Caltech Professor Answers Robotics Questions

I'm Aaron Ames.

I'm a professor of mechanical engineering.

I'm here today to answer your questions from the internet.

This is Robotics Support.

[upbeat music]

Crudbones says, I don't understand the benefits

of these food delivery robots.

You know, this is a job that could be automated

so let's have robots automate it.

They obviously work in some scenarios

but they're not super robust.

I think the bigger thing they're trying to solve

is a proof-of-concept demonstration

to see if we can solve things

like automated package delivery,

last-mile delivery, things like that.

A lot of these are first steps

towards trying to solve this more general delivery problem,

which is actually a really hard problem.

And by the way, putting these robots on the road

has taught us a lot about both what works

and also a lot of what doesn't.

If you've seen these delivery robots, they get stuck a lot.

They can hit things, they can fall off the road.

It does tell you how hard robots is.

And I think that's sort of something to keep in mind

when we're talking about humanoids

and all these other things is look at the robots

that are actually in the field today

and how they sort of still kind of fail quite a bit.

And it tells you we got a lot of work to do.

And by doing that, at least on the robotics side,

we will learn a lot.

From Paradise Nights,

Wtf's with dancing robots?

I thought they were supposed to be our slaves,

before they enslave us, least for a bit.

They're just partying.

It's pretty good.

Yeah, WTF on multiple levels.

First and foremost, why are they having robots dance?

First, there's amazing progress

that's happened in humanoids,

among other things, backflipping, running.

I mean really the behaviors we've achieved

is in the last year or two has been absolutely remarkable.

We figured out a very nice pipeline

in which to get robots to do this.

And the way you do it is you start

with a human doing those actions.

The reason why the dancing looks so human-like

is it's a human dancing.

It's really just puppetry,

very beautiful and advanced engineering puppetry

but puppetry nonetheless.

So a human puts on a mocap suit or you use cameras

and the human dances, not me 'cause I'm a very bad dancer,

but, you know, a good dancing human.

And then you take that data

and you get the trajectories from that data,

that is the motion of the human over time.

And then you train a reinforcement learning algorithm

on the humanoid robot that basically mimics

or copies that human data as much as possible

with the morphology of the robot.

And the end result is it dances like the person

that was dancing,

and as long as everything

is just the way you expected it to be,

meaning the environments like it was

when the human did the thing,

we can get robots to do that now.

But you asked about this,

aren't they supposed to be helping us

sort of before they party?

And the answer to that is that we still don't know

how to solve that problem.

That's sort of the not-very-secret secret

is that all of the hard problems of getting robots

to be truly autonomous and in our homes

are still hard problems that are unsolved.

The next question is from Projectguy111,

What percentage chance do you think we'll end up

in a Terminator future?

So I think there's two answers to this.

One is, what's the chance that AI in learning

will do bad things?

And I think that probability's actually fairly high

if we're not careful, and now's the time to be careful.

If you trust AI to be the decision maker,

if you're not very careful about having guardrails

for that AI, it will make bad decisions.

I mean, if you've ever used ChatGPT or LLMs,

you see that it can produce really nice answers sometimes,

and it's impressive, but then sometimes it just is wrong.

So we cannot trust AI.

In my opinion, you can never trust AI,

but you can use AI as a powerful tool.

Just like if you search something online on Google,

you get a lot of results back,

gives you a lot of information,

but you have to verify and double check.

So I think if we put, like, AI in charge

of our, you know, weapons or something silly like that,

then, you know, bad things will happen.

At the same time,

the second part of Terminator was it became sentient, right?

It actually learned to think on its own,

and I don't think we're anywhere near that.

So I have no concern of sentient AI.

Right now, AI is not intelligent.

They say AI, artificial intelligence,

there's no actual intelligence.

It has no notion of what it's saying or doing.

It is simply pattern matching at a scale

we've never pattern-matched before.

From I_got_too_silly,

What benefits do legged robots have over wheeled

or tracked vehicles?

Legs are inherently beneficial if you wanna operate

in environments for which they're built for humans,

and more importantly, where things are not flat.

So wheels are massively efficient

in how you can move around environments

as long as there's not uneven terrain.

If you've ever been in a wheelchair

or wheeled someone out around in a wheelchair,

you realize very quickly how flat the world is not.

Even what you perceive as flat,

even in a city environment

where there's sidewalks and everything,

you realize there's curbs that don't dip enough,

there's big breaks in them,

and all those little things for wheels become big problems.

They become sort of sticking points

and legs have the inherent ability

to walk more robustly over multiple terrain types.

Quadrupeds being one example,

I mean, where your dog or your cat can go on four legs

is pretty incredible.

Bipeds, of course, are sort of the ultimate expression

of mobility in human environments,

'cause the environments are built for us.

So if you need to get into a small space

and get up a small set of stairs or something like that

or climb up a ladder, only a biped can do that.

Give a ladder to a quadruped,

it's not gonna know what to do.

Give a ladder to a human,

as long as they're reasonably healthy,

they can climb up that ladder no problem.

From SamMee514,

What are robot dogs actually being used for?

So we've actually come a long ways in robot dogs,

which are actually technically called quadrupeds,

'cause they have four legs.

It's really amazing what's happened

in the last sort of decade

where we've gone from these robot dogs

being in really lab environments

and research environments

to being things that you can buy at insanely low prices.

The hardware has made immense progress.

I think the practical use cases are still a little thin.

They're thinking about doing them

as things like inspecting buildings,

sending robots ahead in disastrous scenarios, right?

And for that, legs are definitely better than wheels.

The next question is from Leovan21,

Is there any attempt to put ChatGPT inside a robot?

Yeah, there's lots of attempts.

So right now, there's many humanoid robot makers

that have that as a layer in their humanoid robots.

And so, the way to think about this more generally

is that as we're making robots do things,

there's no one thing

that's gonna make them do all things, right?

So, imagine your body, you have a brain,

you also have a spinal cord, you have proprioception.

So you sort of intrinsically have multiple computers

running in your body at any given time.

Like, your spinal cord is a computer in its own right

and those computers run different algorithms.

So at the highest level, at our cognitive level,

that's where LLMs will sort of play a role, if you'd like.

So you wanna ask the robot a question

and have it perceive something about the environment

and make a decision.

So ask the robot, you know,

Where is the red apple on the table?

It can use an LLM to parse that language into code,

use computer vision then to detect all the apples

on the table and return based on that an estimate

of where it is.

When you actually wanna reach for the apple,

an LLM is not gonna reach for the apple.

That's where traditional robot control will come in.

Or where reinforcement learning will come in,

which is a whole nother type of learning

that's different from LLMs.

And those are what's actually running on the robot, right?

Those are what are making the robot go.

Just like your spinal cord

is really what moves your body most of the time

without you having to think about it.

From Iwannabefree10,

Are autonomous vehicles mobile robots?

Yeah, I mean if you wanna ask

what the most advanced robot is today,

you would take autonomous cars

and mobile warehouse robots like Amazon has.

Autonomous cars, I think, are one of the prime examples

of the farthest that we've pushed autonomy.

So yeah, it's absolutely a robot,

it's an advanced robot

that's a beautiful work of engineering.

From Vnightpersona,

How does the Roomba know what's what?

I think what they're asking is how does the Roomba

have a semantic understanding of the environment

is the way we would technically phrase this question.

That it's able to identify things around the room

that it sees and correctly identifying them.

So basically, we've been able to take a bunch

of training data,

a bunch of examples of pictures on the internet,

videos on the internet,

and teach robots how to correctly identify those.

By teach, what we mean is we can train up

basically this big neural network that takes images in

and produces what's in the images out.

And as a result, we can now correctly identify

most things in an image.

And so what Roomba does is, simply, it has a camera

that's perceiving those things in the environment.

It's checking with the internet based on these large models

and then identifying the things in the environment

and using that information to tell you what's going on.

The next question is from fireplacerror,

Why make robots humanoid shaped

when they could have six arms?

Humanoid robots are the most suited to do the most things,

even if they're not the best suited to do any given thing.

Again, the world is built for us.

We've built the robot for something of our shape

and our function.

Doors, stairs, narrow corridors, all these things.

And if you want a robot that can slot into any scenario

where a human can go,

which is pretty much most scenarios we care about,

not all but most,

you want somebody to start making you sandwiches

in your kitchen

or you want a robot

that can be dropped into an existing factory

and automate some tasks, right?

I wanna say all that as a preface to the fact

that humanoid robots are not the best form

for, again, a given application.

For example, right now,

the largest owner of robots is Amazon.

It has over a million.

And as a result,

one of the largest sector of robotics is warehouse robots.

Robots with two wheels that go under these pallets

and then move them around warehouses.

In particular, what Amazon does

is they have the robots go pick up a pallet

with the thing you order from Amazon,

and it moves that over to a human who's filling the order.

So the human can just stand there,

and they never have to move.

And basically all of this stuff comes to them.

Now you could theoretically have a humanoid robot do this,

like push these big pallets around the warehouse.

But the question would be why?

Like, these robots are very good

at doing what they're meant to do

and that environment is designed

to work synergistically with the robot.

So you could have six arms on a humanoid

if you had an application that determined

that six arms would make a big difference.

Maybe you have a firefighting robot,

and you find that two arms is not enough

and you want four more arms 'cause you wanna hold the hose

and you wanna hold a camera

and you wanna hold fire extinguisher.

That's the great thing about robots, by the way,

is we can make them any way we want.

A Reddit user asks, When it comes to automation,

how close is Amazon to actually automating most,

if not all, warehouse work with robots?

The answer is very, very close to parts

and then other ones are further away.

So in terms of automation of warehouses,

Amazon is by far the leader in this domain.

Amazon robotics in particular

has been working on warehouse robots,

specifically robots that move around warehouses

to move pallets around,

to move your order to a person filling the order

for, you know, 20-plus years now.

And they've really refined that process

so they maximally and efficiently can move packages

to the sorter and then the sorter packages them up.

So the question becomes what remains?

And what remains is then the part

that the human's currently doing,

which is grabbing these things off the shelf

and putting them in boxes.

And so this is much more an open-ended work,

where they're testing out solutions right now.

So for that, you can use things called local manipulators,

meaning you have robot arms maybe on mobile bases

and maybe you use suction cups or soft graspers

and you would pick up objects and put them in boxes.

And you can do this with varying success levels right now,

but you're not at the success level to truly automate that

where you don't need a person present for all objects.

'Cause you imagine all the objects that go in Amazon boxes,

all the different geometries and how they feel and look,

I mean, it's a very complicated problem

to automatically and autonomously load all these objects.

So that problem's sort of open-ended.

KaleidoscopeInside asks,

If you could commission a robot to be built

with no financial barriers, what would your robot do?

So if I had no financial constraints,

I would try to sort of solve the exoskeleton problem.

Like a billion could get it done,

a billion in my mind could develop something

that would eradicate the need for wheelchairs, essentially.

A Reddit user asks,

Why do most of the 4-legged robots,

see Boston Dynamics,

have their knees look backwards?

It's the inverted leg morphology.

There's potentially mathematical advantage

to having your legs like this.

It's actually a twofold thing.

It's not just that they're inverted,

but that they're very light.

So it turns out the way you control robots

and the way you get really stable locomotion behavior

on robots is by having very light legs

and having most of your mass centered in one spot.

There's lots of mathematical reasons for that,

that leverage that as sort of an assumption.

And what that means is as you're walking,

if you're legs are very light,

you can place them very quickly

to catch yourself as you fall.

A lot of the robots were designed based on that morphology.

And in addition,

the inverted leg is actually what birds have.

'Cause birds actually have this type of morphology

where they have very heavy big bodies and very light legs.

And as a result,

birds have some of the most robust locomotion out there.

They can actually step down huge holes

and then step up out of the hole

without ever changing how their main body is moving.

And so if you can get that kind of behavior on robots,

they'll be very robust and able to go all the places

where you want to take robots with legs.

A Reddit user asks,

What is so special about the Mars rover Curiosity?

I mean, what's not special about a Mars rover?

Pretty much everything, it was on Mars.

I mean the reality is that these robots

are incredible engineering feats.

First, just getting it to Mars is special.

And then the rover itself,

all the rovers that have gone,

Curiosity and all of them,

they had these amazing engineering sort of gems in them

that they built up.

You know, their suspension system is specifically built

so that it can handle rough terrain.

Their wheels are specifically built so they'll be robust

to going over rocks and things like that,

and you won't blow out a tire, right?

So the entire structure is built to maximally be robust

to this really harsh environment.

I mean, there's the electronics

which have to deal with this very sandy

and windswept environment

where there's dust storms, you know.

There's power limitations.

Just the power alone,

I mean, the sun doesn't come through at the same intensity.

So you have to have good solar panels

that actually charge the electronics.

Electronics have to be battery-efficient, right?

And then after all of that, you have to do science with it.

And then, oh, of course,

you have to talk back with Earth at the same time

and make sure you don't lose that signal.

I mean, the list goes on and on.

But to build a system like that,

I mean, every one of the rovers we've sent to Mars

is an immense and remarkable engineering feat

that sort of just makes me happy as an engineer.

The next question is from FKruzewski34111,

Wait, how do autonomous cars even work?

Like, are we trusting robots with our lives now?

Not sure about this.

It turns out that autonomous cars

have gone through some things.

About a decade ago, when it seemed to be closed,

there were all these startup companies that formed,

and they all tried to do autonomous cars

many different ways.

And then exactly what you're worried about happened.

There were crashes, people got hurt.

Despite the fact that they'd already been working

on making the system safe,

but they really realized,

and I think that the entire autonomous car industry

pushed more and more to make sure

there was really rigorous safety methods.

So both in the algorithms to keep 'em safe,

and then also protocols around that,

software, architectures, et cetera,

and then data to support it.

And so in that way, there's a lot of evidence to show

that autonomous cars, if they're currently deployed,

have very rigorous safety standards,

and there's a lot of data to back that up

with the limited number of crashes that they've had.

I'm not saying they can't get better,

but, you know, after a decade,

autonomous cars have done the hard work.

A decade ago, we could have autonomous cars

drive around on the street,

but the difference between driving around on the street

for, you know, a demo and actually driving around

and all of a sudden other cars pull out

or you know, a ball bounces into the street

and a kid chases after it, right?

Those corner cases kill and safety kills,

meaning if you're not safe,

you die, both as a company, but also as a technology.

The next question is from TimeIsGrand,

Noob question:

Why does Elon Musk think LiDAR is not a good idea?

I don't know Elon Musk's mindset,

but in my mind, LiDAR is awesome.

So you have really two main sensing modalities in robots,

autonomous cars, et cetera.

I mean there's many more,

but in terms of perceiving the environment,

the two most popular things,

they're using cameras, of course, which is what Tesla does,

and you have LiDAR.

Now, LiDAR sends out a bunch of laser pulses

and those bounce back and tell you what's around you.

So it's sort of 3D radar, but now it's using lasers, LiDAR.

Cameras, of course, have a lot of information present.

They can tell you not just how far objects are,

if you can properly estimate distance,

but what objects there are

and where they're at in the environment.

So you can get semantic understanding,

the environment and things like that.

LiDAR does an incredible job

of precisely identifying everything in the environment

with like a full 360 view

so you know exactly where all the objects are,

but you have no idea really what the objects are.

LiDAR is super effective on robots and on autonomous cars

because what you wanna do

is make sure you're not hitting anything

and that means anything.

It doesn't matter if you're sort of hitting another car

or hitting, you know, the guardrails

on the side of the road.

You don't wanna hit any of that stuff.

And for that, LiDAR can work really quickly

and really robustly

to do things like dynamic collision avoidance.

Cameras, again, give you this semantic understanding.

And so I imagine that what they're thinking at Tesla,

generally speaking,

is if they can solve the camera problem

and make them as good as LiDAR,

then this will extend to a lot of other application domains.

But I think what I found on robots

is you wanna use every sensor you can get your hands on.

And so for that, for safety critical applications,

which Tesla cars could still improve on, to be honest,

you want LiDAR in the loop

so you can quickly and rapidly respond

to dynamic changes in the environment

and not have to deal with the latency

that's present in a lot of perception-based representations.

We've got a Reddit user who asks,

Why is a surgical robot better than a surgeon?

The reality is that robots are really good

at certain things.

They're very, very good at precise small motions

and doing those precise motions

repeatedly again and again and again.

There's also places where humans

are infinitely better than robots.

Basically, any place where you have to interact

with the environment in a soft and tactile way.

And so here's where actually surgical robots get complicated

because our skin is soft, our organs are soft,

our bodies are soft.

And so as this robot is moving through your body

to perform surgery,

you need to be aware of those soft

and sort of compliant interactions with the human body.

And that's why a surgeon is in the mix.

The surgical robot can do those precise motions

while the surgeon operates the robot,

and they get all this haptic and tactile feedback

while they're doing the robotic surgery.

So they can take advantage of the precision of the robot

while still being able to do what humans do really well,

which is operate in these complex environments.

This is from theballisrond,

Why is it so difficult to make a clothes folding robot?

The problem with clothes

is that clothes are really difficult to model.

They move around, there's fabric.

I mean, it's very difficult to handle.

So a robot has to interact with something in the environment

that we can't put in a computer easily.

And so that's why this is a big challenge

and why there's been a big push

to use things like machine learning and AI

to try to understand this clothes folding problem.

That is, you train a robot

with the humans folding clothes,

like you teleoperate the robot

so that it folds the clothes a bunch of times

and then you try to teach the robot that task

without it actually having a model

of the environment itself.

But just these reference trajectories

that humans generated through teleoperation.

The next question is from SurpriseNew763,

Whatever happened to the NEO robot?

It's interesting because there was this big announcement

that you could buy like a humanoid robot

as soon as this year in fact and have in your home.

Now, to give them credit,

they completely were transparent about the fact

that this robot couldn't actually do a lot

without teleoperation.

There's even in the app,

they show you how you will book time on the robot

with a teleoperator.

So this is interesting from a couple perspectives.

One is it's a tacit, at least, acknowledgement

that humanoid robots are not actually ready

to be deployed in your homes.

I mean, that's the clear and obvious ramification.

So the question is why are they going to put a robot

in the home that's not actually ready,

meaning it can't do a lot of tasks on its own, very few?

The reason why is right now,

the idea in the robotics industry

is that the only thing that's missing is data.

So we have internet-scale data

and that's what made LLMs possible.

So what they're betting on

is if we can get sort of internet-scale humanoid robot data,

we will be able to solve all these problems

just like ChatGPT solve the problems for language,

we'll be able to solve in this other way.

So the idea with the NEO robot

is not so much that they think

they're really gonna sell these robots

and make money off them by having them teleoperated,

but rather this is an amazing way to do data collection.

You put all these robots in homes,

and then people agree to let you teleoperate them

in their home.

They collect all this data,

and then they're gonna use this data to train a large model

to determine how to make the robot

do these tasks automatically.

If you can get early adopters to sign up for that,

because it's cool, you'll be able to generate enough data

to actually train large models

to learn how to do these tasks.

But I wanna say one final thing

is that I don't think it's gonna work.

I don't think that there's enough data

you can collect in this way

to solve this problem with just human data.

People tend to confuse LLMs and text with robots,

which are fundamentally different things.

In particular, the amount of data needed to understand

how a robot moves, these trajectories,

think about them as trajectories as the language of robots.

And they include position, velocity, force,

all this rich dynamic information.

But they're so variable

'cause there's so many degrees of freedom

that it makes language seem simple.

And in fact, the way to really understand this

is look at evolution.

We developed language in a very small fraction

of our total evolution.

Most of the time we spent evolving

was to do these other basic things, right?

Things that require motion of our body.

That's why they say you only use 10% of your brain.

What they mean by that

is that a lot of your brain is being used

for these very complicated motions that we do, right?

And so the point of this all being

is I don't think that data's gonna work.

I think that what we're missing is not more data,

not that it can't be helpful or help to polish things,

but we need to understand physics as well.

You need to merge physics with human data

and that's the only way you're gonna solve

the general intelligence problem on humanoid robots,

because robots are not language.

That's it. That's all the questions.

Hope you learned something along the way. Thanks.

[gentle bright music]