transcript

Speaker 1:
[00:24] Welcome back to the Nature Podcast. This week, the robot that could beat you at table tennis and the latest attempt to test the strength of Big G. I'm Benjamin Thompson. When it comes to humans versus artificial intelligence, there's a long history of researchers designing machines to beat humans at games. Things like chess, backgammon, Go, we've covered several of these kinds of stories on the pod before. But things are getting a little more physical with the latest paper we're covering. The story of an AI-powered robot that can beat elite players at table tennis. Reporter Jeff Marsh spoke to Sony AI's Péter Dürr to find out how and why.

Speaker 2:
[01:19] If you look at the history of AI, scientists have often used games as benchmarks. And in a way, games are very interesting because they allow us to directly compare the performance of an AI agent to a human's performance.

Speaker 3:
[01:40] Okay, great. Do you want to introduce yourself then?

Speaker 2:
[01:43] Yes, very nice to meet you. My name is Péter Dürr, director of Sony AI in Zurich and also the lead engineer of what we internally call Project Ace, which is a robot that can challenge human athletes. And I'm a roboticist by trade, and so this is really a dream project for me.

Speaker 3:
[02:07] What is it that makes table tennis per se such a challenging sport to compete in?

Speaker 2:
[02:12] So when a professional player hits a fast ball, it can reach velocities of 25 up to 30 meters per second. They also impart a massive amount of spin on the ball. So this lets players influence how the ball flies. This also causes the ball to bounce differently on the table and then causes the ball to bounce differently from the opponent's racket. And the spin that you see in professional games can exceed 150 or 160 rotations per second. So that's really challenging actually to measure in real time and this was one of the challenges we solved with our robot system.

Speaker 3:
[02:59] So that was one of the first things Peter and his team had to build for Ace, a perception system that could accurately locate the position and spin of a small and very fast moving ball in real time. I guess about as difficult as it sounds.

Speaker 2:
[03:12] And we did that with two separate system components. On the one hand, we used a number of very fast cameras that we set up outside of the playing arena. This lets us triangulate the 3D position of the ball in space. This gives us a measurement of where the ball is, roughly 200 times per second. However, as I mentioned in table tennis, spin is very important. So we built a second system, and that system uses image sensor technology that's called event-based image sensor. And that gives us the ability to see things that move very fast. So it gives us a very high temporal resolution. The optics use two little mirrors that we can move. And using our estimate of where the ball is in space, we could orient the mirror such that we can see a very close zoom-in image of the ball at all times as it's flying through the air, which allows us to measure the magnitude and also the axis of the spin in real time, which then gives our robot the information it needs to anticipate where the ball is going to go and plan how to hit it back.

Speaker 3:
[04:39] But before we get on to the robot arm and how it physically hits the ball back in the real world, we've got to go back a step because the neural network under the hood of Ace first had to be trained and it didn't learn the ropes in the real world.

Speaker 2:
[04:52] The training for our control system happens exclusively in simulation. So basically, we shoot a ball in simulation at our robot, let it do random things at the beginning. It doesn't know how to react to a ball that flies toward it, but eventually it will hit the ball and maybe be lucky enough to hit the ball back on the table. And at that moment, we can say this was a good idea to do that kind of thing again. But we also want to hit the ball fast and with lots of spin, either topspin, backspin, sidespin, etc. So we add all these components to the reward function that we use to train the neural network. And by repeating this for hours and hours and hours and hours of simulated table tennis, we end up with a control system that can play at a weak level at first, and we get stronger and stronger performance.

Speaker 3:
[06:02] So allowing the neural network to effectively experiment in simulation is key to teaching Ace how to play with its unique arm. But obviously that arm is a relatively heavy, very fast moving lump of material, so you can't just give Ace completely free rein, especially when it's eventually going to play against soft human opponents.

Speaker 2:
[06:21] So what we did is put together the neural network with more mathematical optimization based layer of control that gives us the opportunity to have an escape plan. So what we do is we let the neural network do whatever it wants to do. But we constantly also generate a plan how to stop the robot safely. And then if we predict at any point in time that the next action that the neural network wants to do is going to lead to a collision in the future, we can just ignore that command and instead fall back to the escape plan that we computed in the last step.

Speaker 3:
[07:14] So that's the perception and control systems covered. The final component that allows Ace to compete at an elite level is its hardware.

Speaker 2:
[07:22] So, when we started the project, we bought a few industrial robot arms from different manufacturers, but we quickly found that there is nothing on the market that would let us play at the level we wanted to play. So we decided to build our own custom robot platform. So it has an arm with six joints, six degrees of freedom, that's on a linear stage that let it move in the X-Y plane. And we managed to build a robot that can actually move the racket at over 20 meter per second velocity. We also put a little cup at the end of the arm that lets us toss it for the serve.

Speaker 3:
[08:12] All right, so that's how you got everything in place, Peter. Obviously, Ace was designed to compete against humans. So tell me about the human trials with Ace.

Speaker 2:
[08:21] Yeah, that's very interesting and it's a long and tough experience. So we started playing table tennis in very generous terms quite early in the project. So we managed to build something that could bounce the ball back. And some of the researchers and the engineers practiced enough to be able to shoot the ball toward the robot in a way that the robot could play back. At some point, the system became more and more competitive and we started to beat beginners. And this was a major milestone for us when we were able to challenge the beginners and then beat the beginners convincingly. And I'm very happy to report that just last month, we managed to beat a woman who is in the top 25 in the world ranking and also a man who is in the top 200. So our robot continues to improve. We invited a former Olympian to watch the games. And he commented that one of the shots was so unusual that he thought he'd never seen a professional player play anything like that. But he also thought that players might actually try that and use that to their advantage going forward. So that was a really interesting observation that we loved to hear.

Speaker 3:
[10:05] Impressive as Ace is, and it really is, you should go and check out the film on our YouTube channel. I wondered if Peter thought that the ultimate goal of this sort of physical AI system was to beat humans at sports like table tennis.

Speaker 2:
[10:17] This is very much a research project, and it was built on the hypothesis that it would challenge us to push the individual component technologies to their limits, rather than solving table tennis as an application. I think the sensing and control techniques that we demonstrated with Ace have other applications. The ability to combine this novel low latency perception with the adaptive decision-making and physical control that we invented for Ace can also be used for other applications of robots, like in industrial settings, in service robotics, et cetera, et cetera.

Speaker 3:
[11:08] Before I finished up, I wanted to get some outside comment on Ace to see what others in the AI robotics community made of it.

Speaker 4:
[11:16] My name is Esther Colombini. I'm a professor of AI and robotics at the University of Campinas in São Paulo.

Speaker 3:
[11:24] And presumably you haven't been involved in sort of like table tennis robots before.

Speaker 4:
[11:30] No, not table tennis. I've been involved with soccer player robots, but not table tennis.

Speaker 3:
[11:37] Okay. So here's another researcher in this field, dabbling in sports. What gives?

Speaker 4:
[11:42] If I can have robots playing soccer well, I would say that they can interact with humans safely because you wouldn't want a team of robots playing against a team of humans and injuring them. And you don't want a robot in your home injuring the people that are in the home. So I don't need robots to play soccer. I don't need robots to play table tennis, but I need the skills and the abilities of these robots learned in these environments that are easy for us to see how they are evolving. So sports are just a proxy from what we want. And a very good proxy, because if you want robots to work in environments where humans are living and that require interaction, you need the skills that usually you can learn in sports. So sports are really a good place for that.

Speaker 3:
[12:34] Okay, and is there anything you would like to see in robotics in the next sort of 10 years? Maybe is there something more utilitarian for old people? Or what kind of gets you up in the morning?

Speaker 4:
[12:46] I think that assistant robots in terms of helping people like with the loss of muscles and so on and movement, people with disabilities, the products of robotics, they are very good for society. They can be really bad as well. But I would say that the helping part of robotics is what make my eyes look to different solutions. And I would really like to see this being like the way that robotics is going to evolve in the next years. So just for helping people.

Speaker 1:
[13:31] That was Esther Colombini from the University of Campinas in São Paulo, Brazil. You also heard from Peter Dürr from Sony AI Zurich in Switzerland. And the reporter was Jeff Marsh. As he mentioned, if you want to see a large robotic arm swinging around on a large mechanical base and performing some quite impressive table tennis shots, then head over to the Nature YouTube channel. And we'll put a link to that in the show notes, as well as a link to the paper. And a News and Views article, written by Esther. Coming up in the show, how the latest measurements of the strength of gravity has thrown up some surprises. Right now, it's time for the research highlights with Dan Fox.

Speaker 5:
[14:17] There's a mystery around Venus. Just beneath the planet's acidic clouds lies a 20 kilometre thick layer of haze. And while researchers have known about its existence for a long time, its origin has remained unexplained. To get a better picture, researchers modelled Venus' atmosphere and found that some properties of the haze can be explained by a steady influx of cosmic dust. This origin would explain how the haze has sustained itself. And the model suggests these tiny particles from space could explain the Venusian atmosphere's ability to absorb ultraviolet light. The authors suggest that understanding cosmic dust could be crucial for our study of other planets with thick atmospheres. Dust off that research in Nature Astronomy. An outbreak of plague in Switzerland between 1665 and 1670 fell hardest on low-income workers who could not afford to stay home even during an epidemic, echoing disparities seen in the COVID-19 pandemic. Researchers studied the skeletons of 15 people buried at a cemetery on the grounds of what used to be a municipal hospital in Basel. DNA samples revealed that at least five of the people had been infected with the plague-causing Yersinia pestis bacterium. The best-preserved DNA sample was recovered from an individual buried with a clay pipe, and an artisan's insignia on the artifact helped to date the burials more precisely. Those buried in the graves were young. Eleven of them died before the age of 20, and a large proportion of them had lesions in the spine or the shoulder, injuries that suggest that they performed physically demanding jobs and probably belonged to lower social classes. You can read that research in full in Antiquity.

Speaker 1:
[16:21] Next up on the show, we're talking physics, and specifically working out the strength of Big G, the gravitational pull between two objects. And a new experiment has thrown up some surprises as to what that strength might be. Now joining us to discuss this is a, well, a regular G, Nature's Lizzie Gibney. Lizzie, thanks for being here.

Speaker 6:
[16:43] I like that. Thank you for having me.

Speaker 1:
[16:45] Yeah. Well, listen, I've given a sense of Big G. Maybe you could tell us a bit more about it.

Speaker 6:
[16:49] So Big G is the least well understood of any of the fundamental constants of nature. So when you're trying to work out the gravitational strength between two objects, you'll have like the sun and the earth, and then you'll do Big G times the mass of one, times the mass of another divided by the square of the distance between them. It's everywhere in basic physics. But we actually really don't know the value of Big G very well at all. It's 6.673 or 4 times 10 to the minus 11, depending on who you ask and which experiment is exactly. And whilst, you know, that sounds maybe still decent, like only having three significant figures on a number like this is quite extraordinary compared to everything else, you know, that we've been able to know with great precision in nature, in the world.

Speaker 1:
[17:37] And as you say there, it's the least understood. Also, it's not actually that strong, right? It's a pretty weak force. I can overcome the gravitational pull of the earth just by lifting up this coffee cup here, which this one's over here. So that makes it quite difficult to figure out.

Speaker 6:
[17:50] Exactly. So gravity is so much weaker than every other fundamental force. So you've got the strong nuclear force, weak nuclear force and electromagnetism. They are trillions of trillions, perhaps even more times stronger than gravity. So, you know, when you are picking up that coffee cup, you're fighting the entire gravity of a planet, the size of earth, and you can do it very, very easily. So that is part of the problem is it's so small. It's also a problem because what we're trying to measure, if you're trying to measure this strength, you have to, in an ideal world, you protect that experiment from all other influences. But gravity just goes through everything. You can't shield against gravity. So instead you just have to factor in every other influence. And that just makes these experiments extremely hard to do.

Speaker 1:
[18:37] Right, and folk have been trying to work out the exact number of Big G for hundreds of years.

Speaker 6:
[18:41] Yeah, it goes back to the 1790s and they did it by putting two masses like on the end of a rod that was then hung by a wire. And then they had these masses that were sitting outside of those and there'd be like a pull between them. And as they were pulled, the rod would twist and they would measure the twist on the wire that it was hanging from. And that's the way of trying to get a value for Big G. And experiments like that, a lot more high-tech, a lot more precise are still happening today. And what I've been writing about is this experiment to do exactly this. So because measurements of Big G are so hard over the years, the values have just kind of scattered around and people's error bars are not overlapping like you'd hope. So the value that physicists around the world have agreed on is kind of a mishmash between all of them where you factor in as much as possible how well the experiment was done in their own judgment of their uncertainties. But there was one number that was from a very prestigious lab, the International Bureau of Weights and Measures in Paris, that was a bit of an outlier. And so a group at the US National Institute of Standards and Technology, NIST, they tried to, they were kind of almost made to by the international community, to replicate this experiment. They said, take the exact kit, ship it across the Atlantic, redo this experiment and see what value you get because it's just not really acceptable to have a number that no one can agree on.

Speaker 1:
[20:06] And of course, replication in science is something we talk about a lot and they've really gone the distance here to try and do it. And this experiment went on for what, 10 years? Is that right?

Speaker 6:
[20:14] Yeah, it was a full decade. And what else was, I thought, very cool about the way they did this, it was completely blinded. So they got somebody who is independent to come in and actually take an offset off the masses so that the experimenters themselves didn't know what numbers they should be getting from their experiments. So they didn't actually know what masses they were using. And so this meant that they were going to do the calculation. There was no sense that they were going to keep looking for discrepancies until they reached the number they knew they should be getting, which would be such a human thing to do. Obviously, no one would do that on purpose, but it's incredibly difficult not to do that if you know what you should be aiming for. But they were completely blinded. They didn't know the masses they were using. And so it all came down at the end. They had the experiment. It was kind of an envelope that they opened to see what the offset was and to figure out actually whether the value they had matched or not.

Speaker 1:
[21:01] And don't leave it in suspense then, Lizzie. Did the value match that of the French experiment?

Speaker 6:
[21:05] It did not. It was very, very far out, in fact. And you can see that as good or bad, really. So the experiment was also really thorough. So it found some potential reasons for the discrepancy between the French value and the US one. Like some of the exact shaping of these masses was not perhaps quite so precise in the original version. And there was like a pressure from gas inside the device that they also didn't factor in. But the upshot of it is that it seems like the French value was just a bit too high because this value was quite a lot lower. So on one hand, they don't match, but maybe that's good because maybe because they found these potential sources of error, that actually means that the newer value is a better one. It's good to have that explanation for why the previous value may have been a little bit of an outlier. However, the new value also doesn't match with the world's agreed value, which is created by bringing together all of these different dozen or so experiments that have been done over the years around the world. It also sits quite outside of that. So it definitely isn't the end of this. So every four years, there's a group called Codata who decide on what we're going to agree on as these values, and they will include this. This is such a really, really meticulous, well-done experiment. And so this will definitely, definitely go into that calculation. It will be a new data point. But whether it's really a useful data point from that perspective, I don't know, because again, it doesn't match. It's quite far from the other one. No one really knows exactly how this is going to affect things. All we do know is we still definitely, absolutely cannot measure big G with the extreme precision that we'd like to.

Speaker 1:
[22:38] Right, because you say the French value and this reproduced US value differ massively. But really, it's only what, 0.02 something percent, which on the face of it isn't much. But in terms of physics experiments, a millimetre is a mile, right?

Speaker 6:
[22:50] Exactly. I mean, that's enormous when you compare it to the precision with which we know other fundamental constants. Yeah, the error is about, I think, it's 1.5 thousand, something like that. That's pretty big.

Speaker 1:
[22:59] And so what does this mean overall then for our understanding of the universe, right? If we still can't get an idea of what the strength of gravity is, does that muddy things a little bit, do you think?

Speaker 6:
[23:11] Yeah, don't worry, we're not going to have satellites falling out of the sky or anything like that. In fact, this is one of the reasons probably that we don't know Big G that well is because it's not that useful. So metrologists, people who really care about precision have been studying it, but it's not had the big bucks thrown at it because actually, most of the time if we are trying to calculate the orbit of planets or yeah, the trajectories of satellites going around Earth, you actually just need Big G times the big mass in that calculation, and that you can measure with much greater precision than we know Big G. So in all those cases, you don't actually need Big G and there's a lot of other instances where it's going to just like cancel out. So most of the time, it's not actually that useful. It might be useful in the future, having said that. So cosmology is getting a lot more precise, like looking at the origins of the universe and how structures formed, and all of that would have been affected by the strength of this kind of universal gravitational pull. So there could be a day where the numbers we're getting out of cosmology are so precise that we need to have equivalent precision from Big G. And also, at the moment, Big G is just a constant. We measure it, but we don't have a way of calculating it from first principles. If someday theoretical physicists figured out there is a value that Big G should be from theory, then of course we'd also want to have a really, really precise number to match that with. So it might not stay useless forever, but for now, I asked Stefan Schlamminger, who was the lead scientist behind this big replication effort, and I said, you know, why do you do it? And he said, well, it's like Everest, isn't it? Why does anyone climb Everest? It's because it's there. It's because of the challenge.

Speaker 1:
[24:48] I mean, I do get the sense from your article that the researchers themselves are aware that maybe working out Big G isn't a priority for the physics community, and they're quite sanguine about the results and their efforts to get them.

Speaker 6:
[24:59] Absolutely. And, you know, a lot of the time, these experiments have piggybacked off other experiments people have been doing. This is very rarely the only thing that a physicist will be doing. They'll be doing a bunch of other experiments and have the Big G as the thing that they just do, you know, because of pure passion on the side. Somebody called it the hardest experiment in the world. It is hard to convey just how difficult these experiments are. And again, for a certain type of person, that makes it very appealing. And once you've done this, you could do anything, right? Like the physicists are doing these kind of experiments are often metrologists, like measurement scientists, and they're going to apply these incredible skills that they've honed to other areas of metrology.

Speaker 1:
[25:38] Wonderful. Let's leave it there. We'll put a link, Lizzie, to your article in the show notes, so listeners can read even more about Big G and the quest to figure out what it is. Lizzie Gibney, thank you for being here.

Speaker 6:
[25:47] Thanks very much, Ben.

Speaker 1:
[25:48] Nature's Lizzie Gibney there. And that's all we've got time for this week. Don't forget, you can reach out to us on social media. We're at Nature Podcast. And of course, we're on email too, podcastatnature.com. I'm Benjamin Thompson. See you next time.