Download the Nature Podcast 29 May 2024

In this episode:

00:25 What the rise of AI language models means for robots

Companies are melding artificial intelligence with robotics, in an effort to catapult both to new heights. They hope that by incorporating the algorithms that power chatbots it will give robots more common-sense knowledge and let them tackle a wide range of tasks. However, although impressive demonstrations of AI-powered robots exist, many researchers say there is a long road to actual deployment, and that safety and reliability need to be considered.

News Feature: The AI revolution is coming to robots: how will it change them?

16:09 How the cockroach became a ubiquitous pest

Genetic research suggests that although the German cockroach (Blattella germanica) spread around the world from a population in Europe, its origins were actually in South Asia. By comparing genomes from cockroaches collected around the globe, a team could identify when and where different populations might have been established. They show that the insect pest probably began to spread east from South Asia around 390 years ago with the rise of European colonialism and the emergence of international trading companies, before hitching a ride into Europe and then spreading across the globe.

Nature News: The origin of the cockroach: how a notorious pest conquered the world

20:26: Rare element inserted into chemical ‘complex’ for the first time

Promethium is one of the rarest and most mysterious elements in the periodic table. Now, some eight decades after its discovery, researchers have managed to bind this radioactive element to other molecules to make a chemical ‘complex’. This feat will allow chemists to learn more about the properties of promethium filling a long-standing gap in the textbooks.

Nature News: Element from the periodic table’s far reaches coaxed into elusive compound

Subscribe to Nature Briefing, an unmissable daily round-up of science news, opinion and analysis free in your inbox every weekday.

Never miss an episode. Subscribe to the Nature Podcast on Apple Podcasts, Spotify, YouTube Music or your favourite podcast app. An RSS feed for the Nature Podcast is available too.

TRANSCRIPT

Benjamin Thompson

Hi everyone, Benjamin from the Nature Podcast here. Something a little bit different this week, we're going to do a deep dive into some of the stories that have appeared in the Nature Briefing. And well, we've got an all-star cast to do it. Number one, Lizzie Gibney, Lizzie, how you doing today?

Lizzie Gibney

Hello, very well thank you.

Benjamin Thompson

Excellent. And number two, Flora Graham. Flora, hi.

Flora Graham

Hi. Great to be here. Thanks very much.

Benjamin Thompson

Lizzie, why don't you go first? Because, well, speaking of deep dives, that’s something you've been doing for a little while you've been looking at robotics and AI, and you've written a Feature about it.

Lizzie Gibney

That's right. So most kind of robots have some form of AI, but not the kind of AI that has been sweeping the world in the last couple of years. So the basic idea behind this story is asking the question, if we put foundation models, which are the kind of very generalised models that are behind the chatbots that we have online and the image generators, these kinds of very general, very powerful, very large models. If we put those into robots, could we finally someday have the kind of robots that we've maybe all been dreaming of? You know, I think many of us when we were younger, we expected by now that there would be some kind of equivalent of a, you know, robot butler-thing like that. And instead, you know, even the very, very best robots we have right now tend to be excellent but at a very specific thing like in a factory at a very specific task. And they might even be just an actual robot arm rather than the whole body. Even, you know, the really good ones that we might have seen doing parkour or showing us some funky dance moves —there's a company in US called Boston Dynamics — even those, which are incredible robots, are to some extent, pre-programmed, they like pick the moves that they do from a library within them. So this is looking at what if we could have robots that actually learn and therefore become much more adaptable than the robots that we have today.

Benjamin Thompson

So how do these type of AI’s work in this context then?

Lizzie Gibney

So well this particular idea of foundation models, so these kind of large language model-type models, what they do is they learn from text data and image data and then they also have all these many, many examples of robots in action that they learn from. That can be robots which are being teleoperated they’re actually operated by a human, or they can be doing it like reinforcement learning, they can be doing it again and again, and getting some kind of feedback on successes and failures. And they learn it’s very like large language model, they essentially absorb all of that information and they start to make these statistical predictions as to what comes next. So the ultimate result is that you have a model that says, I can see this image around me, you know, robots view eyes of the world, this is what I see now, I know what I want to do, or what the plan is and I kind of make a prediction as to how to achieve that. So it's very analogous to how the large language models come up with this extremely convincing sometimes unnervingly good sentences, it works in very much the same way.

Benjamin Thompson

And you said, the aim was to make these robots more adaptable, because doing stuff is hard, right? I mean I opened the door here balancing a cup of tea, and my laptop and all the rest of it. I mean, that's easy for me, because I've got a wealth of experience for how to do that and I can adapt. But that's maybe not quite so easy for a robot.

Lizzie Gibney

It's incredibly difficult. So you were holding a laptop and a cup of tea, I think almost every robot we have so far would probably completely fail at that challenge. Just opening a door on your own, you know, you've got to balance the forces, you've got to figure out the doorknob, what kind it is, what do you have to do to open it, when you pull it, you've got to make sure you don't topple backwards. There are a huge number of different things that go on in order to make a robot like that work. So it's something that we've been trying to tackle for years and years. But even you know, really good hands, like a human hand don't exist, like even just aiming for one kind of robotic replica of a human. So there are an awful lot of different parts that go into it. And it's very, very difficult. So what they did here was, as I said, kind of use this like large language model strategy. And that enables you to bring in a kind of element of common sense to a robot. So if your model learns on those things as well as robotic actions, you can teach it much less information because it gets a kind of general common sense kind of knowledge, like the language models do. So for instance, one of these robotic foundation models from Google, after they had trained it on both the Internet and on a bunch of videos of robots doing actions and information about the commands they were given, it was able to do things like move a Coke can onto a picture of Taylor Swift. Now, it might have never seen Coke can, it definitely had never seen Taylor Swift you know these things in action in a robot, but it kind of had the whole of the Internet to draw from. So up until now you need to train a robot for each different scenario and scenarios that look very similar to us might actually be very, very different to a robot because it doesn't know what's important and what isn't. So it might have been completely thrown by the fact that it had this image before it had never seen this image. Now, it's saying, okay well, I know who Taylor Swift is, that's all good. I can even switch out if you ask me for Pepsi rather than Coke, I'm very happy to do that, because I know the difference between those things as well. So that's kind of radically reducing the number of actual experiences that the robot needs to have seen, observed and learnt from in order to function very well in everyday life.

Flora Graham

I mean, you had me a robot butler, Lizzie, I've got to say. Often AI is being suggested to replace a lot of things that I'm quite happy to do like writing and making music. But I think that this is the exciting potential for a robot to do other things I don't want to do like do the laundry.

Lizzie Gibney

That’s right because those things are actually really hard for a robot to do. And that's partly why we haven't got to that point yet is they are really, really difficult.

Benjamin Thompson

I guess the dexterity aspect is a key one, right? Like you can train the robot on what a popstar looks like, or what can of fizzy drink looks like. But in terms of actual dexterous movements, I imagine there's not that much data to train these machines on.

Lizzie Gibney

There is very little. So in the past, we've had issues like the fact that each robot is completely distinct. So the way they work the kind of their embodiment, you know what they look like, that's completely different. So if you're trying to learn from robotic data, that hasn't always been possible to get the volume of data that we need. So what's going on at the moment, in order to try and build this internet of robotic data that doesn't exist right now, there are a few different ways of doing it, some groups are just pooling all of their data. So there are loads of robotics labs out there in the world and they're getting together and they're putting on all of these demos that the model can learn from. And they're bringing all of that data together, it's got their robots and loads of different environments. And that makes the robot — that's learned using that model — function much better in new environments, but it's still not enough. So they're putting together different kinds of robot bodies. So the learning from those different robots that are working in different places, and that have even different body parts themselves, they're trying to build models that can absorb all of that data as well. They can also learn from human videos, some of them, if they're humanoid robots, then they can learn an awful lot just from watching videos of humans doing things, which thankfully, we do have a lot of. And the final way they're doing it is simulation. So if you can build a world that is very physically similar to the real world, you can put a simulated version of your robot in that world. And it will interact in a way that you'd expect in the real world as well. So it gets valuable training data from that. This is happening a lot at Meta, Facebook's parent company, they're working in this area. And they have built a very rich simulator-environment and they're training a robotic dog called Spot to be able to put back different items that it finds where it thinks they should be in a house. And again, it sounds quite simple but they can do that over and over again, with lots of different kinds of environments, lots of different objects, they do it in simulation, they can do it much, much faster. No robotics gets worn out by doing that. And then they apply it to the real world and they've had quite a lot of success doing that.

Benjamin Thompson

I mean, it's got to be tough, though right? As someone who's played a lot of video games that are physics reproductions are quite good. But sometimes they're not quite there and you know stuff will fly off into the sky when it shouldn't. So unless the physics is one-to-one to the real world, to this virtual world, mistakes could be made, I suppose.

Lizzie Gibney

They could. And we're going to always need real data. And I think that comes to one of the potential problems with this method is, especially when it comes to quite finicky tasks, you know, things are very, very dexterous. I got the sense from talking to sources for my story that the people who are working in AI and applying these kinds of models that have been so successful up to now to robotics, they just think, oh, that's a detail, we'll kind of get to that down the line. And then there’re the people who worked in robotics, their whole career who say, actually, that stuff is really, really important. You can't just use this video data or even simulation data because we need to be inputting things feel like, how do people react? What are the forces on all of the different parts of the robot when something happens? That kind of data is very, very hard to get, very expensive and time-consuming to get. And it might be that this whole approach gets limited by the fact that we actually don't have that data. And maybe we get robots that can, bit like language models, can really seem to understand things. But really, are they able to function in the way that we need them to without having these failures in the real world?

Flora Graham

Interesting, it makes me think about the safety implications because that's always something that we talk a lot about with AI, but in this case, a use case that's often mentioned is rescuing people from disaster zones, places where humans wouldn't necessarily want to go because it's too dangerous. And in that case, you might have robots interacting with injured people or potentially making things worse than they were before. So I imagine that the possibility of failure, if you're talking about a robot that's designed to, let's say, assist an older person in the home, or someone who needs assistance with mobility or something, it's something that they probably do keep in mind.

Lizzie Gibney

100%. So as you say, we've talked about the issues around safety with language models many times, and they can be racist, they can be sexist, they can make things up, they can get things completely wrong. You know, we're now giving that effectively a physical body in the real world. So as harmful as all those things are, this is a different scale of harm, potentially, that they could cause. And I think the people who I spoke to for the story are well aware of that. They're obviously still working in this area so they're optimistic that we can overcome those problems. A lot of the work that's going on in AI safety at the moment can be transferred straight over to this robotics field. They are talking about applying things like Isaac Asimov's Laws of Robotics, you know, they've been around for a very long time, but you know, ways of overriding a robot so that it doesn't cause harm. For now, they're doing things like just telling it do not interact with anything living do not even try. And you know, they can make those much more kind of higher-level commands within the robot to overcome anything that it's learned. But basically, for that reason, they're not going to be set on the loose anywhere for a very, very long time. Right now these kinds of models many are being applied in places like factories where it's all extremely strictly controlled, and they have a lot of safety measures already.

Benjamin Thompson

On the other side of it as well. I mean, I know with so many robotics efforts, like the robot didn't work 99 times but the 100th time it did, these things aren't necessarily ready for primetime in that aspect yet.

Lizzie Gibney

Totally. So the people who I spoke to with my story, they are still doing things like I told you about moving a can of Coke from one place to another. It’s the fact that it can do that reliably in lots of different environments is a pretty big deal. And there are some that are about to roll out in factories. But I think there's also a lot of kind of showboating in this area. As soon as you start talking about humanoid robot, people get very excited, people are already very excited about chatbots and large language models. And this, what I feel is people anthropomorphizing these models and seeing understanding where really what they're doing is coming up with statistical association. So we have some demonstrations, which look kind of incredible. I'm talking in particular about Figure, which is from a collaboration between a robotics company and OpenAI, they've got this robot, and they asked, you give me something to eat, and it just picks an apple up and hands it over, which actually takes a lot of thought something to eat, that could be an apple, that's the only thing I have here to eat. I'm going to hand it over the dexterity of that. But we just have a video that shows us that that was put out by the company online. We don't know how reliable it is, we don't know if there was a different background, or the table was a different colour, or is a different piece of fruit. Or the man who asked the question moved his hand slightly. If any of those things would have caused it to fail, we just don't know. And when you talk to roboticist, they say, oh, yeah, like most of things fail most of the time. And that's why they are very sceptical.

Flora Graham

Well, one of the topics we've written about it in Nature many times sounds like science fiction, sounds like maybe even possibly a bit of paranoia. But I think there are people genuinely worried about it is autonomous or AI controlled weapons of war. I was just curious what the researchers working on this advanced level of robotics that you spoke to whether they had any thoughts about that? Do they consider their work totally separate from those concerns? Or, I mean, of course, anything can be used for good or ill not just robots, but it does seem like robots have a kind of special place in our fears sometimes.

Lizzie Gibney

I think because this is quite early stage and because people working in this area are so optimistic, I don't think it's directly on their radar. But I think if the kind of success that they foresee happening does happen, I think it's rapidly going to have to be on their radar, just like all of these issues about AI safety is something that they're going to have to confront pretty quickly.

Benjamin Thompson

Well one more on this then. You've talked a lot about how AI could improve robotics but I wonder about the other way round, could robotics improve AI in some way?

Lizzie Gibney

Totally. This was something that I found fascinating when writing the story, but really didn't have much space to fit into the piece in the end. But there's this idea out there that AI at the moment is going to be fundamentally limited unless it is able to learn and absorb and interact with the real world. You know, at the moment, it all takes place in a digital space. And that is not how humans grow up and learn. And there's this idea about, they call it the embodied AI, which is to say that maybe AI is going to have to have a physical form to get to that kind of human-like intelligence that we have. So already we could see that just spatial reasoning could be radically improved within an AI if it has an actual physical format in the real world. But it's possible that going beyond that that will become something that actually takes us from having these AIs that seem to understand and seem to be able to reason but probably aren't quite doing that. To something that has a kind of general, genuine intelligence. But yeah, we'll have to see, it's a theory lots of people are pursuing.

Benjamin Thompson

Absolutely fascinating. Thank you, Lizzie. And I have a feeling that AI and robots and robot-AI are something we might touch on again, on the Nature Podcast in the future. And Lizzie your Feature is out now, and we'll put a link to it in the show notes. But let's keep going. We've got a couple of quick stories. And I'll go next. And I've got a story about an animal that people generally aren't fans of. And that's the cockroach and it's the origin of the cockroach, how it's kind of conquered the world. It's a story that I read about in Nature.

Lizzie Gibney

Cockroaches, we're really gonna be putting a lot of people off, aren't we today.

Benjamin Thompson

Yeah right. And in particular, I'm talking about the German cockroach as its common name is, but it turns out that it actually didn't originate in Germany. Okay and this is based on a study published in PNAS, which suggested that the creature actually originated in South Asia and then spread globally because of its affinity for human habitats.

Flora Graham

And this is the cockroach that's kind of the ubiquitous fella that you might find all over the world. Is that right? So it might be called the German cockroach, but let's not hold that against anybody because this is the one that we've kind of said, how did this bug conquer the Earth? Is that right?

Benjamin Thompson

That's right. Its Latin name is Blattella germanica, okay, and this was first described in 1776, in Europe by Carl Linnaeus. And so it's got this name, and I think everyone made assumptions about its origins, okay. But say, researchers really want to know evolutionary, what the story was, okay,.where this cockroach came from? And so a team of researchers analysed the genomes of 281 German cockroaches collected from 17 countries, including Australia, Ethiopia, Indonesia, Ukraine, and the US, okay. And they looked at the similarities and differences between the genomes to calculate when and where the different populations might have established. And it turns out, the closest living relative to the German cockroach is probably the Asian cockroach, which has got a slightly different Latin name, Blattella asahinai, okay. And that's still found in South Asia and there's two species probably split off they recon maybe 2,100 years ago. And then fast forward a little bit of time so about 1,200 years ago, the German cockroach began its travels when it hitched a ride west to the Middle East.

Lizzie Gibney

So is its history linked with human history and travel then, if that was obviously a very rich time for trade within the Middle East, did the cockroach kind of go forth and conquer from there?

Benjamin Thompson

You’re absolutely right Lizzie. It seem like it hitchhiked on kind of commercial and military traffic. But also, it didn't just spread there. It began to spread east from South Asia around 390 years ago, so quite recently, and that was with the rise of European colonisation and the emergence of international trading companies like the Dutch and British East India companies. And so around a century later it hitchhiked to Europe, and from there spread around the world to Europe. And you know, there it is, it's ubiquitous now and that's where it got its name.

Flora Graham

One more of the treats that we brought to the world, then, on our trading ships in colonialism.

Benjamin Thompson

And one of the researchers quoted in his article said that it's kind of interesting to be able to combine genetic data with historical events to work out how this insect dispersed itself around the world, as I say its become abundant. But without using these modern genetic tools, there was no way of knowing that it's not actually a native European species.

Lizzie Gibney

And is there a reason why this cockroach is so successful?

Benjamin Thompson

Well, this one is particularly good, I think it's fair to say. So these cockroaches readily adapt to modified environments, particularly, you know, human occupied niches. They have short reproduction cycles, they're incredibly opportunist. And so, hitchhiking to a new place where people and access to food. One of the researchers quoted here saying, “that's a perfect combination of ingredients for making a species very successful in a human-shaped world”.

Flora Graham

We really had to think about this story when we put it in the Nature Briefing. It's so fascinating, and it did turn out to be one of our most popular stories of the week. But it does involve a big picture of a cockroach. And we really debated whether having an email drop into your inbox with a big picture of a cockroach on the top would be too unappealing for most people. Now luckily, in the story, there's a couple pictures and one of them is a very lovely glossy brown cockroach on a green leaf, you know, living its best life in its natural habitat, not in my kitchen. And we thought, well, this guy or gal looks like an insect we'd be more than happy to see in the woods, so we decided to go with that one.

Benjamin Thompson

I'm happy to say I've had limited experiences with cockroaches. But anyway, let's move on. We've got one more story Flora, you’re up.

Flora Graham

Oh, this is a story that I definitely would not have let us miss because this is really basic science at its best and most exciting. This is one of those rewrite the textbook moments. This is all about an element called promethium. Fantastic name, named after Prometheus of course. This is one of the lanthanide family, which is a row of 15 metals that kind of sits marooned down in the southern territories of the periodic table. It was discovered in 1945 so it's fairly new to us. And we think that there's something like less than 1 kilogram of promethium naturally on Earth. So that just gives you some idea of how incredibly rare it is. But because it's so rare, it's so mysterious. And this is element number 61 in case you want to look it up on your own periodic table at home. This is the first time that chemists have put promethium into what is called a chemical complex. Now, basically, in a nutshell, this is a compound in which it's bonded with some other atoms, and it's in a solution with water. And this kind of completes the set of these lanthanide elements. So previously, every time we've tried to analyse this family of elements as a group, we've always had to say, well, except promethium, we've never we don't know about it. So obviously, we can investigate it in other ways, but having the ability to see how it interacts with other molecules offers key information about how this element actually works, how it reacts.

Benjamin Thompson

If there's so little of it, and it is so difficult to work with, how do they go about achieving what they've done for the first time, as you say?

Flora Graham

What they did is they took a waste generated during the production of plutonium. And they harvested a radioactive isotope of promethium, called promethium 147. And they worked their magic with other things which were able to cog on to the promethium ion, which created these complexes of promethium oxygen bonds.

Lizzie Gibney

In what ways did they then prod it and poke it to try and understand promethium better if this is the first time that I've actually had it in this chemical complex?

Flora Graham

Yeah, so they used X-ray spectroscopy. And they also ran simulations in order to see how the oxygen on molecules connected to the promethium. And it's going to lead to a better understanding of promethium in general. So this is an element that is radioactive, it can provide power in certain situations like in pacemakers and potentially this research could lead to better control of the element. So you could do things like separate it from similar elements in nuclear waste, for example, in order to harness it more easily than in its naturally-occurring form.

Benjamin Thompson

I mean, that's a cool story, right? Because it, you know, we will learn about the periodic table, and we can picture it in our minds, I'm sure. But there is still more to discover about the elements inside it. And what the researchers say about this achievement? Because as you say, it's taken quite a long time to get to this stage.

Flora Graham

Absolutely. They're saying that this was such tricky, difficult work. Polly Arnold, who's a chemist at the Lawrence Berkeley National Laboratory in California called it a “tour de force”. So this is the level of impact it's having among chemists who are working on this level.

Benjamin Thompson

Well, that's a good upbeat note to end this week's podcast. And listeners, for links to all of those stories and where you can sign up for the Nature Briefing to have even more of them delivered directly for free to your inbox, look out for links in the show notes. And all that's left to say is Lizzie and Flora, thank you so much for joining me.

Lizzie Gibney

Thank you, Ben. Gonna try not to have nightmares about the cockroaches.

Flora Graham

Thank you, Ben. I'll be looking out with some promethium.