ΑΙhub.org
AIhub coffee corner: Are humanoids the future of embodied AI?
The AIhub coffee corner captures the musings of AI experts over a short conversation. This month we ask whether humanoids are the future of robotics. How has the technology evolved recently, and what can we expect in the years to come? Joining the conversation this time are: Joydeep Biswas (University of Texas at Austin), Sanmay Das (Virginia Tech), Rina Dechter (University of California, Irvine), Ioannis Erripis (AIhub), Sabine Hauert (University of Bristol), Ella Scallan (AIhub), and Marija Slavkovik (University of Bergen).
Lucy Smith: There have been huge advancements in humanoid robots over the past couple of years. Today’s topic is particularly timely, as RoboCup 2026 recently took place. I thought it would be interesting to start with an overview of recent developments at RoboCup.
Joydeep Biswas: This all started snowballing a couple of years ago, and we suddenly started to see this huge inflection on the capabilities. New humanoid robots were showing up at RoboCup that didn’t fall down as much. Suddenly we were able to deploy robots without humans following them around. I want to highlight that this is not just one key technical innovation, it required the confluence of a wide number of things: much better hardware, significant advances in algorithms, and advances in how we train and learn. Once upon a time, humanoid robots required classical control. However, most modern humanoid robots are using learned controllers for walking. So the rise of doing sim-to-real, like training these models in simulation and transferring them to the real world, are new algorithmic innovations.
This year we had 55 humanoid robot soccer teams at the international competition, which is a lot for an international competition. We had a demo game of 11 vs 11 humanoid robots. Every year we have the RoboCup trustees play against the robots to see where we’re at. It used to be trivial. I won’t say it was hard this year, but I will say it’s getting less easy for the trustees as the years go by.
RoboCup has recognized that we are at the inflection point, and at the trustees level we’ve been thinking about how to restructure our competitions to focus on these big advances coming up. We’ve simplified our soccer leagues significantly – in 2028, there will only be humanoid league soccer. Once upon a time, we couldn’t have a fully integrated humanoid robot capable of doing everything. Even though it’s not perfect today, we have the ability to integrate control, perception etc. We feel that from the scientific perspective, it is worth exploring having everything integrated into one robot form.
Sabine Hauert: I started out excited about robotics because of RoboCup. I was in Manuela Veloso’s team doing the aibo (robot dogs) league. And that got me excited about robots working together. And then from there, they went on to the Nao robots and then to the humanoids. I guess this is a way to shake things up and to get it to the next level.
Generally, I think the humanoid shift has been fascinating. And looking at ICRA and IROS, just the sheer shift in what’s on that exhibit floor over the years is really quite drastic. It used to be all drones and then it was all robot dogs and now it’s all hands and humanoids. There are also some of the more industrial robots that have been there for a while. But that shift is very striking because you walk around and there’s just so many of them. And at the robot parade you just see all these humanoid form factors and it’s very exciting from the hardware perspective. In addition there have been beautiful advances in the control to make that hardware and shape form work. I’m very excited for other robots, actually, not necessarily humanoids, because some of that technology can be transferred.
I don’t know that we need humanoids for everything. I’m still very much in the category of people that think we need task-specific robots so that we can build them for the things that they’re meant to do. Maybe we get a general form factor not through having a human form, but by having lots of different robots with different form factors that work together, giving you the same capabilities as a human, but through multiplication rather than putting everything in one agent. But humanoids are part of the ecosystem, and I think they’re boosting the field, and so there’s a lot to gain from that.
Ella Scallan: Are humanoids the most effective shape to get tasks done?
Joydeep: No! The argument has been that everyday environments are designed for the human factor. So if you could have a working robot which has a humanoid factor, they could be a drop-in replacement and be able to use the tools and the environments which humans can. There is one big presupposition there: that we have functioning humanoid robots. As of today, that is absolutely not the case. In fact, if you want to have something work well today, there are very few cases where you could argue that the humanoid robot is the best solution. I think being a researcher in this area, it’s great to actually think a little bit about how we get to that eventual goal. Humanoids are rarely the correct solution for an application. Simple tasks, which are what we would call kinematic planning (for example, they are position-based – you want to pick up a block and put it somewhere else), are better with a wheeled robot, because the base is stable, and much easier with simpler planners, which have fewer degrees of freedom. You add a humanoid, suddenly the complexity increases. Now you’re not just worrying about position, you’re worrying about position and forces and velocities and accelerations which has much higher dimensions for planning. And it’s just not very easy to do. There are many more reasons too, that’s just one example.
Marija Slavkovik: Can I give a historical anecdote here in response to that? A humanoid shape is not necessarily the best thing. Douglas Bader was a pilot in the Royal Air Force in the Second World War. He was considered to be a better pilot because he didn’t have legs, which he lost in an accident. And because he didn’t have legs, he was able to better handle high acceleration. There are some pretty horrible examples of children being used because their smaller bodies were better suited to the task, like kid jockeys on camels, children in mines, or in weaving. Of course, these are all examples of people, but their bodies are not what you think about when you say humanoid robot.
Sanmay Das: I mean, it’s an interesting set of questions. I was wandering around one of these big conventions recently, and it was all about AI and robotics and very government tech/dev tech-focused. But one of the things they had that people were really interested in was this robot barista. You could ask it to make a coffee and it was going around doing things with the arms, but it wasn’t producing particularly good coffee from what I could tell. And I was thinking I have a semi-automatic espresso machine that I press a button on that does a really good job of making good coffee. I don’t see why we need this extremely complicated thing. But at the same time, people like it, right? How much of this is that people are more comfortable engaging with robots in this form? I think that one of the main things that we need to be careful about is anthropomorphization. I think for robotics, especially as they become more and more capable, this is really going to become a question. People will anthropomorphize these humanoid robots in ways that they will not your Roombas and your robot dogs and so on, and do we want that? And I think that that’s going to be an interesting question to think about.
Ella: Even creating humanoid robots is an act of anthropromorphization.They aren’t the most effective shape for the task, so we’re sacrificing some function for form.
Marija: Speaking of what we want, I’ve seen videos of robots that roll into a bathroom and clean it. They use the same brush for the toilet and the sink, but that’s a separate problem. My question is: how realistic are these? Because they’re just promotional videos: “invest now, and in one year, these are going to be real”. I even see some folding laundry, which is the holy grail, right? Robotics wise, are we there yet?
Joydeep: I think you bring up an excellent point, which is truth in advertising. We are definitely seeing a lot of new capabilities that we didn’t have before, but we have to be honest about what they are and are not capable of. For many actors, it is in their best interests to promote the best of their platforms and hide the worst just because of the financial incentives. If you think a little bit about what went into setting up the demo, how many takes did it actually take you to do it? How over engineered was it for that very specific scenario? It worked with a particular brush, what if I gave it a slightly different brush? What if the light was turned off and it was on? And I can tell you the answer here is that none of these solutions are really robust. They’re definitely not general purpose. The best advancements are definitely over-optimized to very specific scenarios. Also the ways in which they’re being trained are just not scalable for doing the wider broad set of things that we want to do. So I think as a community, we need to demand and require truth in advertising. And the downside of this huge influx of companies in the space is that the incentives for scientific advancement are not necessarily well aligned with the advancement of companies today.
Sabine: I think there’s a sense that people think it’s been done because they’ve seen it in a video, when a lot of it is either remote controlled or very fine-tuned for the scenarios that they’ve been designed for, as Joydeep mentions. There isn’t a “get your humanoid out-of-the-box and have it doing something useful in an environment it’s never seen before” because the data’s not there, it’s not been trained there.
Joydeep: Yes. Even for a wheeled robot, we don’t have that capability.
Rina Dechter: Thinking about one use case, robotics for social interaction with humans, for example if you wanted healthcare assistance or similar, psychologically it might give a better impression if you were assisted by a humanoid than a box. It may make a difference to people. But in other cases I agree, the best form for the robot depends on the tasks.
Joydeep: I 100% agree. I think also it’s the challenge of the latest capabilities of our current research algorithms, because behind the capabilities that we’re seeing today are things like reinforcement learning and learning from demonstration. This is great for non-human interactive stuff like manipulating objects, but it doesn’t really scale well to human interaction. You can’t collect a very large number of demonstrations of a robot interacting with a person because there’s no one person. Everybody is different. And you can’t simulate a human realistically either. So the capabilities that we are using to bring these advances are just not very good for the human interaction capability that we need for the important problems.
Sabine: At the Bristol Robotics Laboratory we have the assistive living research facility and there are all these robots that are beautiful for human-robot interaction. So probably there’s a nice space for the humanoid form in those caring settings as long as they’re not deceiving in the way that we worry about from an ethical perspective. You also expect more from your humanoid in terms of its capabilities somehow because of its human shape. So somehow the cartoony ones that weren’t that capable were maybe better for those settings, also safer. I’ve seen so many videos of these humanoids toppling over in crowds.
Rina: What I found really impressive was when robots started to talk using LLMs. It really was amazing. And this is because it’s more going towards the human. It has moved from just writing text to generating it and voicing it. I think it’s easier to be influenced and believe something that’s a human-like object.
Marija: I wonder how much of this humanoid obsession comes from science fiction’s original picturing of humanoid robots? That sets the shape of the robot and the term robot. So everything is already pre-biased to be in that shape. It makes for a better story. Whereas if it is some kind of a box that does something, you have to do a lot of explaining as to what the robot is and what it does. I remember watching the Roomba clean the floor for the first time, and it was a very annoying experience because it’s not doing the things the way you would have done them.
Sabine: Also, because AI has had such advances with LLMs, I think people just see humanoids as the next obvious thing that will definitely happen. This is now seen as the normal form factor for a more general form of intelligence. And so there’s this obviousness to it. In some discussions, people have really been pushing me for an answer on when humanoids are going to work the way we expect. And I don’t know. RoboCup has an answer for that, right, because they have an aim to get them playing at human level against the trustees.
Joydeep: Well, I won’t say Robocup has an answer. Robocup has a challenge to defeat the human world champions by 2050. That’s more an open conjecture. You raised an important point that not everybody in the general public is aware of. There’s an equivocation of the capabilities needed. People see the capabilities of chatbots and you think that a robot picking up a laptop is trivial. But that’s because they’re equivocating what is hard as a human to what is hard for a robot, but it’s absolutely not true. In fact, doing the so-called very high-level planning in some sense can be easier for a robot than interacting with the embodied world. This is Moravec’s paradox. Lots of roboticists have spoken about this, but I think there’s a gap in the public perception of what is easy versus hard and the disconnect between what is easy and hard for a human versus what’s easy and hard for a robot.
Sabine: And there’s also just the data. We had buckets of data for language that we don’t have in the physical world.
Joydeep: We don’t have the internet of physical interactions.
Marija: Well, we now have people being paid to do manual movements that are supposed to be somehow used for training afterwards.
Joydeep: Yeah, so there’s a huge effort right now across so many companies spending so many millions to billions of dollars on what’s called teleoperation. And that is a human controlling a robot to do a task because that data can be used to train a robot. And yes, it is true, it is useful data, but there’s a divide even in the robotics community. There are some people who believe this is all you need – you just need to scale up. There are other people, like me, who don’t buy the argument that more teleoperation data is going to get us to general purpose robots. And I think the reason for that is rather simple. It’s that the complexity of the tasks that we want to do has combinatorial explosion. And we’re never going to have that level of demonstrations. And the other thing here is that we’ve seen this craze before with self-driving cars. If you think back, in 2014, maybe a little bit earlier, a common narrative among self-driving companies was that we just needed to go out and build HD maps of our environment and that’s all we needed. Like driving would be solved. Guess what? It didn’t turn out to be the case. This current craze about collecting data and mimicking humans feels very similar to that era. That’s not to say it wasn’t useful. But I don’t think that alone is going to solve the problem.
Marija: I agree with you as a non-expert in this, because humans do things the way humans do them, and machines do them the way machines do them. Birds fly, airplanes fly, but not in the same way. Again, with my non-able body examples, people without legs don’t “walk” like people with legs, right?
Sabine: One reason the humanoid form might make a lot of sense in robotics from the data perspective is human video. There’s loads of human video. And I’m guessing the reason for that is they collect the data from the humans rather than have them teleoperate a robotic system. And then things need to have a similar physical frame for those videos of humans then to translate to the capabilities of the robots. We recently did tests with VLMs [vision-language models] and it was harder to use that data that’s already available in these trained models because data is human-centric, and our robots are not in humanoid form.
Joydeep: I think what you said is absolutely true. We have a ton of videos on YouTube of humans doing tasks. The mismatch of course is the data; the way that data was collected was using very different sensors than what robots would have. So there is a mismatch there beyond the mismatch of the embodiment. But many of us agree that that data is useful.
Ella: If the solution to get general purpose robots isn’t more data, do we need to wait for better AI and better learning methods?
Joydeep: All the time. I mean, there’s one thing that’s clear. I think many of us believe that experience from interaction is crucial. You don’t learn just by mimicking what someone else did. You learn by getting some initial understanding based on seeing what other people did, but then trying it out and gaining intuition. And this is true for learning algorithms as well. If the expert shows you how to do a thing, and if you can do exactly what the expert did, you’ll be good. But invariably what happens is that if you try to do exactly what the expert did, you make one mistake and suddenly you’re in a state which you did not have in the expert demonstration. Now what do you do? You don’t have any training data for that, right? So this is why experience is useful and important.
Rina: From a causality perspective, the language we use is that there is observational data and interventional data. Causality cannot be gleaned just from observation, you have to interact in order to see what’s going on.
Joydeep: Yeah. I think interaction is the best way to exploit some of the behaviors that Marija was talking about, like getting behaviors optimized to the form and functionality that you have.
Sabine: I do worry though about this idea that now we need better robots or better AI to complete useful tasks. We were so close to getting our robots that were simple working really well, with enough computation on board, good sensing, that they were useful in lots of different ways. And now somehow we’ve moved the goalposts towards robots that are maybe not as green, not as easy to verify and test etc because they’re in a more complex form factor. And so I worry that we’re going to keep moving the goalposts instead of benefiting from all the stuff that already works in the robotics field and trying to get that out into applications. And the fundraising kind of shows that: the humanoids are raising and it’s harder to raise if you’re not in this next wave of humanoids or physical AI. We need to make sure it doesn’t end up being the only way forward.
Joydeep: In a talk I give I have a slide about this. I talk about the open challenges for robots and say that you need the following to have a successful mobile robot: a sufficiently detailed map of the world, an accurate predefined notion of action costs, a well-defined task objective, a curated list of entities of interest, no surprises, and enough onboard sensing and compute. This is a very large caveat list, but in settings such as manufacturing and logistics, these can all be achieved. But the moment you take any of these caveats out, suddenly you’re in the research landscape. And that’s kind of where we’re at.
Sabine: The other thing I heard is that people are calling everything humanoids, so the word has taken over. It’s got a lot of excitement, it’s helping to raise funds. And then you’ll have a wheeled humanoid, or a humanoid with six arms, for example. And there will all be forms of humanoids. So it’ll go full circle back to the different form factors.
Marija: If you think about it, a computer playing chess does little for solving general AI tasks. But it acted as a catalyst to attract attention, funding and resources so that they can be used to actually develop deep learning and so on. So maybe it works out in the end. You create this magic for people to believe in and invest in, and something else comes out of it. Although I’m not going to say I’m advocating this, it’s definitely borderline fraud sometimes, these videos of depicted abilities. It’s just easier to sell, right?
That was the trick, Sabine, that whenever somebody asks you when we are going to see useful or capable humanoids, you can ask them what they mean by useful or capable. Because then you turn around the question. People understand that we already have very many useful and capable things.
Sabine: Yeah, so they were asking when the breakthrough was going to be. I say the breakthrough is when we see all those robots that work in everyday environments, not necessarily humanoid, and all the things we’re pushing research-wise out in the world. And I think that’s the breakthrough.
Joydeep: I think even before ChatGPT was released December 2022, and folks who were playing with these large language models, GPT 2, 3, etc, we were seeing capabilities which were very hard to do beforehand. What really made it break through in the public consciousness was having a public-facing website – that was an inflection point. I doubt that for the robotic space we’re going to see a similar inflection point where suddenly we’re able to see generational new capabilities. I think we’re probably going to see slow adoption of robots in limited capabilities and specific scenarios, increasingly more and more and more. It’s just going to be a very slow adoption. We’re not going to suddenly have a robot capable of doing it all.
Marija: If any of you see a robot that folds laundry or irons laundry, wake me up in the middle of the night. To me, that’s the breakthrough point. If I see one of them, I don’t care how much it costs.
Joydeep: Well, again, the question is, under what conditions? If you are okay with a few hours to fold your laundry and specific kinds of laundry.
Marija: It can take as much time as it wants, let’s make it one week maximum. But no caveats. You have to take it out of the dryer and fold it in a neat stack, unspecified laundry. I mean, think about it. Would you care how long it takes as long as it’s done? How long does it take me, the lazy human to do it? Three days it’s going to live in the dryer until I get to it.
Joydeep: Yeah, so the challenges here are robustness. So you said it should not matter what the laundry is, and you want a certain degree of accuracy. You don’t want that every fifth attempt, it makes a catastrophic failure, and then you have to go rescue the robot. Now, not only do you have to do your laundry on your own, you also have to rescue the robot. I think as humans we have very high expectations for autonomy, which we don’t explicitly mention, which is fine, but I think that’s the challenge. In research papers, we say that an 80% success rate is phenomenal. No, 80% success rate as a product is awful. It’s a failure. Or even 99%, it’s still not good enough.
Ioannis Erripis: I have two small questions. Firstly, I’ve seen recently at least three companies that are advertising household humanoid robots. What do you think about those?
Sabine: The remote controlled ones.
Joydeep: Yeah, there’s certainly not anything which can do all household tasks autonomously. They’re actually using you for data collection.
Sabine: There’s a company in San Francisco that will pay you to do your household cleaning, because they collect data from your house.
Ioannis: My second question is more maybe superficial. What do you think about the design of most of the humanoid robots? And let me rephrase this a bit more specifically. One of my favourite robots was maybe Honda’s ASIMO, which looked like an astronaut toy. If you see behind the glass, it is smiling and has a very cute face. It’s a robot that you would easily give as a gift to your child, for example. But almost all humanoid robots are very menacing. They almost look like a villain from a movie. And actually, at the same time, they aspire to be adopted by consumers. What do you think about this? To me, it looks like a very unorthodox approach.
Joydeep: I think you’re absolutely right, but there are companies which are trying to build humanoid robots which are human approachable. Fauna Robotics is one of them. They are very much leaning into exactly what you’re saying.
Ioannis: I guess that’s an exception.
Joydeep: I agree it’s an exception, and I think partly that is because the capabilities that we’re really missing today are these fine-grained motor skills. People are trying to think about whether we can just build a form and understand how we can solve these fine-grained motor skills, and then we’ll talk about the interaction. But you’re absolutely right that we should think about them jointly.
tags: AAAI, coffee corner, RoboCup
AUAI is supported by:
