How good are robots at soccer?

[00:00:00]Friends, the 2026 World Cup is here. And that's big news, of course, for the sports world, but also big news for the robotics world. Why? Because of viral marketing campaigns. That's right.

[00:00:12]>> The Atlas robot making a halftime appearance.

[00:00:14]>> If you're a humanoid robotics company or a research lab and you want to get attention for your robot, what better way than make a soccer video. Show your robot doing something soccer related.

[00:00:23]Kick a ball, juggle a ball. And it's also a chance for us to watch these videos and be like, "How good are robots at soccer? What are they good at? What are they not good at? Why is that? What are they doing?" And look, I want run a robotics research lab. I I love doing robot videos. This is my thing. I love making them. I love watching them. So, let's watch a few of these. It should be a good time.

[00:00:48]Doing a fancy kick.

[00:00:53]And It's good shot. Nice. Okay, so it's pretty cute. It's doing this crazy little kick. So, this is the Atlas robot by Boston Dynamics, which is this famous robotics company who has produced some of the most compelling robotics demonstrations in the history of the world, if you ask me, and just put them out on YouTube. And this is a fullscale humanoid. It's about 6 feet tall, 200 lb. So, it's it's got some heft behind it. So, like stay out of its way. You don't want it bumping into you. One thing that's really cool about this robot is that a lot of its joints will rotate around 360 degrees. And that's not easy to put on a robot because it actually creates a problem for its wires. So like imagine you have like a wire here for your robot, right? It's it's transmitting power or communications or something and you have to have this go through a joint. If it has if it spins like more than once, what'll happen is it'll start to like, you know, get bunched up and then as a result we'll kind of yank the wire. Um that's a bad thing. So you have to design it special. So with special rings that transmit power, so it makes it more challenging to build. Cool stuff. So they're doing a trick here that is increasingly common for humanoid robots these days. It's called motion retargeting. The idea is you take a motion of from a person doing something and then you retarget it to a robot. So it is mimicking it. It sounds like it makes complete common sense that people should have been doing this for a long time, but it's actually very recent. For the longest time, it just wouldn't work if you just took a human performance and just put it on the robot because the robot would fall down because the robot weighs different than a person. It's got different proportions than the person.

[00:02:25]It wouldn't know how to balance. And what we really managed to do is find a way to make the robot both mimic the performance and not fall down and like only like the last three years or so.

[00:02:34]So, what's interesting here is that they apply this to a much heavier robot. And that's much more challenging to do because lighter robots can whip their arms around much more quickly and their legs. That allows them to catch themselves. The fact they're doing motion retargeting in such a fluid way with this robot is is is quite impressive. I I I do like that. Now, what does that mean? That means that Boston Dynamics had a person do this trick, this what they call the ghost raona. This this this I don't I don't know, soccer. I is is that a thing to do this trick and then they captured it with a motion capture suit that will measure all the motion all the joints and then they target it onto the robot.

[00:03:11]What that means is that they train the robot in a computer simulation to say do that move and don't fall down. And they will try that thousands and thousands of times on the computer and then take the resulting policy that's like the strategy and put it on the robot. Now what's happening here? There's a couple things that are a little bit unclear.

[00:03:31]Like for instance, is that robot actually seeing the ball? So like here we got the ball. Okay, I don't have a soccer ball. This is a texture ball for babies. This is the closest thing I can come to for the World Cup. So, is the robot just swinging at the ball, knowing where the ball is, meaning it can see with the cameras and aim at it, or are they teeing up the ball in a precise location, and then the robot is just swinging through in the exact same way each time trying to get the kick, which would make sense as sometimes it's missing. Or it could be using its cameras to spot it using image segmentation. It's not fully clear. One little nitpick about the video is that they make it seem like when the robot's failing and then it's getting better that the robot is learning, learning, taking the experience and then changing what it's doing as a result. Almost certainly isn't. It's probably just doing the same routine over and over again and they're just taking the ball and just kind of putting in the right place. You know, they're trying to just line it up better. Most of the time, these robots are not learning when they're operating. All the learning is probably done by the time the robot is turned on. There are some research labs, of course, that work on this exact problem. They want continual learning.

[00:04:39]It definitely exists, but the vast majority of these neat robot videos, they're not learning. It's It's cute.

[00:04:45]It's cute. Uh but it's a far cry from it doing that in this one stage demonstration than from like putting it on a soccer field. It's like, oh, it's going to run over and it's going to Oh, there's the ball. Oh, what am I going to do? Do the ghost raon, you know, and knock it into the goal. It We're a ways away from that.

[00:05:05]Okay.

[00:05:11]>> Sure.

[00:05:12]>> Okay, friends. This is fake. This is a visual effect. It's not a real robot.

[00:05:16]So, what someone did here, they took a video of two people doing a shootout and they literally just replaced one with a robot using some visual effects tool, maybe some AI, generative AI tool that automatically does it. It's fake. And but understandably, it's going to be harder and harder to tell the difference between a real video and a fake video because the real videos are also capturing human performances and replicating them. A lot of the subtleties of the smoothness of this video kind of give it away for me, but it's going to be hard. And yes, you can do some like amateur video effects analysis, be like, "Oh, the shadows look wrong and blah blah blah." You can do that, but be careful because I've definitely seen people microanalyze real videos of robots and say it's fake.

[00:06:00]Unless you really know what you're doing in terms of analyzing fake videos, you know, from a visual effect standpoint, it's a dangerous game. You can get it wrong. So, it it might actually make sense to rely upon experts to tell you if it's real or not. Hey, you could you can ask me in the comments. I'll tell you if it's real or not to the best of my ability.

[00:06:33]I love the way these things are cut. Um, you can, if you've done robot videos like this, you definitely see the way that the shots are set up. It, by the way, cool. This is this is this is neat.

[00:06:42]A lot of this stuff is like cool, but what are the limitations, right? Um, so this is robot juggling. This is the Bolt humanoid robot, and I've seen it around before. So, this is pretty impressive because one way that you could have done this is you just put the robot through a pre-established like kneeing routine for the juggle. You could just say, "Oh, left foot, right foot, left foot, and then sort of just put the ball on the knee and then just have it like hopefully you get like three or four bounces off of it." And sure enough, you do see that like there's a quick cutting. It doesn't get any more than maybe four, technically, maybe five juggles, which is good, but probably better than me, right? Um, but it is very clearly reacting to the ball because you can see it kind of lingers on one foot a bit longer when it's waiting for the ball to come down. So, there's a lot of what we call feedback control. It's aware of the ball and it's reacting to that in order to juggle. So, that's good. That's good stuff. Now, how is it doing that? How is that getting that information? Is it using its cameras to do that? It's not fully clear, but there are a lot of hints that it isn't. For instance, if you look in the background right right here, like you can see these are motion capture cameras and they are tracking for sure the robot's position. You can see these little markers there. Those are reflective markers to tell where the robot's feet are. So, the way these cameras work is they shoot out infrared light and if they hit a reflective surface, it bounces back to the camera and they use that to report the position of whatever it is they're tracking. And you have cameras all around the robot so you always have a good view of it, right? So, it's definitely tracking the feet of the robot. It's probably tracking the ball position, too. In fact, one thing you'll notice is that the ball is all blacked out. You know, it's it's all covered up with black.

[00:08:17]That's probably to make sure it doesn't throw off reflections that you don't intend, and it would throw off all the measurements of where the ball is, which would be a bad thing. Right now, I don't quite see the markers on the ball, but there has to be some reflective tape on that ball somewhere. Otherwise, why are they using motion capture for this? And the cameras are probably not just recording the data of where these things were for the test for posterity sake. is probably reporting them to the robot in real time up to a thousand times per second. You can do it. So the robot can use that information to like, oh, that's where the ball is. I can go I can hit it here, hit it here. So more evidence that it's actually reacting to the ball position and helping it juggle. So, you know, good stuff. But you can definitely tell that they got like here's our four takes that we got of the robot doing the juggle from these angles and everything else like the ball rolling up to the robot and then under the foot. I mean, that's all sort of a stage thing just for, you know, dramatic effect. No, no shade on that. But the real juggling there. Hey, not bad. Also, this is a precarious setup. Look how ready to go they are with those crash pads right there to catch the robot. So, yeah, they probably just barely got the footage they needed. So, it's neat stuff, but it's obviously not going to be able to do this on the soccer field without those fancy cameras.

[00:09:27]Although they don't have moves like Messi and they can't bend it like Beckham, these synthetic strikers are helping researchers at Google DeepMind to rewrite some of the rules of robotics. So, I love this video. I love this result, this is so cool. It came out about 2 years ago. And while this might look a lot to you like two little toddlers, and sure enough it does. But there's a lot of potential here. And one thing I love about it, other than the fact that it is a real soccer match.

[00:09:52]These two robots are actually playing soccer against each other and it's fully autonomous. There's no one with a remote control telling it where to go. What I love about it was how it was coded to play soccer or more specifically not coded to play soccer. I'll tell you what I mean. So, they're using a technique called reinforcement learning. And it's a really hot technique right now where you take a computer simulation of your robot and then you have to give it what we call a reward function which is something that you want to train your robot to do. You want to say hey you did this well you get rewarded you get like a treat right and a lot of these robots are using this including these robots that are doing these nifty kicks and things like that but the thing that they're being rewarded for is mimicking the person. Did you mimic the person more closely? If so good good reward right? where these robots, all they're being rewarded for is winning. They're just told to win the game. And everything else they figured out, the kind of the goofy way they kind of put their arms out to the side, the way they kind of toddle around, that's all figured out on their own by the robot.

[00:10:52]It's really incredible what they were able to figure out. Again, they figured it out on the computer and you put it on the robot itself, but it's impressive nonetheless. Now, a couple of little caveats here. They had to do a couple of things to make the robots figure it out.

[00:11:05]And you can go into the paper and you can read about them. One of which is that when the robot falls down, it just picks itself back up. It's just told to do that. And this is how you get up if you fall down. That's it, right? And that helped it learn. Also, they kind of had to encourage it to run toward the ball and not run at the other robot.

[00:11:23]That's another way they managed to get the robot to do something to kind of guide it in that direction. But the rest it's figuring out, and that's really impressive. I love that. One thing about this setup is it's not using any cameras. It's using motion capture around the arena to tell the robots where the other robots are, where the ball is, where the goal is. That way, it doesn't need to figure that out on its own because visual processing can be hard to do on such a tiny robot. It's not impossible, it's just tricky. And they want to show off how the robot would think using this reinforcement learning technique. And it's really a great result.

[00:12:01]Everyone loves robots falling down.

[00:12:03]Everyone laughs. It helps they look like children, I will say.

[00:12:11]Bring out the stretcher. Oh man. So, uh, adorable, right? And you see these robots falling over and it's always funny. I I always feel bad for the roboticists involved, but it is kind of objectively funny. One of the funnier things at the end is that you bring out actual stretchers. You use things for humans to get these robots off the off the field. It's for instance, it's also very common when you're trying out new robots to use an actual patient lift from a hospital to prevent the robot from falling down that are designed to prevent patients from falling down. So that's adorable. This So this is robot team soccer. And robot team soccer is not a new thing. It's been around for almost 30 years. It started way back in Japan in like 1997.

[00:12:51]And the goal of the Robo Cup was to by 2050 have a robot team that could play against a human team. Look, obviously there's a long way to go here, but it's pretty hard what they're trying to do.

[00:13:03]There's very little help that the robots get from the outside world. All of the planning, all of the moving, all of the vision has to be done by the robots. A lot of these other videos, they'll get like a little help. They'll get like a motion capture device to say, "Oh, the ball is there. The opponent's there."

[00:13:21]But here, the robots have to do all the vision. They have to know where the goal is. Sometimes the robots will kick the ball toward the wrong goal cuz they get lost and confused. That's that's something that will happen.

[00:13:30]>> And now there's two red ones being confused. So that's a good chance for the blue one. One confused.

[00:13:35]>> So there's some algorithm that's saying, "Hey, go here. Go here. Go here." And then the algorithms on board the robot are trying to keep it standing up and from falling over. And you might wonder why it's like shuffling around like that. That's kind of moving its feet constantly and kind of that way. It's actually a stability tactic to make sure it doesn't fall down. Because the idea is that if you have one foot in the air for only a little bit of time, you can't fall much to the side in that direction.

[00:13:58]So your foot comes right back down.

[00:13:59]Stabilize it. Now you go back to the other side. So it's not very efficient.

[00:14:03]You probably burn a lot of energy on the robot that way. But it does keep the robot up at least a little better.

[00:14:20]So that's pretty impressive that robot dented the wall. And look, robots are getting really powerful. But the question is, how powerful are they really when compared to say people, right? Because you can't really take into account the fact, yes, it dented the wall, but like how thin is that wall? I mean, this is an office space.

[00:14:36]It's not an indoor soccer arena that's designed to take an impact, you know?

[00:14:40]So, what we need to look at are the kicking speeds of the robot. And helpfully in the video, they actually did like a radar gun test and showed that the robot is kicking at about 70 kilometers an hour. And I look at the video more carefully to see if that was like a really kind of a reasonable speed based upon the apparent distances and timing involved. And it seems about right. So, I I'll trust the number. And where does that compare to people and hopefully people of similar size to this robot? Thankfully, know a lot about this robot. This is the booster T1 series humanoid. You can buy these from Booster Robotics. It's a Chinese robotics company. And it's about 30 kg, 66 pounds, which is about the weight of a 10-year-old. So, how fast can a 10-year-old kick a soccer ball? Well, helpfully, people have actually done studies on this. They've actually measured Brazilian soccer kids, like elite level academy trained kids for soccer, see how fast that they can kick the ball. And it turns out this is like right on par with how fast kids kick a soccer ball. Like really good soccer kids. That's impressive. It is performing on par with really good human performance. And this is something that was predicted pretty recently in a paper just before this major humanoid boom hit us in like 2024 that was analyzing why robots were not yet as good as people at doing various locomotion tasks, running, you know, maybe soccer, could be one of them, right? And the conclusion of this study was that the robots are actually strong enough in terms of their actuators, their motors, the materials, their sensing, every component of these robots is good enough to beat a person at stuff, right? However, the thing that they weren't able to do is coordinate all these things together. The control was the problem. And sure enough, in those couple years since, the control has gotten a lot better. The advent of deep learning has made this a lot more feasible. And sure enough, now you're seeing it as predicted. Now we have the components and we have the smarts. These things can perform as well as people.

[00:16:34]And this is this is one such example of that. So these robots, they're pretty powerful. I mean, they don't leave us in like dust in terms of capability, but like they're like strong, capable people when they work reli. They're not always reliable, but when they work, they can do cool stuff.

[00:16:52]So a lot of these videos are pretty impressive. you know, they can do things in isolation with some caveats attached, but lot of it's pretty good. And it kind of raises the question, could we have an entire team of robots competing in the World Cup? RoboCop says they want this by 2050. And they said this way back in like 1997 when who knows what what the path of robotics would be back then. But right now, looking at it today, that's not crazy. It really is not. If you look at the robot hardware, these physical robots, they're strong. They can be made agile. They can jump around, run around, and that was a big limitation for robots in the past. They just were not capable.

[00:17:29]But that's no longer the case. It's really the algorithms that have to get better. And they are getting better.

[00:17:34]They're getting better on an individual level. These robots are better at balancing. They can do some skills, but they're not yet integrated with vision super well. There's some preliminary work that we can talk about. There's a lot more videos, by the way. There's a ton more videos out there of some preliminary research experiments that have been done. some papers that were literally published like days ago. So if you want to know more about robot soccer, we can do some more. If you have some videos you find are interesting, just let me know. But if the algorithms get better both on the individual level and at the team level especially, that could potentially happen. I mean, think about it. You have an actual robot team playing against people and you know, the robots score a goal. You know, they high-five each other, they celebrate, they do some kind of crazy dance. Is is that a thing they do? Do they do the crazy dances at the World Cup? I I've seen crazy dances. I don't know if they do it at the World Cup. I I don't know much about the World Cup. I'm sorry. I'm I'm learning fast. Look, this was the closest thing to a soccer ball I had in my home and it was for my baby. Okay?

[00:18:30]You know, let me know. Do they?