Sergey Levine is one of the world’s top robotics researchers and co-founder of Physical Intelligence. Recently I talked to him to dig into where we are today with humanoid robotics.
He also had a lot to say about the Chinese robotics ecosystem. I asked him questions about that and about competitors such as “what if Anthropic and OpenAI started their own robotics labs?”
We went over a bunch of questions about predictions on future timelines. I asked him what a concrete roadmap might look like so we could get a sense of how deployment will happen.
Hope this episode is helpful in learning where we are today and where we will be when it comes to humanoid robotics.
Check out the episode wherever you get your podcasts: YouTube, Spotify, Apple Podcasts.
Timestamps
04:20 - Most surprising capabilities so far
07:03 - The most inspiring real world robotics
08:36 - If OpenAI or Anthropic got into robotics
13:15 - Will one lab breakout from the rest
16:59 - Thoughts on a concrete roadmap
21:03 - Generalization and demonstrating it
26:04 - Types of data and which is best for robotics
34:34 - Why humanoid robotics differs from Waymo
37:10 - If humanoid robotics failed here is why
39:55 - Are there hot take modeling architectures in robotics
42:05 - Thoughts on AI safety in robotics
46:44 - Top robotics research paper recommendation
49:35 - Why is Boston Dynamics less top of mind
53:47 - Advice for his younger self
Transcript
00:37 — Where are we today
Ryan:
[00:37] LLMs have caused unprecedented impact and investment in that area. Physical AI and humanoid robots could potentially be even bigger. And so I wanted to ask you today about where are we today with humanoid robotics, and how do you foresee this technology actually being deployed into the world?
Sergey:
[01:01] I guess with machine learning, what we’ve learned over the last few years, or the last decade rather, is that it works when you do it at scale. And this is very obvious now, but it wasn’t always obvious. But there’s a caveat, which is you have to scale the right thing. And initially when people started working on models for language, for example, the dominant design was LSTMs. Some people remember what those are.
[01:29] They were kind of okay; they were a lot better than what came before that, but they didn’t really scale as well. And then the big thing with transformers was not that transformers were somehow particularly mathematically elegant or anything like that. It’s just that they scaled better. So they were easier to train on very large amounts of data with lots of parameters. So the technology proceeds in phases.
[01:48] First you figure out what you can scale, basically what is the scalable technology. And then you pour on a lot more of kind of an industrial-scale effort and adding lots of data, adding to model size. And that’s when kind of the magic happens. So when we’re doing kind of more fundamental technology development, the key is to understand what are those scalable levers. So figure out the design, figure out roughly the mixtures, and that by itself does something pretty cool.
[02:14] But that’s not the thing that actually changes the world. It’s like when you start pulling that lever that things actually change. So with LLMs, when the first GPT models came out with GPT-2, it did some stuff, but it was sort of like a parlor trick, so you could get it to synthesize a story about unicorns in Peru or something. And it was coherent English, but it wasn’t a thing that would solve lots of real-world problems.
[02:39] But the folks that worked on this kind of recognized that, hey, there’s something kind of magical that happens because as you add more data and you make the model bigger, this stuff gets more coherent and more effective. So they could see that if we do a lot more of that, then it’ll become a lot more powerful. So to come back to your question, what I would say about robotics is that it’s not in the GPT-4 to GPT-5 stage, where it’s an industrial-scale effort to kind of make the model bigger and get more capability out of it.
[03:10] It’s in that stage where we’re establishing the fundamental technologies. And because of that, what one should expect to see right now is not necessarily that each month the model gets bigger and more powerful by some predictable kind of scaling curve. It’s that the scaling properties themselves are evolving as we develop the right technologies. So to bring this back to something closer to reality.
[03:36] I’m very happy with the demos that we’re doing here at Physical Intelligence. And I think that a lot of the results that other people are coming out with are really cool. But these are, to put them in context, we should not expect these to be the things that are actually illustrating the power of scale. We should expect them to be developing the fundamental technologies that will be scaled up after that.
[03:55] So where I think we’re at now is that we’re actually kind of getting all those puzzle pieces in place. And I think it’s actually very close. I think a lot of the puzzle pieces are falling in place, but it’s not like. What makes it so hard to prognosticate about where the technology is going to go is that it’s not yet at that predictable scaling stage. It’s at the stage where we’re figuring out the puzzle pieces, which I think is really exciting.
[04:15] But it means that it’s also very hard to foresee sort of what the coefficients on that will be.
04:20 — Most surprising capabilities so far
Ryan:
[04:20] What are the most astonishing emergent capabilities you’ve seen so far?
Sergey:
[04:25] That is the thing that is the most fun. And certainly we’ve seen a lot more of that happening as we progress. In the very beginning, it was kind of little things, but they were kind of magical because in robotics, basically prior to 2024, the stuff never happened. So the little things that happen, and this was maybe at this point about two years back, we would see things like, okay, we train our policy for folding laundry and it takes individual shirts out of the hamper and tries to fold them.
[04:55] And then one very vivid memory I have in late 2024 is we were watching one of the evals, and it takes out two shirts at the same time. And I’m watching this, I’m like, okay, it’s done for. There’s no way you can possibly do this. And then it puts the two shirts on the table, disentangles them, puts one of them back, and then starts folding the other one. It’s like, wow, okay. In retrospect, you can do some detective work and figure out where it got that from, some piece of training data.
[05:22] But that was one of those moments where I don’t think anybody watching that eval thought that the robot was going to do this. It’s like a little thing. It’s exhibiting the common sense you expect people to have. But actually, the thing that I find more interesting recently is some of the mistakes, because one of the things that’s pretty remarkable about LLMs is that once they got good enough, even the mistakes kind of made sense in the sense that they weren’t like crazy mistakes where just outputs, like, zzz all the time.
[05:50] But their mistakes are sort of semantically sensible. We had an evaluation last year for π0.5 where the robot was cleaning up a kitchen, and it’s told, like, put away all the utensils. There are some spoons, spatulas, et cetera. And it tries to open the drawer where it thinks the silverware goes, and it can’t really get the drawer open, so it slides over and opens the oven, which is right next to it, and then starts putting this stuff in the oven.
[06:16] It’s like, you can sort of imagine that if you ask a child to clean stuff up and put it away, they might decide to do that because, okay, it’s like a container, and you can put stuff there, and nobody sees it. Another experiment we had is washing all the plates. So it would pick up the plates, wash them with a sponge, and put them on the drying rack. And this was an experiment on memory, because it has to keep track of everything that it’s doing and has a scratch pad of memory where it’s writing down, hey, I had three plates.
[06:40] I cleaned the gray one, I cleaned the green one. And then it drops one of them on the floor and drives the base over so you can’t see. And I was like, okay, I’ve cleaned the gray plate. It’s done. So, I mean, obviously these are not the things that we want to see. But it’s kind of interesting that some of the mistakes, they’re almost like what you would associate with a child trying to do the task.
[07:00] So now it just needs to grow up.
07:03 — The most inspiring real world robotics
Ryan:
[07:03] You mentioned the advancements. They’re kind of happening all over the industry. And I know there are a lot of people building humanoid robots. There’s Figure AI, there’s Tesla, and there’s many other competitors, and you have the expertise of what is hard and what isn’t. When you look at the competitors, has there been any advancement or achievement where you think, oh, that’s really admirable and that’s impressive?
Sergey:
[07:29] I actually think that one of the most inspiring things to me in the industry is to see the kind of takeoff that autonomous driving systems have had. Because one of the criticisms that is sometimes leveled against robotics researchers is, it’s like nuclear fusion. It’s the technology of the future, but it’s always in the future. But that’s what people said about autonomous driving, too. And now we’re in San Francisco.
[07:53] You can go outside and you can take a Waymo, and it will actually take you to your destination. And there was no driver sitting there. So, without getting too much into the technical details, I think what’s really inspiring about that is just this case in point that, yes, you can actually have one of these technologies of the future, and it actually does land. And I think it’s not an accident that it’s landing now in the mid-2020s, because a lot of the puzzle pieces for large-scale ML are getting to the level where we can put them together with actual physical systems.
[08:24] And there’s a lot of differences between driving and robotic manipulation, of course. But I think that the illustration that, yeah, we can actually land learning-based technologies in the real physical world, I think that’s really inspiring.
08:36 — If OpenAI or Anthropic got into robotics
Ryan:
[08:36] If OpenAI or Anthropic started investing more heavily into robotics, how do you think that would impact the industry? Do you think competitors would be worried about that?
Sergey:
[08:49] Robotics is an area where, to be fair, the ecosystem hasn’t been as healthy as it has in other areas of machine learning. And what I mean by that is that computer vision and NLP are things that sort of lend themselves naturally to a machine-learning-based ecosystem, because there’s freely available data. People kind of have a general acceptance that they’re going to be using learning. There aren’t concerns as severe concerns about safety, at least physical safety.
[09:20] Right. You know, people rightfully are concerned about AI safety, but it’s not the same as like a physical device causing some physical harm. And because of that, I think it’s a bit easier to spin up a very serious large-scale ML effort in those areas. Robotics is not like that. Robotics traditionally is not a discipline that really embraces sharing of data, for example. So I think that the more activity there is around learning and robotics, the more I think it’ll shift people’s thinking toward this kind of future where we accept that robots will be controlled by learned models, not by hand-designed controllers, that there will be data, that data will need to be shared because there’s no way that somebody can build out a true foundation model in a single vertical and basically kind of shift the entire thinking around robotics to look more like how we think about vision and NLP as opposed to traditional factory automation.
[10:12] So I think in that sense, much as I’m proud of the work that we’re doing at Physical Intelligence, I think it’ll take more than one company to kind of shift everyone’s thinking in that direction.
10:22 — Chinese robotics
Ryan:
[10:22] As a bystander, I see on Twitter these really impressive demonstrations of Chinese robotics. It almost feels like they’re ahead in some sense, but I don’t have that deep domain expertise. So I was curious about your thoughts on what you think of China’s robotics, and are they further along?
Sergey:
[10:45] I think that one thing that is very useful and constructive for us to do, those of us that work on these things in the United States and in Europe, is to ask what is the lesson to learn. And to me, one lesson is that it’s important to have a healthy ecosystem. And ecosystem means that there should be obviously good researchers, good engineers working on these things. There should be healthy open source.
[11:18] But it also means that the different industries that contribute to robotics need to individually be very healthy. And those industries are not just the computer science, ML, and model building stuff. It’s also supply chains, manufacturing hardware, R&D, these are all very important things. And aspects of those things are things that the United States does quite well. Other aspects of them are things where the United States has sort of let things go a little bit.
[11:52] And I think that what we should do is we should look at what’s going on in the world, look at some of the excellent results that Chinese labs are doing, that labs in other countries are doing, and we should take away that lesson that we should strive to build a healthier ecosystem. And that means investing in all the different facets that contribute to this. So I’m not much of a business person, I’m not much of an investment person, so I can’t claim to know how to do this, but I think that it’s important to sort of fully embrace that this is a holistic thing and not something where we can do just one piece of it and outsource everything else.
[12:25] Basically.
Ryan:
[12:27] Of all the pieces of the ecosystem, are there parts of it? As a robotics researcher in the US, if it was better, that would have the biggest impact on the advancement of robotics.
Sergey:
[12:39] Yeah, I think certainly availability of reliable, low-cost hardware is a big deal. And right now, I mean, certainly for hardware used for robotics research, a lot of that does come from China. And it’s good, it’s relatively inexpensive, it’s of high quality, and meets the standards that people generally need. But it would be awfully nice to be able to source all that domestically as well. And I don’t think there’s anything impossible about that.
[13:07] I think it’s just a matter of embracing the fact that that entire ecosystem needs to be supported rather than just one piece of it.
13:15 — Will one lab breakout from the rest
Ryan:
[13:15] You can imagine that the lab that is first to get to scale will be the first to break out. There may be this exponential growth effect where once you deploy, deployment helps you grow faster, and growing faster helps you deploy, and this flywheel. And so do you believe that that will happen in this industry, where one lab, whether in the US or China, will hit some breakout point and they’ll kind of jump everyone else?
Sergey:
[13:47] There is a lot of truth to that. I think that there’s an important detail to keep in mind. The detail is that you kind of have to scale the right thing. So I think that basically the statement that the way that I would phrase this is having an effective positive feedback loop where more deployed robots translates to more model capability. That’s kind of the key, and that makes total sense. The trick is that there are lots of ways that it can be done wrong.
[14:23] So I’ll tell you a few obvious ones. One obvious one is let’s say that I’m a car company and I have a robotic arm that is welding cars, and it’s there on the assembly line, and it welds cars every day, and it gets a million welds every month, and so on. If I just use that as my data flywheel, I’m unlikely to get something more capable than a robot that welds cars. So since the robot is already there and it’s already welding cars, presumably the marginal improvement for that is not all.
[14:54] That valuable. So that’s an example of how you kind of have to scale the right thing, because data is not quite as fungible. It’s not like electricity or oil. You can’t just buy more of it. It has to be heterogeneous. So I think that basically it’s right, but you have to scale the right thing with the right technology and the right kind of source of diverse learning, like data is more like an education program for your robot than it is a fungible commodity.
Ryan:
[15:19] If you think about that data flywheel and, I guess, putting out the proper platform that would generate data that matters, that would then improve that platform, when do you foresee this kind of happening in the world?
Sergey:
[15:34] I think, on the technology side, things are advancing very rapidly toward that. And I think that maybe a particular balancing act to strike that would significantly determine that timeline is how much structure somebody’s willing to admit. So on the one extreme, you could imagine jumping straight to fully unstructured deployment domains like home robots. And that could be really exciting because then you have a lot of diversity right off the bat, but the bar is a lot higher to be effective enough in that domain to be safe enough.
[16:04] Safety is a much bigger issue there because it’s around people in their home. On the other extreme, you could imagine much more structured tasks, maybe not quite the welding robot, but sort of like maybe the robot down the hall that does something more unstructured. And that could be a much easier domain in the sense that there’s much less safety concerns because it might be around trained humans, the task might be more predictable, and so on, but the marginal value of each bit of data you get in that domain is lower because there’s less variety.
[16:36] So it’s like you could take off earlier, but with a smaller slope, or later, but with a larger slope. And you kind of have to calibrate that. But my sense is that whichever end of that extreme we’re talking about, it’s in the single-digit years rather than the double-digit years at this point. So it could be that the more structured ones might be happening now or next year. The less structured ones might be a few more years out, but probably not like a decade.
16:59 — Thoughts on a concrete roadmap
Ryan:
[16:59] A lot of people talk about timelines, and it’s kind of more nebulous. And I’m wondering if you were to put down, I guess, somewhat of a roadmap, but not so concrete, but just milestones toward that north star of a robot in my home doing.
Sergey:
[17:16] Yeah, that’s a really good question. So maybe to preface this answer, I think that there’s one thing that is very important to say about robotics that is very easy to miss from the hype cycle and the demos that people put out, which is that the hard thing in robotics was always generalization. But when somebody shows a demonstration of their system, the demonstration alone usually doesn’t make it clear what level of generalization is being shown.
[17:48] The highly acrobatic robot demos, for example, are really exciting to look at. But typically, if it’s something that is a little bit more staged, sometimes literally on stage, if it’s a show that is obviously rehearsed, and that’s okay, because it’s literally just a show, but it’s not the same as doing a task every time reliably in any home. And the generalization piece often doesn’t look that impressive when viewed in isolation, because generalization is sort of a property of many trials, not of one trial.
[18:20] So you might see the robot doing something fairly mundane and unimpressive. But what’s exciting about it is that it’s doing it with an object that it’s never seen before in an environment that has never been tested before. And that is actually harder than doing an acrobatic backflip that it’s practiced millions of times. So with that said, my answer to your question is that the roadmap is all about both achieving better generalization and kind of that second-order effect of having a mechanism to get more generalization as you generalize.
[18:53] So one of the steps on that roadmap is to have a very concrete demonstration of a robotic system that gets better with autonomous experience that is collected in a setting that it wasn’t originally trained for. So I do whatever I do in the backend in the lab and whatever. I get my model, I get my adaptation algorithm, I put it in a new setting. Maybe it’s a home, maybe it’s a factory, whatever it is, something where it’s doing something real, and it does okay.
[19:24] But then over time, it gets better and better. And it gets better and better to the point where it reaches sort of practically relevant levels of robustness without sort of capping out at like 50%. I think that would be a major milestone, because now that says, okay, if this is truly an automated process, it’s improving, it’s getting better, even as it’s collecting useful experience. Now, I can take it and I can put it in lots of different domains, collect useful experience, do something that people actually want, and it’ll improve the model.
[19:51] So that, I think, is a really major step on that road. I think another really major step is to demonstrate a very concrete and practically useful way to transfer knowledge, to transfer common sense, to achieve robustness. So that’s like kind of the other scenario we don’t get to practice. If you’re driving your car on the road and you see a fire truck and you see a bunch of traffic cones, even if you’ve never been in that situation before, your common sense tells you, like, hey, I should slow down.
[20:20] Maybe I don’t know how to react optimally, but I shouldn’t just barrel on through the cones and upset the firefighters and so on. So that’s common sense. And if you can apply that common sense to effectively recover from unexpected situations, like when the robot put the spatula in the oven, it should probably open it up, take it out. It knows that this is not the thing you do semantically.
[20:40] That, I think, is another important step because that tells us that we can use common sense to fix mistakes.
Ryan:
[20:46] I could imagine a more narrowly scoped robot. Let’s say it’s a humanoid robot. That’s one step in assembly line, in manufacturing. And my understanding, you’re less interested in that because that’s not really a step toward that general intelligence.
21:03 — Generalization and demonstrating it
Sergey:
[21:03] What I would say, it’s not that I’m less interested in it, it’s that I think that the real world has these leaky abstractions that make that kind of stuff a lot more complex than it seems. Let me try to explain this with an analogy. So in the ‘90s, when people really started working kind of full steam autonomous driving, there was this idea that we could avoid a lot of the hard problems by kind of instrumenting the environment a little bit.
[21:31] Like, we’ll have magnetic sensors along the highway and so on, and cars will have a little sensor on them and a little transmitter so they can tell where each of the cars are. Kind of like the way you do it with aircraft, basically. And then people thought, well, we don’t really need fancy AI. We’ll just have these sensors and it’ll just work. And that just basically didn’t go anywhere because the real world has so many messy exceptions and special cases that even if 99% of the time, the magnetic sensors and all that stuff just allow the car to drive the 1%, when someone steps in the middle of the road or there’s a piece of trash or whatever, it just messes everything.
[22:05] Up. So the thing that actually worked was when Waymo said, hey, we’re going to not try to avoid the hard problem. We’re going to actually try to deploy our cars not in the middle of nowhere, but in San Francisco, very messy, and let’s just deal with it head-on. And that allowed that community to make progress. And I think robotic manipulation is going to be the same way, that past the fully structured world of the factory, if you want to go even a little bit outside of that, even if 99% of the time it’s all pretty straightforward, that 1% when something weird happens means that you really need the full scope of the problem basically to be addressed.
Ryan:
[22:41] This kind of reminds me, I remember Figure AI had this demo where they live-streamed the robot sorting packages. When you watch that demo, does it demonstrate generalization in your opinion?
Sergey:
[22:54] Yeah, I think it does. And by the way, this is something that I find very encouraging is that, like I mentioned, it’s hard to show generalization in a video, and it’s clear that lots of people are thinking about that and are thinking about how do you present something that somebody can watch and take in kind of at a glance what generalization is. I think you can see a lot of creative steps towards that.
[23:16] Live demos, these kind of really long time lapses. I think it’s a great idea, and I think that’s a really nice way to move towards elevating the importance of generalization in people’s consciousness. When we were working on the π*0.6 project, the RL project that we did late last year, we wanted to do some longer horizon experiments. In some cases it’s obvious. We had our robot assembling boxes at Dandelion Chocolate Factory.
[23:46] So they’re like, it’s an actual chocolate factory, so they need the boxes. So we ran it for several days. But we had this coffee task, which was the robot using an espresso machine to make espresso. So what we did is we ran it for 13 hours making espresso drinks. And we are, we try to be very, I guess, environmentally conscious about this. So we didn’t want to throw out the coffee.
[24:05] So after 13 hours, everyone in the office was a little wiry because someone has to drink the coffee. But it ran for 13 hours, and it was pretty cool. It screwed up a few times. It’ll spill all the coffee grounds, and then it needs to go and get a cloth and wipe it down, but it does it and nothing exploded. Thirteen hours went by. Probably the most negative consequence was loss of sleep from too much caffeination.
Ryan:
[24:29] But when it spills the coffee grounds and cleans it up, it did that by itself.
Sergey:
[24:34] Well, so the way that that experiment was done is that there is a high-level prompting. So roughly the prompt is updated maybe every five minutes or so in between semantically coherent tasks. So you tell it, like, make espresso, clean up the machine, et cetera. So those steps, the actual clean up the machine, was commanded by a person. In principle, we could automate that. In fact, one of the things we’re spending a lot of effort now on is improving our high-level policy that does those commands.
[24:59] But for that experiment, every five minutes, somebody basically updates what it’s being asked to do, the way we intended. It was like the commands would be like, if you go to an actual coffee shop, you say, “Oh, I want a latte, I want an espresso.” That was supposed to be the prompt. Except then you also have to tell it, “I want you to clean it up before you do the next one.”
26:04 — Types of data and which is best for robotics
Ryan:
[26:04] It sounds like on the way to generalizing data is a very important part of that.
[26:10] And I was reading there are different types of data. There’s simulated data you could collect, like physical interactive data. And I want to hear your take on what’s the best data to get, what’s the worst, and what are the pros and cons.
Sergey:
[26:25] This is, by the way, a question that is, I guess, quite. There’s a lot of discussion in the robotics community about this question. And some people have very opposite opinions on it. My own take on this is that a lot of different data sources are easier for the model to internalize if it can ground them in a thorough physical understanding of the world. So let me try to explain what I mean with a few examples.
[26:53] If you want to learn to fly an airplane, you will probably use a simulator, at least during part of your training. But the simulator makes a lot of sense to you, because when you start using the simulator, you have a lot of world knowledge that you can use to ground what’s going on. You know that when you are using the flight simulator to learn how to fly the airplane, you’re not just playing a video game.
[27:15] You’re trying to acquire knowledge that you will then use with a real airplane. And you understand that there’s sort of an abstraction there. Same thing if you’re playing a really cartoony Atari game or something. You know that all the symbols on the screen, you can sort of connect them to things that you’ve experienced in your life, and you can make an analogy there. So a lot of that, even though it kind of seems like these simulated environments reflect aspects of the real world, to us, they make a lot of sense because we kind of bring to bear a lot of our own prior experience, and we fill in the blanks that the simulation has.
[27:52] And also, if you want to use data yourself as a person of somebody else doing something, if you watch someone, let’s say, cooking a meal, even though you don’t experience every movement they’re experiencing, you have a lot of that knowledge that you bring to bear. And you’re like, okay, I see they’re picking up the salt shaker. I’ve put salt on things before, so I kind of roughly know what’s going on there.
[28:13] And I can file it away at this level of abstraction of add salt without having to figure out all their muscle movements. So my point with this is that once you have that understanding of how you do things physically with your own body and how you experience the physical world now, all these other sources of knowledge can be connected up to it, because that foundation you get from your experience helps you ground everything.
[28:34] So where I’m going with this is that if we have a robotic foundation model that is trained on lots of real embodied data, that provides that grounding, and it might actually be much better able to absorb other sources of knowledge. And this is actually a little bit upside down relative to how some people think about it, because it’s very tempting looking at the success of Internet data for LLMs to say, well, maybe we should do the opposite.
[28:56] Maybe we should start with YouTube videos and then put robot data on top of that. But I think it’s actually the other way around, and I even have a little bit of evidence for this. So my colleague Suraj Nair, together with Simar from Georgia Tech, they had a project together a while back where they took our robot foundation model and they added human video data. But they didn’t start with human video data.
[29:22] They actually started with a model trained on robot data and then added video data on top of it. And what they did is they looked at the representations inside the model. Basically, how does the model represent human experience versus robot experience? And they found that if you use a small model with a small amount of robot data, predictably the human experience and the robot experience are fully separated, meaning that the feature representations are different.
[29:42] But if you train on lots of robot data from lots of different robots, then the features are grouped much more by what task is being done rather than by whether it’s a human or a robot. And when we looked at the feature plots, it was just mind-boggling because literally when you crank up the amount of robot data to 100%, they just line up perfectly. You do this t-SNE embedding, you see the shapes of the embeddings, and it’s just all task identity and minimal sensitivity to embodiment.
[30:10] And to me, that’s kind of mind-blowing because the base model wasn’t trained on any human data at all. But once you start adding human data, it represents it exactly the same way. And I think that’s really exciting. And I think that to me is one of the strongest indicators that if you have that good foundation of robot experience, you can put everything else on top of it. It’s actually better at absorbing that.
Ryan:
[30:29] Is it important that that base model has data that was collected using that specific set of motors, specific set of joints?
Sergey:
[30:39] So far we’ve obviously put a lot of effort into cross-embodiment models that can handle many different robot types. But generally, you do need data of the robot you’re going to be deploying on to get good performance. So kind of the metric of generalization there is not can you zero-shot a new robot, but it’s mostly can you get away with less experience from the new robot and transfer skills from other robots?
[31:02] Okay, so that’s the current state of things. Now there is a little bit of a surprisingly positive read on that, which is even though you need data from these robots, the amount of special stuff that the model is doing is kind of minimal. When we started doing all this, I had a big long list of all the cool research I wanted to do to better accommodate different morphologies. Can you factorize the model’s representation in some way so that there’s a six degree freedom arm head and a seven degree head and a gripper head, et cetera?
[31:32] We didn’t do any of that. The model just outputs a big vector of numbers. If the robot has fewer degrees of freedom than the number it outputs, it just zero-pads it. There’s just nothing fancy, and that’s it. And then it just trains on all the robots and outputs the correct actions based on what is seen through the camera. But now, to your point about whether you can handle new robots.
[31:55] So far, the thing that we focused on, and I think this is showing some promise, is being able to transfer skills between robots. And this is actually where the particular choices in how the model works seem to matter. For example, you can have a model that does some intermediate thinking, and that thinking can be done in different modalities. So you can think in text, and thinking in text is really good for transferring high-level behavioral structure.
[32:27] So that’s basically how you understand that, hey, if I want to clean the kitchen and put away the silverware, first open the drawer. That’s kind of a semantic inference. And you can transfer that very well because obviously that’s largely agnostic to any embodiment or anything like that. But even lower level things can be transferred if you use the right representation. So one experiment we did is we had a thinking stage that is expressed in images, where you basically dream up an image of the next milestone in the task.
[32:55] And with that, we can actually get a robot, the UR5 robot, to fold a T-shirt, even though we didn’t have any T-shirt folding data on the UR5. Because while getting the arm motions correct is very hard, because the robot basically requires very different joint angles to do the task, cooking up an image of what it looks like for it to fold a shirt is not that hard because you’ve seen the robot arm in all sorts of different poses, you’ve seen the shirt in all different stages of being folded and unfolded, roughly where it should hold it.
[33:21] So getting a good generative model to cook up that image is pretty straightforward. And once you have the image, then, from that backing out the correct actions is easy too, because you can just look at the synthesized arm angle and just back out what the angle should be. So it’s not changing the problem, but it’s just introducing this intermediate step that makes it easier to solve. Just like if you’re solving a math problem, if you figure out the right intermediate step, kind of the answer is obvious from that intermediate step.
[33:47] And I think that’s really exciting because now that shows that this level of generalization across robots, and I’m sure other generalization too, can be facilitated with thinking, just like in large language models, but with a twist that you have to think in the right modality.
Ryan:
[34:00] Interesting. So it outputs a, I guess that image is what its video sensor is seeing, and it’s like the next step.
Sergey:
[34:09] Yeah, you can almost think of it like image editing. You can do the same thing with video. You can do it with video prediction. But the key is to imagine what it would look like to progress on this task. That seems like a very human thing to do.
[34:22] Some things you plan semantically and some things you plan spatially. If you’re doing rock climbing, you’re probably not thinking like, “Hey, left arm to rock, 37 centimeters to the left.” You’re probably more imagining your arm reaching for the rock.
34:34 — Why humanoid robotics differs from Waymo
Ryan:
[34:34] Earlier in the conversation, you mentioned that Waymo was very inspiring, and their kind of path to productionization is proof that you can do real-world generalized robotics. And if I recall correctly, when I was a lot younger, it was kind of this early promise of, “This is going to happen,” and then in reality, it took a lot longer. So I guess my question is, in the case of humanoid robotics, what would make you say single-digit years, it’s coming, versus a long tail and policy challenges as well?
Sergey:
[35:17] I think one big difference between how robotic foundation models address the problem and how more traditional engineered systems address the problem is that the stack is really thin. So it’s not easy to train a foundation model. Obviously you need to get the right data. There’s a lot of work that goes into curating, labeling, all that other kind of stuff. But the actual software that runs on the robot is very, very simple.
[35:42] So you might have some kind of thinking or reasoning stage. You might have the model produce actions; it needs to be fast enough. But if you think about it in terms of raw lines of code, it’s much, much lower than a more traditional AV stack. And partly that’s because modern autonomous vehicles, the work on that started a lot earlier with very different technologies and evolved over time. Partly it’s also because the problem is more safety-critical.
[36:13] Like, yes, you don’t want a robotic manipulator to drop a fragile object, but at the end of the day, that’s a lot less bad than having a car hit somebody. So that is not to say that the safety challenges with robots are not real. They’re very real and it’s very important to tackle them. In fact, it’s probably one of the harder ends of the problem. But they are not as much of a hard stop to practical deployments because you can come up with tasks and environments and domains and also physical hardware where those problems are a lot less severe.
[36:47] So I think that combination, radically simpler software stack, plus less drastic software challenges, actually make it a lot easier. And because, to your earlier point, there is this kind of flywheel effect, that there is a positive feedback loop, that starting to get things out in the world, even under some constraints, will actually facilitate getting them out more and more.
37:10 — If humanoid robotics failed here is why
Ryan:
[37:10] I think a lot of people are familiar with this idea of postmortem, looking back on why something failed. But in this case, I’m curious, what would you say to a pre-mortem in the sense of if humanoid robotics did not succeed in single-digit years, what do you think would be the most likely reason why humanoid robotics failed?
Sergey:
[37:34] Ultimately, for these things to be truly useful, they do need to reach a level of reliability and robustness and generalization that is higher than what we typically expect from LLMs, for example, or generative AI for images and video. Because typically these tools, they are very much human-interactive tools. You get an LLM to do something, and it doesn’t do quite what you want, so you sort of revise your prompt, and you basically iterate with it.
[38:01] And that’s why even the earlier LLM tools, like the first version of ChatGPT, even though they were much more primitive than what we have now, they were still already useful because somebody could just keep hammering at it until it basically solves their problem. Just like if you’re using a search engine, you type something in the search engine, you don’t get quite what you want, you revise your query and then you get what you want.
[38:20] Whereas with a robot, the full value of it is unlocked when it’s actually doing the thing autonomously. So to have somebody constantly iterate for every single task is almost antithetical to the benefit that you’re getting. So I think a lot of the risk has to do with how easy is it to get that level of reliability and robustness. And that’s again where some of the demos might be a little bit misleading, because if someone shows a demo of their robot doing something cool, I mean, obviously if everything is presented in a forthright way, that could still be a very good indicator of progress.
[38:52] But it doesn’t make it obvious how far or how close it is to reaching that practically relevant level of robustness. So I’m personally a big believer in using techniques like reinforcement learning that can actually benefit from autonomous experience to kind of fine-tune those last few percentage points to make it go from 95 to actually 100%. But that’s really important, and it’s not yet a solved problem.
Ryan:
[39:18] If it did take longer than expected, it’s because the bar is higher.
Sergey:
[39:23] Because the bar is higher. And in particular, those last few, kind of the last inch, so to speak, is something that requires not just really good models, but also new innovations in technology. I mean, I don’t think that I’m not the kind of person that would say, oh, we should throw out everything that we know about foundation models to start over. I don’t think it’s that at all. I think that roughly the puzzle pieces that we have are actually very good puzzle pieces.
[39:46] But still, we should acknowledge that right now, the methods and the models need more work to cross that level of robustness.
39:55 — Are there hot take modeling architectures in robotics
Ryan:
[39:55] In LLMs, it feels like everyone is doing kind of the same thing, but different flavors in the robotics industry. Is everyone doing kind of the same thing? Are there any hot take architectures that are different direction?
Sergey:
[40:09] I actually think that there’s a lot more heterogeneity than it might seem. One big dividing line that I think is maybe not as obvious from just looking at the results is the distinction between fully embracing the foundation model ethos, so to speak, versus focusing on specific vertical areas. And I think this is hard to tease out sometimes because obviously everyone’s going to say, “Oh, I’m doing the thing that LLMs did,” because that’s the cool thing.
[40:40] But the foundation model ethos fundamentally is something like this: if you have a particular problem you want to solve, it is better to train a more general model that can use data from a breadth of problems. And if you do it right, it’ll actually be better at the specialized problem you want to solve than a narrow specialist. So again, to come back to the LLM analogy, if you want to do machine translation, don’t build a machine translation system.
[41:05] Build a language model that understands all language tasks and throw it at machine translation. And in robotics, I think that is actually very deeply uncomfortable to people. Because if someone is actually working on an application, they’re doing warehouse automation. It is very awkward to think, oh, if I want to do warehouse automation, let me collect data of putting away silverware in kitchens.
[41:27] It just sounds bizarre. But that is the foundation model lesson: if you have enough breadth, if you collect data from a wide range of different tasks, then you will acquire those generalizable skills. And if your model is built correctly, it will repurpose those skills for whatever situation it encounters. So I think it is actually true that even if you want to build a warehousing robot, you are better off collecting a breadth of data and will be better at handling all the weird edge cases you might encounter, even in that warehouse domain.
[41:55] But this is not something that’s easy for people to accept because it’s just so antithetical to the principle of building kind of a traditional, vertically integrated robotic system.
42:05 — Thoughts on AI safety in robotics
Ryan:
[42:05] I noticed this new, interesting phenomenon with these AI companies, which is if they’re wildly successful, it creates this, I guess, worry or new set of things. So, for instance, when Anthropic had a very powerful model, then the government comes in and there’s these worries about safety and risk and all that. And I’m curious how you think about that. Like, if Physical Intelligence this year had a phenomenal, incredibly capable, generalized model, how do you think about those kinds of topics that might come up?
Sergey:
[42:48] Working on AI safety is not a new thing. My colleague at UC Berkeley, Stuart Russell, was talking about this stuff over a decade ago, and lots of people spent a lot of time working on it. It’s just that the trouble is when the technology moves so fast, the important problems are not just a function of the core principles. It’s also a function of how society reacts to it, what kind of tools are adopted, and so on.
[43:11] And I think that’s very, very hard to anticipate. So I don’t have a very satisfying answer here. In terms of how we are approaching it, our philosophy around all this stuff is basically one of empirical experimentation. Let’s get stuff out there. Let’s see what happens in the real world. Let’s see what goes right and what goes wrong, so that we have as much of a preview for what the technology can do, what are its weaknesses, what are its strengths, and so on.
[43:42] But at the end of the day, you kind of have to just keep your eyes open, see what happens, and adjust as you go. It’s very hard to anticipate. And I think your question, though, is very spot on. Even though I don’t have a great answer for you because, yeah, if we’re having this much concern and issues with AI systems that are basically limited to using computers, we’re presumably going to have strictly more concerns and issues with AI systems that can do everything in the physical world that we can do.
[44:14] Right. So the issues are real. It’s just you kind of have to see what happens and then adjust.
Ryan:
[44:19] And that’s kind of in the scary path, but in the happy path. If everything goes well and we have incredibly capable models and robots, in 10 years, is the North Star, that’s the end of human labor.
Sergey:
[44:37] I believe it’s a mistake to think of robots as mechanical people. Computers at some level are kind of like mechanical brains. But when personal computers really took off in the ‘90s, early 2000s, et cetera, it’s not like the first thing that happened is that people replaced their brains with computers. Rather, what we saw is actually a proliferation of very different kinds of computers. We saw kind of ubiquitous computing.
[45:03] So you would have a computer on your desk, but you might also have one in your pocket. You might have one in your refrigerator and in your car. Because computing became so accessible, you could have a little bit of computing in everything. And I don’t think that’s what the people that first started thinking about this stuff in the 40s and 50s would have imagined. They would have imagined room-sized computers whose job is to control the policy of an entire country or something, rather than a little bit of computer in everybody’s refrigerator.
[45:32] So I think by analogy, we might imagine there might be a little bit of Physical AI in everything, and it might just be lots of everyday things that you have to do yourself. Now you can get a little bit of help with it. I think the other example that’s worth thinking about is modern coding agents. So I think that this is something where, of course, the jury is still out as to what the end game of coding agents is.
[45:57] But certainly from the experience of software engineers today, it kind of seems like it’s fair to say that most would consider coding agents to be more empowering than somehow causing them to panic. I mean, obviously some people might panic, but in general, at least from the software engineers that I’ve talked to and from my own experience, it’s more empowering to be able to amplify how much work you can do with AI tools.
[46:28] So I think from that, and that maybe is a pretty direct analogy, because that is straight-up an example of an actual real job where AI has entered into it and has actually provided more leverage to the people doing that job. So I think that’s another example that we can look to. But the truth is that I think it remains to be seen.
46:44 — Top robotics research paper recommendation
Ryan:
[46:44] So in LLMs, there are a few seminal papers that, if you’ve read those papers, you kind of get a sense of the lineage of the breakthroughs that mattered in understanding where we are today in the robotics industry. Are there a set of top papers that you really think kind of show the breakthroughs that people should know about if they’re curious about the state of the art in terms of humanoid robotics?
Sergey:
[47:11] One thing I would point out, and this is partly a shameless plug, because I am a coauthor on that paper, though candidly, 99.9% of the work on this was done by Tony, who was the lead author, is the original ACT paper, the ALOHA paper. It’s an interesting example because in some ways the ideas weren’t really that new, but they were illustrated in a really nice way. And the idea was that, hey, if you set up the right kind of low-cost robot setup, in his case it was based on these robot arms from Trossen Robotics that are, they’re like $7,000 hobbyist arms.
[47:48] He set them up in a bimanual setup with a leader-follower teleoperation device. And he showed that actually if you do it right, without really any particularly fancy tricks, you could easily collect teleoperation data of extremely dexterous tasks that people had previously thought would require very sophisticated hardware and all sorts of really expensive stuff, and then set up a fairly straightforward transformer-based model, and it could actually do a lot of those tasks.
[48:13] And it’s kind of like an interesting thing because usually in academic research we put a big premium on do you have some sophisticated new mathematical thing or some sophisticated technical insight. And in that paper, which I think at this point has been hugely influential, the insight is really just, yeah, just put together the right pieces and have a little bit more faith in what a simple robot could do, so to speak, equipped with a good intent learning system.
[48:39] And he showed things like replacing batteries in a remote control. He even got a mannequin foot, and he showed that he could put a shoe on it for an assistive task, sort of. Some people need help getting their shoes on, so that’s good. But what people found, I think, so interesting about that paper is just how far you could get with relatively simple building blocks. And at this point, he open-sourced the code for it.
[49:06] And the ACT code has been used by lots of people. Sort of like if someone wants a very basic starter kit for robotic learning, that’s usually what they grab. And I think it’s worth for somebody who wants to get into the field to go through that paper and really understand what’s going on there. Because even though in some ways it’s not that sophisticated, I think it provides a bit of calibration on what matters.
[49:31] The details matter, but the details don’t have to be complicated.
49:35 — Why is Boston Dynamics less top of mind
Ryan:
[49:35] Before all this large-model robotics kind of wave, prior to that, Boston Dynamics had these really impressive demonstrations and tons of mind share. I guess I wasn’t even in the field, but I think, wow, they’re really doing incredible robotics. And then in the last, I don’t know how many years, I don’t really hear about them much anymore. Is there some shift in the industry that made that so, or is that something you could explain?
Sergey:
[50:12] So the way I would explain it is this: robotics at some level is about building complex systems. So even though it’s very tempting to say, oh, there’s different areas of AI, there’s LLMs and vision and robotics, one of those is not like the others because for robots you actually need all the parts, everything from how you wire up the robot, what the power source is, what does the actuator look like, all the way to how does it do high-level planning to determine what tasks to do next.
[50:46] And even though we could look at these things and say all of these different videos and different companies and different demos, they’re all robotics, they’re really different parts of the stack. A lot of what the classic Boston Dynamics results show is very sophisticated hardware, very carefully designed hardware with a traditional control approach, with very smart controls engineers setting everything up, but with comparatively less emphasis on the kind of decision-making aspect.
[51:20] And I think that at a particular point in time that actually made a lot of sense. Because if we can’t build the physical body, it doesn’t matter what kind of decision-making system is running on it. But to our earlier discussion about generalization, at this point we’re at a stage in the development of these things that even though we can do more on hardware, in many ways it’s good enough. And the big challenge is how to have the decision-making loop that actually works and that reacts intelligently to everything in the environment.
[51:52] And the place where I would draw the dividing line between those is decision-making loop doesn’t mean symbolic decisions. It could mean low-level decisions. The question is, do you need to take the rest of the environment into account, or are you just dealing with a robot? So if you want to do a backflip on flat ground, you most have to deal with a robot. But if you want to pick up a coffee cup off of a table, even though that’s maybe in some ways simpler than doing a backflip, you really have to understand what’s going on in the rest of the world rather than just your own body.
[52:20] And that dividing line, I think the way that technology has panned out, I think it’s fair to say that that is the dividing line between AI and controls. Controls is when you have to control the robot body. AI is when you have to take into account what goes on outside of the robot. And I think that that’s why you see this divide, because I think a lot of the demos where you mostly needed to deal with the robot itself and not the rest of the world, really good controls, could allow you to admit a very good solution there.
[52:51] And another thing I would say here is, okay, if there’s a lot of controls work that goes into doing some particular skill, well, there is actually something to learn from that. Because if you can hand-design a controller that performs a sophisticated behavior, very likely you can also learn that controller. So just that proof of existence that the thing is possible, and not only possible, but also simple enough that a person could build it.
[53:15] Because remember, people are, at the end of the day, even with code these days, the kind of complexity that people can handle is not as high as the kind of complexity the AI can handle. So if a person can hand design something to do a backflip or do some acrobatics, that’s a really great proof of existence that there exists some relatively parsimonious control law for doing that skill. And parsimonious does mean generalizable.
[53:36] So if it’s simple enough for a person to design, probably there’s something fairly general in there. And if you can learn it and automate it without having to have the human controls engineers in the loop, that’s good news.
53:47 — Advice for his younger self
Ryan:
[53:47] And then last question for you is, if you could go back to when you just entered the industry and give yourself some advice, knowing everything you know now, what would you say?
Sergey:
[53:57] One thing that I’ve learned over the last few years, which I think is a little different than my original mindset, is that addressing robotics effectively requires using very broad prior knowledge. And I think there’s this idea that a lot of people in robotic learning have, which I think I shared initially, that since people learn things from scratch, maybe robots should learn things from scratch too.
[54:26] So, for example, in some of our early work on large-scale robotic learning at Google, we had this, what we call the ARM Farm project. We set up a bunch of robot arms in a conference room, actually, because we didn’t have a proper lab. It was a conference room. And we had them all grasping objects. And the idea was, well, if they grasp millions of objects, they’ll learn very general grasping strategies.
[54:51] And it basically worked. They could learn to grasp objects, but it was very hard to take it further than that to the next level. Now I can pick up anything, but so what? It didn’t serve as a very good stepping stone for more complex skills. And I think part of that was that we were approaching this from a very blank slate. Let’s start from zero and see if, knowing nothing in advance, the robot could start picking up behaviors.
[55:12] But I think that it’s much, much more practical to get all this to work if you can combine robot experience with knowledge that you can pull in from other sources. For example, I was very skeptical initially about the utility of language. And I think scientifically this is defensible, which is that, hey, animals can do some pretty impressive things. Monkeys can do really cool stuff. But monkeys, as far as I know, can’t speak, at least not very eloquently.
[55:39] So maybe our robots should also be able to do stuff. And they don’t necessarily need to understand language. But I think the subtlety there is what’s important is not language, it’s prior knowledge that you can put in as a scaffold on your learning process. And you can pull in that knowledge in all sorts of ways. Humans don’t necessarily pull it in entirely through language. Humans and monkeys certainly don’t.
[56:01] They do it from observation, from observing other people, other creatures, and so on. So there’s lots of sources of prior knowledge, but the point is that you got to get that prior knowledge in there. Otherwise, you’re actually faced with a harder problem than what humans and animals have to solve. Because if a person had to figure out how to assemble IKEA furniture, but they’ve never actually encountered any article of furniture in their entire life, okay, that would be pretty difficult because they don’t even know what the point of this is or what the end game looks like.
[56:28] So, yeah, prior knowledge is important. And while I’m still a big fan of learning things through experience, I think that my advice to myself would have been, take prior knowledge more seriously.
Ryan:
[56:38] Awesome. Well, thank you so much for your time, Sergey. I really appreciate it.
Sergey:
[56:41] Yeah, thank you for your questions.
