Horizons from PBS News

I’m William Brangham, and this is "Horizons."

Lifelike humanoid robots have captured our imagination for decades.

Thank you, ma’am.

I wish I had your confidence.

Brangham: And today, those robots work in factories and assembly lines all over the world.

Will they, as some companies promise, soon be working in our own homes?

As this technology speeds forward, are we racing into a new era, separating fact from fiction about our robotic future?

Coming up next.

♪ Announcer: Support for "Horizons" has been provided by... Steve and Marilyn Kerman and the Gordon and Betty Moore Foundation.

Additional support is provided by Friends of the NewsHour.

♪ This program was made possible by contributions to your PBS station from viewers like you.

Thank you.

From the David M. Rubenstein Studio at WETA in Washington, here is William Brangham.

Welcome to "Horizons."

Would you welcome a robot into your life?

I don’t mean one of those rolling vacuum cleaners.

I mean an actual life-sized humanoid robot.

I understand why this is a complicated question, because for every depiction we’ve had of an efficient, helpful robot, like the prim, elegant C-3PO from "Star Wars."

I am C-3PO, Human-Cyborg Relations, and this is my counterpart, R2-D2.

Skywalker: Hello.

Brangham: There are countless other nightmarish ones, like Roy Batty, the murderous replicant played by Rutger Hauer in Ridley Scott’s film "Blade Runner."

Quite an experience to live in fear, isn’t it?

Brangham: And, of course, there’s Arnold Schwarzenegger’s infamous Terminator, a robot built to kill by a rogue AI network.

I’ll be back.

Brangham: But that has not stopped the sales pitch.

Bornich: My name is Bernt, and today we’re launching NEO, our humanoid for the home.

Brangham: The company 1X has promised that its robot, NEO, which it claims can do everything from help you around the house to teaching you new dance moves, will be ready for shipment soon.

Well, wouldn’t it be great if everyone could have their own personal C-3PO R2-D2?

Brangham: Elon Musk, whose company Tesla is building its own line of humanoid robots, predicts that within 10 years, we’ll have a billion of them deployed, and within 20, he says there will be 100 billion.

That’s 10 times more robots on Earth than humans.

But how realistic is this mechanized future?

Is this something people actually want?

And how smart is it to bring these machines, many of which will be guided by AI, into our homes?

So, for a reality check, we turn to Aaron Ames.

He’s a professor of mechanical and civil engineering and control and dynamical systems at the California Institute of Technology.

He works with robots all the time, leading the AMBER Lab at Caltech, which stands for Advanced Mechanical Bipedal Experimental Robotics.

Aaron, so great to have you here on the show.

Thank you for being here.

Ames: Pleasure to be here.

Brangham: Let’s talk definitions first.

We are talking humanoid robots, not those rolling flat Frisbees on the floor that vacuum our houses, not even my car with its semi-self-driving abilities.

We’re talking about humanoids, the ones that look like you and me, heads, arms, legs.

When you see the current conversation that is happening around those robots, what are your initial impressions?

What do you think?

I mean, it’s a surprising mixture of excitement and trepidation, right?

I mean, the excitement comes from the fact that humanoid robots are something I’ve passionately loved since I was a kid, right?

I grew up on humanoid robots.

I grew up reading science fiction, Isaac Asimov, where, again, this was another example of robots that could help us and make our lives better.

And this idea that we could make something kind of in our image that could come with us on our journey and assist us along the way is really, really compelling.

At the same time, I’m nervous, right?

I’m nervous not about robots killing us or taking over the world or anything like that.

Brangham: That’s not on your fear list?

That’s not on my fear list right now.

What’s really on my fear list is this sort of overpromising what we can do with robots near term.

So this expectation game in terms of what people should really expect these robots to do, what their function is, what their purpose is, and how we can actually get them out to do something useful and functional.

And I think there’s a massive gap between sort of the vision of what we’d like with humanoids, because it’s a very natural vision, right?

I mean, people understand implicitly and immediately the value of robots that could go and do all the things that we do off the shelf, right?

But how do we do it in a way that’s robust, reliable, and most importantly, safe and trustworthy, right?

How can we have these things be integral parts of our lives where they’re additive, not subtractive?

Brangham: And you don’t think we’re really anywhere.

You think the conversation has gone far beyond the actual reality.

Ames: I think the expectations misalign with reality to a large degree, right?

In terms of where we’re at, where these robots are going to go, what they’re going to do, and the timelines to do things like have them in our home, right?

I mean, this is the holy grail of robotics.

As a humanoid robot, as your sort of personal assistants in the home, doing your laundry, folding it up... Brangham: Taking care of old people.

Ames: Yeah, yeah, taking care of old people.

I mean, this is the long-term goal, but the reality is just much different, right?

I mean, the technology we need to get there is still undone.

So is it, in your mind, a technological hurdle that is holding us up right now?

Ames: That’s right.

That’s right.

I think at a fundamental level, people conflate the success of AI.

What I really mean there is the success of LLMs.

So I think we need to separate out... large language models.

What I mean is I think people kind of put AI in this big bucket, and I think AI is not a great word, because it’s too big and includes too many things, including standard things that are not AI, right?

I think control algorithms are now called AI on a broad level, right?

So things that run in your car are called AI, but they’re not.

They’re not an LLM.

They’re not intelligent.

They’re not inferring off something.

So I think we need to separate out these pieces a little bit to understand where we’re at, right?

So LLMs, large language models, have really moved the needle for what we can do with computer intelligence, right?

With machine learning.

I mean, it’s amazing when you can talk to a computer.

Brangham: Incredible.

- Yeah, it’s incredible.

It really is.

And I think people think that because computers are doing this very hard thing, like computers can do calculus now, right?

Like it’s nothing.

Brangham: Solve really complex math problems.

Solve really hard problems.

I mean, calculus is an understatement, right?

Literally solve open problems in math.

And that’s incredible, but they’re just computers.

And the problems they’re solving are very complex, but they’re patterned off existing knowledge.

And they take that knowledge, they take this Internet-scale data, and they combine it in a way that they can output things that we understand.

And we confuse that pattern matching at immense scale, but pattern matching nonetheless, with intelligence, right?

The ability to step outside of a domain knowledge and get somewhere new.

I think we also confuse this with physical intelligence.

Like the ability to do things.

So, you know, in terms of the way to understand this is evolutionarily, it took us millions of years.

I mean, it’s hard to sort of benchmark exactly how much time, but millions of years to develop things like the ability to hold a pen, right?

The ability, just the simple thing and take the cap off a pen, right?

And little manipulation tasks, to walk, to move, to ambulate, right?

To balance and to throw balls and describe a way.

So all these physical movements, this physical intelligence took a long, long time.

And we have incredibly complex structures in our body that do this seamlessly.

And so we sort of take it for granted because everyone can do it.

So it seems easy.

A baby gets up and you start to see it happening in incremental stages.

And you think, "Oh, I didn’t even teach him to do that, or to pick it up."

Ames: It seems natural and sort of easy.

On the other hand, calculus, which is hard, seems really challenging.

So we sort of say, "Well, it can do calculus now."

And I’m sort of using this example for all math.

And so it must be able to walk because not everyone can do calculus, but not everyone can walk, but most people can, right?

And so it seems like it should be able to do the easy things if it can do the hard things.

But those things are fundamentally different.

Again, one lives in the realm of language where it has an immense example set, like an Internet scale example set.

The other one is physical motions for which there’s very little data.

And that’s why in the humanoid space right now, for example, the focus is on collecting all this data of people doing things, of robot doing things, because the feeling is somehow if you can get to Internet scale data, you can recreate that lightning in the bottle that was LLMs for language, now it’s physical movement, right?

But that’s the gap is right now we have not bridged that gap.

Right now there’s some evidence, but the gap to actually getting something out there that’s robust, reliable, and deployable is massive.

And this is where the expectations misalign a bit.

So I’ve been in preparation to talk with you.

I’ve been looking at all sorts of different videos of robots doing different things.

And there have been some of these videos of robots changing a light bulb and that people had pointed to that as the dexterity required to pick up that delicate round object and to turn it with just enough force or to pick up a raspberry or something very subtle, that that was a sign of a breakthrough here.

Are you arguing that that’s just a discreet little thing that doesn’t add up to all the other things we would need a robot in the home to do?

That’s a good way to say it.

I mean, these are important steps, right?

So this doesn’t take away.

Just to be clear, I talked about this mixed feelings at the beginning, right?

Excitement.

It’s great that we can have a robot unscrew a light bulb, right?

It’s great that we can have a robot run really fast.

And just the mobility we’ve gotten on humanoids is incredible in the last year.

There’s been huge technological movement.

So I don’t want to minimize that amazing accomplishment that’s happened across the robotic space.

But we can’t confuse doing a singular task once in a lab setting, right?

Versus doing something in your home repeatedly in a diverse way.

Brangham: Right, where the ground is uneven and who knows how my mother might move.

Ames: And the light bulbs change size.

And that light bulb, the one that the data was collected on.

So in that case, there’s the light bulb unscrewing kind of example.

They had a bunch of data.

They look at teleop data, right?

They have people first use the robot to unscrew the light bulb.

This is the classic way that people do it now.

They teleop, and then they take all that data and they train something that does the same thing, right?

So they’re sort of mimicking human motion.

It’s the same with the running and the walking.

A lot of this is based on either human data or some physics thrown in.

Brangham: Or the dancing, too.

Ames: Or the dancing.

I mean, the dancing that you see.

This isn’t organic dancing that the robot generated on its own.

Wait, the robots don’t know the Macarena?

Well, the ones in my lab do, but everywhere else, no.

No, no, no.

I mean, there’s no, yeah, there’s no knowledge.

You have to understand that the way to characterize the current state of technology is we’ve gotten really good at playing back motions on robots.

Brangham: Playing back motions.

So we show the robot how to dance, how to pick up a raspberry, whatever that is.

And that they, after multiple, multiple, multiple tries, can do that.

Ames: That.

Sometimes.

- But that is very different.

Ames: And that’s the key.

A, they can do that, and they can do that with some success rate that’s actually not that great yet, right?

So 50% of the time, right?

Or 70%, or even give it 80%, right?

So what do we need for reliable deployment?

Well, we need to have at least 99, let’s say 99.9% success rates in labs, right?

Because when you go to the field, everything’s different.

Like your home, the light bulbs are slightly different sized, and they might be old light bulbs that no one had in their data set, right?

So you need to generalize, and not just light bulbs, but everything else around the house.

Think about all the tasks that you do on a daily basis that you don’t think about, right?

You put your keys on the counter, right?

You go make a cup of coffee, all the buttons you have to press, all the little things you have to move.

How do we put all of those pieces together holistically and reliably?

And again, to go back, the way to understand where we’re at right now is we can play back motions on robots that we could never play back before, but we can play back those motions.

So what we have is automation, if you’d like, really good automation, but we’re starting to get automation, but not autonomy.

Autonomy is where the robot has to deal with unseen circumstances, make decisions about it, and accommodate it.

And that’s one of the big gaps, right?

Brangham: And how do we bridge that?

Is that using AI?

Is it training an AI to recognize, "I think I’m going for a coffee pot," but the coffee pot’s actually on the drying rack, or do I need to visually recognize where it is and how to hold it, and all of those things?

You just start to sketch out the complexity involved here.

Is it extraordinary?

Ames: It’s extraordinarily complex, and this is sort of what I really do in my job is how do we understand that complexity?

How do we understand that system-level architecture?

And this is what it really is, is how do we build that system at a layered architecture level?

So I think this is the root of the scientific problem, and let me break it into two pieces.

I think right now where there’s misalignment is trying to exactly mimic and recreate the LLM, again, lightning in a bottle.

And for LLMs, all we needed was data and lots of it because language is based on data.

Language is a discrete set, and we have lots of examples for it.

So there was nothing else to do.

There was no mathematics.

There was no physics under the hood.

So the first thing you have to realize is the physical world is not the digital language world.

And so what’s the first thing you need, in my opinion, and this is all now my opinion because this actually differs from the opinion of a lot of people, is that you need physics.

You need F equals MA.

You need math.

You need principles, right?

Like F always equals MA.

If you want to understand how objects move in the environment, Newton figured this out hundreds of years ago, right?

And we can take that knowledge to scale data, because physics generalizes.

We were talking about this generalization because that’s one of the key components is generalization.

Physics lets you go places where you don’t have data, right?

So if I need to know how a shirt falls when it comes off a clothesline or how a stack of newspaper might flop on a driveway, all of those things about the way in which the actual world works, you have to train them.

You have to train them.

And what you need to do is separate out the things that don’t have good physics.

In fact, things like how a cloth falls, physics for that is really hard, right?

To predict that, to simulate that is really hard.

But to simulate like something falling off your counter, like a ball falling off your counter, let’s say, that’s actually really easy, right?

So when you have, and the way to kind of differentiate is when you have sort of rigid objects in the environment that classic physics describes really well.

Brangham: Right, Newton’s apple.

Yeah, we have like Newton’s apple.

We have lots of examples of that.

And let’s take the Newton example with the apple because that’s a really good one.

The modern architecture, if you’d like, the modern way that a lot of people are pushing this to recreate the LLM bit is to, we want to predict how long it takes for that apple to fall, right?

Because we want a robot to go pick it up or whatever the reason.

So what are they going to do?

They’re going to have cameras observing an orchard with apples falling all day long.

They’re going to extract the fall data, and they’re going to train some, you know, intelligent system to predict the time it takes for apples to fall, right?

And this is exactly a point where I think we need a little bit different perspective, which is we know how long it actually takes for the apple to fall.

Let’s seed that physics data into that and then augment the data, because the physics predicts how far, how long it takes to fall, but it might not predict if there’s a gusty wind condition or all those more complicated things.

So what you need to do is combine the two together, is use physics to scale the data, right?

And then collect data to augment that physics.

And you put the two pieces together.

So that’s the first thing I think that actually is sort of being missed right now is that there’s such a reliance on data, because data works so well for LLMs that there’s been sort of a lack of focus on bringing good old-fashioned physics and engineering principles into the mix when it comes to robots.

And these are very powerful methods that we aren’t properly leveraging as a community right now, right?

Because we’re trying to sort of recreate that spark.

Can I go back to your concern, which is I understand that you feel that there might be some hyperbole in this marketplace right now because people are trying to raise money and blah, blah, blah.

Is the concern also that an early deployment of a robot causes a problem that creates a PR nightmare for your entire industry and sets back all the work that you all are doing years?

Yeah, that’s exactly the concern.

That’s my fear.

Sort of I wake up every morning hoping I don’t have a bunch of text messages from people saying that a robot hurt someone.

Like this is what I’m terrified on a daily basis.

And not that it hurt someone willfully, that the AI tried to hurt someone.

That people put robots in an environment where it just, yeah.

Like that boy who was kicked by the Chinese robot.

Ames: There’s that video that I can never get out of my mind.

It’s a unitary robot.

Again, this goes back to the playback thing we were talking about.

Right now these robots are just playing back motions.

There’s not an intelligence there really.

It’s just playback.

And when you play back a motion, that was like sort of a dance kick sequence.

And the crowd got close.

And the robot’s not perceiving the crowd.

This is completely what we call open loop, right, playback.

And the boy got in the way of the kick and got kicked in the stomach.

And this is what’s going to happen because right now, again, there’s not this intelligence layer happening on the top.

And this also gets back to this physics idea.

So think about data and learning as one element of the bigger picture.

And I think when we talked about what is it going to take in the system level architecture, this is what I really mean is sort of good old-fashioned engineering.

What I mean is data learning is going to be huge and important.

But you’re going to need physics to feed into that data.

But this is a really important thing.

You need physics on the back end.

So anything that goes on the robot needs to be checked, checked with math.

We need math to make sure that the robot does not do bad things.

So if we just roll out these things open loop, it is going to happen that a robot is going to hurt someone.

Brangham: Not intentionally.

Not intentionally, to be very clear.

I mean, if you have a big, heavy robot doing the Macarena, and a kid’s there and doesn’t know that the robot’s playing back the Macarena and steps in front of the robot at the wrong time, the kid gets hurt.

Brangham: Right.

Ames: And so what you need is safety systems to prevent that from happening, which means the robot needs to perceive its environment and always make sure that it’s doing the right thing.

And by the right thing, I mean, not hitting kids, not hitting things.

I mean, it needs to be safe.

And we know how to do this, by the way.

We’ve done this for many, many robotic systems.

You talked about your car and the intelligence systems that it has on it or even these ground robots and these mobile robots.

So we aren’t talking about those today, but we have lessons from those industries that we can take and translate over.

We know how to safely and autonomously deploy robots in real-world scenarios, and this requires this principled, structured, mathematical approach that wraps around the learning as a protective layer.

So just because I have you, and I have to ask you this, on this idea of intentionality, you don’t fear a Terminator future?

Ames: I don’t fear a Terminator future.

Brangham: Because?

Because I think that the near-term danger is putting robots in places where they shouldn’t be yet by humans.

I mean, I think the danger is humans being so excited about the technology and they want to get it out there so bad, and they want to deliver on this promise because it’s exciting.

And I want to say that.

It’s incredible to look at.

I love seeing robots in the home.

I love seeing humanoids walking around the street, right?

Like I show off humanoids all the time to kids in my lab.

Like their fascination, again, this is this mixed feelings we talk about.

I mean, I’m deeply passionate about this idea that robots can be part of our lives.

And at the same time, again, my fear is that that excitement will take over rationality.

Brangham: Right.

Coupled with massive market pressures to deliver these things and deploy these things.

And they’re going to put it in places where they aren’t ready.

And they’re not thinking about these things yet.

I mean, there’s just not a proper effort on safety, robustness, reliability.

These are things that sort of fight performance, right?

They slow you down when they’re not done right.

And when they’re done right, they can make you still performant.

But balancing that is really, really hard.

And people don’t want to slow down right now.

They don’t want to put the brakes on.

They don’t want a safety system.

They don’t want to deal with the seatbelts, right?

They want to go unconstrained, go really fast.

And I get that.

Brangham: Like those robots we saw in the Olympics.

They went faster than Usain Bolt, but they also crashed and sparked out in the wall.

Ames: Perfect example of our moment, you know?

Really impressive.

We have a robot that runs faster than Usain Bolt.

Amazing.

Just amazing.

And I want to stress that, right?

Just fantastic moment in robotics that we finally can get robots to run faster than a human has ever run.

Brangham: But?

Ames: But the end is really the thing, right?

They’re doing this and then they crash and they burn and they blow up.

Because, again, the way to think about this and not to get too technical, but I can’t help myself sometimes.

I am a professor.

Brangham: [Laughs] It’s that there’s two axes, right?

There’s performance and there’s safety.

And performance is exactly what you saw at that robot Olympics, right?

Incredible speed.

That’s maximizing the performance.

But zero safety.

Like nothing about that was safe.

Nor was it meant to be, to be fair, right?

They were trying to be safe.

And they knew these things were going to hit the wall and they knew that.

But to get a robot in your home, you need this safety.

And all of the investment and all of the effort and all of sort of the hype is around this performance, what we can do with them.

But if we don’t get them to be safe, we just cannot put them out in the real world.

Brangham: Can I ask you just in the minute we have left, given your evident passion for all of this, do you ever just feel like you’re like one of those now guys who’s like shouting at the kids, "Get off my lawn"?

You obviously love these and want them to go forward, but you’re also counseling us to be careful and to be slow.

Yeah, a little bit.

A little bit.

I feel that way, yeah.

Brangham: You must be a bit of a drag at a party.

Yeah.

I feel like I’m the sober one, right?

That’s like, "This is great.

"You guys are having a lot of fun."

But, you know, somebody’s got to drive home, you know, and then, you know, the sober guy does it.

And it’s true.

I mean, yeah.

And I don’t want to rain on people’s parade, right?

I mean, we’re at this, I mean, I want to stress the progress we’ve had in the last couple years on humanoid robots is nothing short of breathtaking and remarkable.

Brangham: Right.

- I mean, 3 years ago, we had a couple examples of humanoid robots that could do a couple things and maybe walk.

And now we have these robots doing these amazing tasks, right?

- And it’s really exciting.

Brangham: It is.

Ames: But if we’re not careful, you know, if we’re not principled, if we’re not honest with where we’re really at, and the fact that these are fun lab demos, these are not ready for the real world, and people should know that, you know, there’s going to be a reckoning.

And I don’t want that because it will slow us down in the end.

It will slow down our innovation in the end.

Brangham: Aaron Ames of Caltech, such a great conversation.

Thank you so much for being here.

Pleasure.

Thank you for having me, really.

And that is it for this episode of "Horizons."

You can watch us on YouTube or listen wherever you get your podcasts.

If you like what you hear, subscribe and give us a rating.

That’s going to help more people see this program.

Thank you so much for watching.

We’ll see you next week.

Announcer: Support for "Horizons" has been provided by... Steve and Marilyn Kerman and the Gordon and Betty Moore Foundation.

Additional support is provided by Friends of the NewsHour.

♪ This program was made possible by contributions to your PBS station from viewers like you.

Thank you.

♪ You’re watching PBS.