Rob Bernshteyn (00:00):
Morning? Yes. Okay. Well, I'll tell you, speaking of mornings, it wasn't that long ago, a couple of months ago on an early morning, I left my home and an Uber driver picked me up. And he was taking me to an address that I had written down, 396 Treat Avenue in the Mission. Like anyone early morning, you get in the back of the Uber, you get on your phone, you start responding to texts, emails, things that came in over the course of the evening. And 20 something minutes later, I was dropped off, still answering an email. I get out of the Uber, I close the door, respond.
(00:35):
It's dark and I look around and realize I'm kind of in the middle of nowhere. There's really literally nothing around. There's no retail locations, there's no commercial buildings, no receptionists. There's a really shady character half a block away. A little concerned for my safety. Where is this 396? Sergey knows what I'm talking about.
(00:59):
Luckily I kind of move around a little bit and I discover this random door with 396 written on it, real small numbers. In part for my safety and in part because I had a meeting to attend to and no one was welcoming me in, I decided to push the door open and go in. And I could tell you that I literally entered a completely different world. There were robots everywhere. There were people running around like their hair was on fire. Somebody was training a robot to make coffee. Someone in the other corner was training a robot to fold socks. Somebody was training a robot to fold boxes. It was a world that I'd never seen before.
(01:42):
And the most impressionable part of it was when I looked into the eyes of the people that were doing the work. They had a certain sense of purpose about them that was incredibly, incredibly memorable. You could tell they felt that they were part of doing something really, really special. And I can tell you in part, I also felt like I was part of something really special as well.
(02:06):
This was the offices of Physical Intelligence. And the bet that Physical Intelligence made is actually very simple. Can you create one model, can you create one operating system, one brain that can power robots, robots in any different form factor to do all kinds of things? Fold laundry at laundromats around San Francisco, which is happening today. Package boxes in warehouses around the country, which is happening today. All done through one brain, one model, but a different robot, a different form factor, a different set of components that do that work.
(02:54):
Now, Sergey who's here with us today is a co-founder and he's also a professor at Berkeley by the way. And over the last 10 years, he's been thinking about the question underneath all of this. Can we train a centralized brain, a centralized model to learn from all of its physical interactions in the world and get smarter and smarter and smarter? So I'm really proud and excited to welcome to the stage, Sergey Levine.
Sergey Levine (03:31):
We'll try to give you a warmer welcome next time so you don't have to brave the tough streets of San Francisco.
Rob Bernshteyn (03:36):
Well, I'm excited that you told me about the new headquarters as of a few weeks ago. That's wonderful.
Sergey Levine (03:40):
It's in a nicer part of town now.
Rob Bernshteyn (03:41):
I'm looking forward to visiting. I bet the energy's the same though inside. I bet it really is. So look, I remember coming home from, I'm going to age myself, from middle school watching the Flintstones and The Jetsons. And The Jetsons in particular was one I still love the theme song at the beginning. But one thing in particular, I remember the little robot coming out and cooking the food and interacting with the family and wondering to myself, "When are we going to get there? When are we going to have that?" And that was a number of decades ago.
(04:09):
So in your opinion, first of all, why is it taking so long? But also what has changed in your view in the last two, three years that maybe will accelerate this in a way that we could actually have that come to life?
Sergey Levine (04:22):
Throughout the history of AI, there's always been this kind of conundrum where it always seems like things get developed backwards to the order that we might expect, meaning that the things that we expect to be cognitively very complex and demanding actually end up being some of the easier problems in AI. So for example, like chess in the '90s, there was a computer that could beat the world champion at chess, even though we think that chess is sort of the pinnacle of intellectual achievement in some sense. Then 20 more years to beat the world champion at the game of Go. But then after that, another four to six years to be able to produce natural language that sounds coherent. So that to us should seem backwards. But there's a reason for it is that the things that to us feel easier, they're actually easier because we have to be so good at it.
(05:09):
So we have to be good at communicating with each other because that's an essential part of human existence. It's even more important to be very good at motor control. If people weren't good at perception, motor control, kind of the basics of controlling our bodies, we wouldn't survive. So because of that, because we're so good at these things, it's hard for us to appreciate how complex the underlying engineering challenge really is. This is called Moravec's paradox and it's something that the AI community has grown to appreciate over the decades.
(05:35):
In the 1950s, Marvin Minsky famously said that he thought that computer vision would be a summer undergraduate research project that after one summer they would solve and they could move on to the real problems of how to do planning and all this other stuff, and it's still an open problem.
(05:49):
So robotics is very much a more challenging problem than even natural language generation. And furthermore, I would actually say that the problem of physical intelligence, the problem of getting physical systems that are as capable as humans, it actually subsumes all the other AI challenges. Because if you have a machine that can control its body and do all the things that a person can do, well, how do we interact with computers? We interact with computers through our body, right? So if you can do all of those physical things, it actually subsumes the entirety of the AI problem ultimately.
Rob Bernshteyn (06:18):
Well, so you're touching on the complexity of the problem and we as humans love analogies. I know I'm running my own business. It was always important to articulate what my value proposition was relative to something else so people have a frame of reference. And we heard Divesh and Nikesh talk about what you just referenced, large language models. Is it a fair analogy to say that what you're focused on with Physical Intelligence is effectively to build the Claude or the ChatGPT of physical AI?
Sergey Levine (06:48):
Yeah. There's this kind of thesis underlying a lot of modern work on foundation models that generality has a kind of power that you don't get with specialization. So what that means is, to use the language model analogy, if someone was building a machine translation system in the late 2010s, and the way they did is by really focusing only on machine translation, getting paired data of English and French and Japanese and all that stuff, and then language models came along, they would've wiped all that out because language models subsume all that through generality. They're more general, they understand language at a more fundamental level, and they can also do machine translation without needing to be specifically designed for that.
(07:30):
And I think it's very clear to me that robotics is going to go the same way, that the promise of robotic learning is to get rid of the need for meticulous manual engineering of robotic controllers. That's very promising, but it replaces it with another even bigger problem, which is now, if you're going to learn a robotic policy for a specific task, you need to get tons of data of that single task, just like the machine translation example. And that's infeasible and that's why robotic learning didn't really take off prior to the era of foundation models. But if you train a general model on all the different physical skills that your robot could do, now for each physical skill, you might need much less data. And now, it's actually feasible to train a general purpose system to do physical work.
(08:11):
I'll give you an example of this. So I mentioned that in AI, there's this problem that the easy stuff is the hard stuff, right? So there was a researcher, he used to be at EDR at Google, Benjie Holson. He wrote this blog post after the ... There was this Robot Olympics thing in China and after that he wrote a blog post that said, "Well, the real Robot Olympics is not to run the marathon or play table tennis. The real Robot Olympics is all the 'easy stuff.' Can you turn a sock inside out? Can you wash a frying pan with a soapy sponge?" Stuff that's really messy and very physical.
(08:46):
And he had this list of challenges, like the gold medal challenge, the silver medal challenge, all these very mundane home tasks. And any roboticist that looks at this, that is just wild because you start thinking of all the things that can go wrong when you have a soapy sponge in the hand of a robot and there's water going everywhere and it's slippery and so on.
(09:02):
So my colleague, Chelsea Finn, she decided that she wanted to test this new system we had for onboarding new tasks. So we kind of designed a framework where operators could collect data for new tasks and she just wanted to test for it. So the test she did is she said, "Okay, let's go through Benjie Holson's blog post and let's try to see how many of these Robot Olympics tests we could do." She's like, "Okay, you got to budget four hours for each task, go and do it."
(09:26):
And she managed to actually get our operations team to solve nearly all of the tasks. There were a few that she couldn't do because the robot physically couldn't do it. So it turns out that our gripper was a little too wide to fit inside the sleeve of a dress shirt to turn it inside out, so we couldn't do that one. But the majority of, over 90% of these tasks she could do with, I think it ended up being under eight hours of data in the worst case, but usually a lot less than that.
(09:51):
And when she compared this to what happened if you don't start from the generalist model, meaning that you train each test from scratch, same exact algorithm, same exact model design just without the generalist initialization, none of them worked, zero. So that I think illustrates this power of generality.
Rob Bernshteyn (10:06):
Certainly does, but it also brings forth the question of data in general. And then also what I'm curious about is where do you get the data? What's the best type of data to get? When I walked in, I saw people collecting data, creating data. I also know that there are plenty of companies that are creating data and selling that data. I think one of them is actually represented here today. How do you think about data that you would be inclined to purchase versus data you want to create yourself? How do you think about data creation along those dimensions?
Sergey Levine (10:38):
That's a very complex question. So data is a complex question in foundation models in general, but it's even more complex in robotics because there's so many open research problems and the ideal data to get actually changes as the technology changes.
(10:51):
So to give you some examples. Initially, obviously, you have to bootstrap this model from something, so you bootstrap it with good robot data collected by humans, demonstration data. But once you have a model that's capable enough, you can actually instruct it in other ways. One of the things we found is that once our model became sufficiently capable, we could, for example, coach it through language. You could actually ask the robot to do some tasks. We tried to get it to use a new kitchen appliance, air fryer. It tries to do it. It does kind of an okay job. It sort of muddles around.
(11:19):
Eventually manages to use the air fryer, but it takes way longer than it should. Clearly it's kind of making things up as it goes along. And then we basically provide it with substep language commands. We say, "Okay, to use an air fryer, first you grasp the handle, then you pull on it, then you do this, then you do that." And that's a different kind of data, right? That's just a verbal coaching. And because the model already understands physical interaction and can incorporate that verbal coaching information and actually use it to improve.
(11:42):
Another thing you can do once the technology advances is you can use reinforcement learning. The robot can actually learn through trial and error in the real world, so it can leverage autonomous data. Now, of course, it still also needs all sorts of other data sources and it can incorporate human videos and it can incorporate teleoperation data. But the point is the ideal data mix actually changes as the technology evolves. So it's kind of a moving target. So to keep this short, yes, you need lots of teleoperation data, but you also get these new data sources that open up as the tech gets better.
Rob Bernshteyn (12:11):
Now, let me ask you, with generative AI models, one of the things that's being discussed as you know in the media is, okay, can Claude, can ChatGPT, OpenAI do everything? Can it start to eat the deployment layer? And if you draw the analogy to what you all are working on is, to date, you're kind of using third parties for the most part, as I understand, for the robotics for the embodiment of the intelligence that you're creating.
(12:38):
Having said that, are you committed to staying away from going into deployment, to going into the place where you actually produce the robots in certain use case, certain examples? How are you thinking about that decision point that maybe hasn't come yet but may come in the future?
Sergey Levine (12:57):
The range of all the physical work that robots could do in the world is enormous, right? So I mean the kind of very factual answer to your question is we are actually building robotic systems that we are deploying ourselves. To a very large extent, I think we have to because we need to fully understand the entire scope of the problem. So if we are to provide a general purpose model that actually does real things, not just laboratory demonstrations, but actually does real work in the real world, we need to understand the problem very thoroughly and that requires doing a lot of it ourselves.
(13:26):
But at the same time, the range of physical work is so great that there's absolutely no way that a single team can understand all of the different types of robots, all the different types of domains. Some of these domains are extremely specialized. There are companies that have used our models for everything ranging from robotic surgery to controlling an excavator. You really don't want to mix those two. You got to keep that straight.
(13:50):
I don't know anything about excavators. I know very, very little about surgery and I want to leave that to the experts. But if the model has a very general physical understanding, then it could be adapted to all these domains. And I've actually been very, very surprised at what sort of crazy things people can shoehorn this model into.
(14:07):
There was a video one of my colleagues showed me from a company that used it to control a robotic tractor that was bringing feed to cows. They wouldn't have expected that cow robot interaction would be on the bingo card, but it is. And I think what that really shows is that there are fundamental kind of building blocks of physical interaction that once your model understands them, then it can actually be adapted to all these cases efficiently.
(14:32):
So to kind of come back to your point, yes, we have to understand the deployment picture and we have to have very robust systems, but at the same time, we need to be able to leverage the entirety of the ecosystem and provide that physical intelligence layer that other domain experts who know more about those domains than we do can actually use.
Rob Bernshteyn (14:49):
Well, is it fair to say then from a prioritization perspective, it's reach out into deployment to the extent that you need to, but the core driving force is to make the central brain smarter and smarter and smarter as the base for everything going forward, right?
Sergey Levine (15:01):
That's right. And you can almost think of it as a kind of operating system. There's this idea in the robotics community that comes up every 10 years often with some negative connotations of building the robot operating system and it always fails. But I think the reason it always fails is that I think people think about it the wrong way. They think of a robot operating system as a software layer, but an operating system is really something that separates applications from hardware.
(15:26):
And for a robot, that's actually not an operating system in the traditional sense. It's the intelligence layer. If you have a person that's coming to do a job, you don't have to explain to them how to move their body. You just tell them like, "Hey, your job is going to be this and this and this. Let me show you how it's done." You don't need to tell them like, "Hey, fire neuron number 37 in your left forearm in order to pick up this thing."
(15:45):
So the intelligence layer is what's going to allow this decomposition of the problem so that folks that are excellent at building the right hardware can focus on that. Folks that want to build applications can focus on that and they don't have to solve this really, really difficult science problem in the middle. They can use a shared intelligence layer.
Rob Bernshteyn (16:02):
That makes perfect sense. Let me ask you, let's switch gears a little bit and talk about China, which I'm sure questions around China you've had to field in the past. When we were researching the space and looking at what's going on, it just felt like there's a lot going on there and in many ways they've sort of assumed a lot of things of the core language models and are now going into the physical world. How do you think this shakes out in the coming years as it more broadly impacts the United States, our innovation, Silicon Valley and what China's doing and maybe any interplay you see happening between the two?
Sergey Levine (16:36):
Yeah, that's a very complex question. I guess one of the things I would say here is I think it's very important in the United States for us to have a very healthy ecosystem around robotics. And it's fair to say that along some dimensions, the ecosystem in China is further along, certainly when it comes to supply chains, things like that. But in my mind, what I think is very important to have a healthy ecosystem is to have the right tools available to people that want to innovate.
(17:03):
And the trouble is that right now, if you want to build a robotics company, you have this essentially insurmountable problem where you have to build every part of the stack. And that's extremely hard. It's extremely capital intensive. And certainly, there are some companies that have gone very far with doing that, but it is quite limiting. And if there are broadly applicable tools that are available so that somebody who wants to innovate on a particular part of that stack can grab those tools and do the part that they want to deal with, I think that's going to make for a much healthier ecosystem.
(17:30):
It's going to make it feasible, for example, to have an extremely successful business around designing novel hardware. It's going to make it possible to have a very successful business around connecting with a very particular sector of the market without having to also figure out the extremely difficult science problem of intelligence or the hardware design.
(17:46):
So I think that having that middle physical intelligence layer available to everybody is going to greatly strengthen the ecosystem. It's like, okay, with autonomous cars, the same question could be asked for people that work on autonomous cars. Like, "Hey, if you're working on autonomous cars, do you worry about China producing lots of cars?" I mean, some people definitely worry about that, but it's not like that's at the top of the mind for people that want to build self-driving software. And it's because the ecosystem is more decomposed.
Rob Bernshteyn (18:10):
Makes perfect sense. Let me ask you in terms of the metaphor I brought up earlier around LLMs and digital AI, the consensus roughly is that sort of physical AI may be 4, 5, 10 years behind, but that gap will maybe shorten. Do you think it's a fair analogy to say the physical AI is sort of building on the coattails of digital AI and is 3, 4, 5, 10 years behind? What's your sense for that?
Sergey Levine (18:38):
So what I would say is I don't think it's a race so much as it is the natural evolution of AI technology, that basically the physical aspect of the problem kind of subsumes all the others. It's not like there's applications of AI where there's an application, LP, vision and then robotics like the n+1 thing. Robotics actually has to bring together many of those ideas. You need the computer vision, you need to understand language and interact with humans, you need to do planning, you need to do motor control and you need reinforcement learning and all those things need to come together.
(19:05):
But I think even more so because the physical problem is kind of a super set of the digital problem, I actually have this idea, and maybe this is kind of a strange idea, but perhaps in the long run it will prove to be true that if we can actually solve physical intelligence in a very general and conclusive way, whatever we do to figure that out will actually then also enhance AI in all other sectors. Why? Well, because a lot of how we structure the world around us basically stems from our physical embodiment. Computers work the way they do because of how we understand the physical world. We use physical analogies to think about the world. We create interfaces for ourselves that are structured around how we behave and how we think. So I think if we have a machine that understands the physical world the way that we do through embodiment, then it'll be a much better basis for everything else that was designed with humans in mind originally.
Rob Bernshteyn (19:54):
Makes sense. If you close your eyes and you remember that, let's say the ChatGPT moment or that moment when the broad populace was awakened to the power of, let's say, generative AI, have we already had that moment in physical AI? If we haven't, when will we? Will there be a moment? Is it something on a spectrum? How do you think about that in your mind's eye? Because you're thinking about this day to day and many of us are kind of coming in and out of it. What's your best sense for that?
Sergey Levine (20:23):
So I think it's tricky because of that thing that we discussed at the very beginning that for us as people, the things that we find really easy naturally don't seem that impressive. So you see a robot doing a back flip and you're like, "Oh man, that's pretty cool. I can't do a back flip. The robot can do it." But if you see a robot picking up a coffee cup and pouring coffee into it and handing it to somebody, you're like, "Okay, I could do that. That doesn't look so hard."
(20:46):
So I think that perhaps there will be a moment when you can come home and your robot butler is there and you tell it like, "Hey, I would like you to make dinner at 6:00 PM. The guests are coming tomorrow and make sure the kitchen is tidy," and it goes and does it.
(21:00):
But I suspect that it's much more realistic to expect a steady growth capability because the most major milestones will be the ones that won't actually seem that impressive to us. They'll kind of be in the background just working, just doing the thing. So I think people that pay attention might be impressed by what's coming over the next year, but I actually suspect that if we do our job properly, it'll just seem natural that things just work.
Rob Bernshteyn (21:23):
Well, on that note, let me just say that we all at ICONIQ are incredibly fortunate to be involved with you, what you've focused on. You're clearly on the very frontier, on the bleeding edge of innovation. And I can tell you personally, I'm excited to have a robot at home as I saw in The Jetsons, but excited to see how this blossoms over coming months, quarters, years into the vision that you all have there. So ladies and gentlemen, please join me in thanking Sergey.
Disclaimer:
This presentation does not constitute investment advice or an offer to sell or a solicitation of an offer to buy any securities (including, without limitation, any investment fund or investment product referenced herein) which will only be made pursuant to definitive offering documents and subscription agreements.
Certain statements provided in the presentation are made by Sergey Levine, Co-Founder, Physical Intelligence. Neither Physical Intelligence nor Mr. Levine’s views are necessarily indicative of the views of other ICONIQ portfolio companies or their executives, whose experiences may differ from those expressed in this presentation.
ICONIQ did not compensate Physical Intelligence or Mr. Levine directly or indirectly for participation in this presentation, but given his role as a founder of an ICONIQ portfolio company, Mr. Levine may be subject to certain conflicts of interest. ICONIQ holds an ownership interest in Physical Intelligence and maintains an ongoing relationship with the company. As a portfolio company, Physical Intelligence may benefit directly and indirectly from its relationship with ICONIQ, including through access to capital, strategic guidance, operational support, networking opportunities, and other resources provided by ICONIQ. Accordingly, Physical Intelligence and Mr. Levine may have an incentive to make statements that are favorable to ICONIQ in order to maintain or enhance that relationship.
This presentation may contain forward-looking statements based on current plans, estimates and projections. You are cautioned that a number of important factors could cause actual results and outcomes to differ materially from those expressed in, or implied by, the forward-looking statements. The numbers, figures, and case studies contained in this presentation have been included for purposes of illustration only, and no assurance can be given that the actual results of any ICONIQ portfolio company will correspond with the information contained in this presentation. No information is contained herein with respect to conflicts of interest, which may be significant. The portfolio companies and executives highlighted in this presentation are not representative of the entire portfolio or investing history of ICONIQ Venture and Growth. The types, level, and terms of services and resources actually provided by ICONIQ Venture and Growth to portfolio companies may vary. For a complete list of ICONIQ Venture and Growth’s investments please contact us or visit our website.
Certain of the economic and market information contained herein may have been obtained from published sources and/or prepared by other parties. While such sources are believed to be reliable, none of ICONIQ or any of its affiliates, partners, employees, and representatives assume any responsibility for the accuracy of such information.
All of the information in this presentation is presented as of September 1, 2026 (the “Production Date”) (except as otherwise specified), is subject to change without notice, and may not be current or may have changed (possibly materially) between the Production Date and the date actually viewed by you. ICONIQ assumes no obligation to update or otherwise revise any information, projections, forecasts or estimates contained in this presentation, including any revisions to reflect changes in economic or market conditions or other circumstances arising after the date the items were made available to you or to reflect the occurrence of unanticipated events.
ICONIQ is not acting as an adviser or fiduciary in any respect in connection with providing this presentation and no relationship shall arise between you and ICONIQ as a result of this video being viewed by you.
These materials are provided for general information and discussion purposes only and may not be relied upon.
