you is stress and lack of doing things

Lex Fridman (56:12.040)

that you're passionate about.

Lex Fridman (56:13.280)

But in this world, yeah, I mean,

Lex Fridman (56:14.920)

can you comment about why the heck robotics people

Lex Fridman (56:20.720)

are willing to push themselves to that degree?

Lex Fridman (56:26.200)

Is there value in that?

Lex Fridman (56:27.680)

And why are they so happy?

Lex Fridman (56:30.360)

I think, I think you got it right.

Russ Tedrake (56:31.920)

I mean, I think the causality is not that we work hard.

Lex Fridman (56:36.440)

And I think other disciplines work very hard too,

Lex Fridman (56:38.500)

but it's, I don't think it's that we work hard

Lex Fridman (56:40.300)

and therefore we are happy.

Russ Tedrake (56:43.160)

I think we found something

Lex Fridman (56:44.700)

that we're truly passionate about.

Russ Tedrake (56:48.080)

It makes us very happy.

Lex Fridman (56:49.960)

And then we get a little involved with it

Lex Fridman (56:52.280)

and spend a lot of time on it.

Lex Fridman (56:54.600)

What a luxury to have something

Lex Fridman (56:55.980)

that you wanna spend all your time on, right?

Lex Fridman (56:59.140)

We could talk about this for many hours,

Lex Fridman (57:00.800)

but maybe if we could pick,

Lex Fridman (57:03.880)

is there something on the technical side

Russ Tedrake (57:05.480)

on the approach that you took that's interesting

Lex Fridman (57:08.260)

that turned out to be a terrible failure

Russ Tedrake (57:10.240)

or a success that you carry into your work today

Lex Fridman (57:13.800)

about all the different ideas that were involved

Russ Tedrake (57:17.260)

in making, whether in the simulation or in the real world,

Lex Fridman (57:23.400)

making this semi autonomous system work?

Russ Tedrake (57:25.520)

I mean, it really did teach me something fundamental

Lex Fridman (57:30.880)

about what it's gonna take to get robustness

Russ Tedrake (57:33.560)

out of a system of this complexity.

Lex Fridman (57:35.320)

I would say the DARPA challenge

Russ Tedrake (57:37.720)

really was foundational in my thinking.

Lex Fridman (57:41.040)

I think the autonomous driving community thinks about this.

Russ Tedrake (57:43.720)

I think lots of people thinking

Lex Fridman (57:45.580)

about safety critical systems

Russ Tedrake (57:47.080)

that might have machine learning in the loop

Lex Fridman (57:48.920)

are thinking about these questions.

Russ Tedrake (57:50.360)

For me, the DARPA challenge was the moment

Lex Fridman (57:53.340)

where I realized we've spent every waking minute

Russ Tedrake (57:57.480)

running this robot.

Lex Fridman (57:58.920)

And again, for the physical competition,

Russ Tedrake (58:01.440)

days before the competition,

Lex Fridman (58:02.540)

we saw the robot fall down in a way

Russ Tedrake (58:04.440)

it had never fallen down before.

Lex Fridman (58:05.980)

I thought, how could we have found that?

Russ Tedrake (58:10.520)

We only have one robot, it's running almost all the time.

Lex Fridman (58:13.600)

We just didn't have enough hours in the day

Russ Tedrake (58:15.560)

to test that robot.

Lex Fridman (58:17.120)

Something has to change, right?

Lex Fridman (58:19.380)

And then I think that, I mean,

Lex Fridman (58:21.080)

I would say that the team that won was,

Russ Tedrake (58:24.880)

from KAIST, was the team that had two robots

Lex Fridman (58:28.020)

and was able to do not only incredible engineering,

Russ Tedrake (58:30.560)

just absolutely top rate engineering,

Lex Fridman (58:33.240)

but also they were able to test at a rate

Lex Fridman (58:36.080)

and discipline that we didn't keep up with.

Lex Fridman (58:39.600)

What does testing look like?

Lex Fridman (58:41.120)

What are we talking about here?

Lex Fridman (58:42.280)

Like, what's a loop of tests?

Lex Fridman (58:45.000)

Like from start to finish, what is a loop of testing?

Lex Fridman (58:48.800)

Yeah, I mean, I think there's a whole philosophy to testing.

Russ Tedrake (58:51.880)

There's the unit tests, and you can do that on a hardware,

Lex Fridman (58:54.440)

you can do that in a small piece of code.

Russ Tedrake (58:56.360)

You write one function, you should write a test

Lex Fridman (58:58.280)

that checks that function's input and outputs.

Russ Tedrake (59:00.620)

You should also write an integration test

Lex Fridman (59:02.440)

at the other extreme of running the whole system together,

Russ Tedrake (59:05.320)

where they try to turn on all of the different functions

Lex Fridman (59:09.120)

that you think are correct.

Russ Tedrake (59:11.560)

It's much harder to write the specifications

Lex Fridman (59:13.400)

for a system level test,

Russ Tedrake (59:14.520)

especially if that system is as complicated

Lex Fridman (59:17.360)

as a humanoid robot.

Lex Fridman (59:18.460)

But the philosophy is sort of the same.

Lex Fridman (59:21.040)

On the real robot, it's no different,

Lex Fridman (59:24.160)

but on a real robot,

Lex Fridman (59:26.040)

it's impossible to run the same experiment twice.

Lex Fridman (59:28.640)

So if you see a failure,

Lex Fridman (59:32.480)

you hope you caught something in the logs

Russ Tedrake (59:34.380)

that tell you what happened,

Lex Fridman (59:35.620)

but you'd probably never be able to run

Russ Tedrake (59:36.920)

exactly that experiment again.

Lex Fridman (59:39.400)

And right now, I think our philosophy is just,

Russ Tedrake (59:45.720)

basically Monte Carlo estimation,

Lex Fridman (59:47.880)

is just run as many experiments as we can,

Russ Tedrake (59:50.880)

maybe try to set up the environment

Lex Fridman (59:53.080)

to make the things we are worried about happen

Russ Tedrake (59:58.120)

as often as possible.

Lex Fridman (59:59.880)

But really we're relying on somewhat random search