Fei-Fei Li on Spatial Intelligence and Robotics

0:00We are building the next frontier of AI, which is what we call spatial intelligence. At Cynix, we are developing what we call a real-to-sim-to-real pipeline. We can replace all the data, all the evaluation we need in the real environment by using the data that can generate at a scalable way in our digital world. Think about human intelligence. We do a lot of simulation in our head. why there's a very important role simulation plays that real-world data doesn't play, which is counterfactual reasoning. What we are building is a consistent world. Consistence both over space, over time, over different viewpoints, and over different type of interactions.

0:43My new star is I want the robot to work. The world we live in can be multiverse, that we create technology to allow people, builders, developers to act within different spaces. Do you believe we'll ever be able to build robots that have the power efficiency of a human being? How far away are we from this? Is this like five years or this is like never? The TLDR is... Language models transformed how AI understands words. The next frontier is teaching AI to understand and act within the physical world. Following World Lab's acquisition of Scenex, Martin Casado sits down with Fei Fei Li and Yun Zhu Li to unpack the vision behind the deal.

1:25They discuss spatial intelligence, world models, simulation, and why solving robotics will require a new generation of AI built for three-dimensional reasoning, not just language. All right, well, it's great to have you both here. So Fei-Fei, for the listeners that may not have the background, maybe you can give an overview of what World Labs does. Yeah, well, World Labs is a two-year-old startup. I think we should just recognize it's a frontier model lab. We are building the next frontier of AI, which is what we call spatial intelligence. And spatial intelligence is about creating AI that has the ability to generate, understand, reason with, and interact with spaces, whether it's physical or virtual.

2:18And of course, a means to an end towards spatial intelligence is building large world models. And that's what World Labs is mostly focused on. So you've been saying this since the very beginning, which is the machine's ability to perceive and reason about spaces and act on spaces. But I always had the assumption that the acting on spaces was some long-distance future thing, but now you're acquiring a robotics company. And so maybe talk a little bit about the timeliness of this and the intentions. Yeah. So first of all, it doesn't just take robotics to act within spaces or to interact, right? I mean, look at the creative field, whether it's VFX or gaming or design, many use cases, you can create and act within virtual spaces.

3:07And WorldLab's thesis has always been that the world we live in can be multiverse, that we create technology to allow people, builders, developers to act within different spaces. Having said that, the ability to act within the physical space is one of the most exciting and most profoundly important capability of the future AI world. So robotics is very much that. So WorldLab has always believed that robotics is an important application as well as use case of spatial intelligence and world modeling. So by joining force with inviting Cinex and Cinex team to WorldLabs is part of our long-term vision and mission.

3:59We've always committed to that. Amazing. So Yunshu, you're the co-founder of Scenics. So maybe provide everyone with a quick overview of your background and what Scenics does. Yeah, so I'm Yunshu. So I'm currently co-founder of Scenics and also assistant professor at Columbia University. So my research started from my PhD at MIT and then postdoc with Fei-Pei. Really? Yes. That's great. The world is small. The world is small. It is. Throughout my career, my goal has been very simple. trying to help the robots better perceive and interact with the physical world. So I'm a very practical person. I want my robot to work in the real physical environments.

4:38So for Cynics, the unique opportunity we see is that there has been a lot of bottlenecks. Right now, we see faced by the development of general-purpose robots, especially around training and also around evaluations. So at Cynics, we are developing what we call a real-to-sim-to-real pipeline. We want to map the real environments into the digital world that has the best alignments with the real environments. By alignments, we mean that whatever happens in the digital world is also going to happen in the real environments, such that we can replace all the data, all the evaluation we need in the real environments by using the data that can generate at a scalable way in our digital world.

5:19So that is how everything started. in Cynics, we put together a very, very strong and best teams around robotics, robot learning, and also simulation and rendering, trying to build this reel-to-scene-to-reel stack to solve some of the key bottlenecks. It's amazing that you two work together. Yeah, and there is a funny story here because you would think because we work together, he was my amazing post-op, we've been talking about this Cynics and WorldLab integration for a long time. It's actually not true. They came into WorldLabs as a customer. Really? When we released the first version of our generative model called Marble last winter around November, December, Scenics just signed up.

6:02No kidding as a customer. Yes. And I didn't even know what it was. And then I realized this is Yun Zhu's company. I called Yun Zhu. I'm like, wow, this is your company. And then we realized there's so much synergy. Maybe, Faith, could just quickly describe what Marble is? Yeah, Marble is the codename for the base model that WorldLab is been training and iterating on. The fundamental capability right now of Marble that is publicly released is to take a prompt. It can be an image, it can be a few images or a text, and turn that into a geometrically consistent world that can be represented in 3D geometry, whether it's Gaussian splat or mesh.

6:49really what Cynic's team is doing is trying to solve this extremely difficult problem in robotics, which is the lack of data. The lack of data in training, the lack of data in evaluation, this is very, very different from language models, where data is abundant on the internet. And we know that in order for robotics to work, we have to somehow unlock the power of scaling law. But where does that come from? This is something that is a profound problem that everybody's battling with in robotics. It'd actually be great to talk about this energy. You have put together a very, very talented team. You have put together a very talented team.

7:31And so to what extent is there overlap? To what extent is this an extension? Maybe talk a little bit about that. How complementary it is. It's actually the TLDR is very complementary and with a shared mission. So Windu is one of the three technical co-founders. The other two are Changxi Zhen, another Columbia professor who has been a world-class technologist in simulation. And Changxi has his background in also VFX. He worked at Weta. He worked at Tencent. He's been an entrepreneur. Then there's Sonny Hu, who is a phenomenal engineering leader who was also in a startup that was acquired by Amazon many years ago.

8:12So he worked in many different tech stacks in the computer vision field in Amazon. So when we started talking more seriously, I recognized that a couple of things that Scenics has from a talent point of view is extremely complementary to world labs. One is obviously Andrew's incredible thought leadership and just technical prowess in robotics, right? So from really, from hardware, full stack robotics. And even when he was my postdoc at Stanford, at that time, you already had your faculty offer. So you were there only for one year. I wanted you for more than one year, but he had to go become a, have the real job.

8:55So he was a full stack researcher in robotics from modeling to hardware. And, of course, Yunzhu and his students at Cynics was that pool of talent WorldLab hasn't had yet. Then on the Changxi side is just incredible simulation capability, right? He's such a senior researcher and technologist in simulation. And what WorldLab is doing is very much interfacing the world of simulation. So I think what they don't have, obviously, is on the generative model side, as well as the computer vision 3D reconstruction side, we're also very strong at World Labs. So that's a technology that CNIX needs. So together, these two sides come together and make it much more complete.

9:52Fei-Fei's motivation in this is like this is an extension and a compliment to get into robotics. you know having been in your situation which is deciding when to sell a company it would be great to hear from you on like how you think about joining world labs and kind of the fit there and like why you made the decision to do it yeah so at the very beginning we were deciding okay do you want to just keep going but after chatting with feyfe after seeing all the synergies that happen in the middle it just makes perfect sense for the forces to join each other so in any sense as Cynics, what we've been doing is real-to-sim-to-real, is to do dense reconstruction of the environment.

10:27So we capture the appearance of the environment, geometry of the environment, and also the dynamics of the environment, meaning how the environment is going to change when you apply actions. So this dense reconstruction right now is still a little bit on the heavier side. And what World Labs right now has been doing involves a lot of profound capabilities around sparse reconstruction and generations. So we see a lot of opportunities of leveraging like a marble and other like capabilities as World Labs in order to do very efficient reconstructions and modeling of the environment. So can we expect a foundation model for robotics from World Labs?

11:04World Lab is building a foundation model, as you know, Martin. We're building a base model. And as the technology has been evolving some of the most exciting base models are omni models, right? They take multi-modal input. They have multi-modal outputs. And what is a foundation model for robotics? It's very likely going to involve actions. It's very likely going to involve the output of actions in addition to the state of the world. And we're definitely not ruling this out. Yeah, great. So, for example, for the foundation models, it essentially needs to be a multimodal model. So it has to take into account frame text, image, depth, and different kind of modalities.

11:52And action is a very, very important part of that modality. So if you think about frame actions as an input, that essentially affords simulatory. That is going to predict how the environment is going to change when you apply a specific action. When the action is output, this is essentially a policy model. That is trying to predict, giving a specific goal, what should be the action you take in the real environment to get you closer to that goal. So this kind of omni models actually can benefit a lot and actually provide huge amount of values for the robotics communities in trying to understand how to model the environments and at the same time, how to act in the environments.

12:27And this can also act as a backbone for you to fine-tune into specific robotic applications to making sure it's really live up to the reliability and efficiency that's expected by the clients. You know, Yun-Chi, if you don't, if you don't mind a kind of a lay investor question, I see a lot of robotics companies. And a very popular approach right now for the robotics companies that come in is like, we'll use a video model, you know? And like, you know, that's the predominant method where this is, you know, 3D and simulation. It's a very different approach. And so maybe you could contrast, you know, this popular approach of just using video only versus kind of what the ambition here is.

13:07Yeah. So in order to create words with the robot handler, the words, as I mentioned, need to capture the essential structure of the problem. And one of the very important necessary requirements for those words will be consistency. So that is where I actually see there's very, very strong synergies with Marble, because what we are building is a consistent word. Consistence both over space, over time, over different viewpoints, and over different type of interactions. And Marble, the generated words from Marble, also provides an infrastructure, components of that entire world that we believe is necessary for the robot owner.

13:42Imagine if a robot is pushing an object forward, the object just magically disappears, which has been a problem of many of the existing video prediction models. This one provides good enough signal for the robot to know what is the right thing to do. But obviously right now, there has been a lot of investigation on building better and better and stronger and stronger video models. So we actually see a way where some of the infrastructure we build can provide as initial momentums. And to go through this data flywheel of going from this like a more simulation driven models into like a robot policy models, which is going to do the execution in the real environment, collecting new data, the data will come back in.

14:21Where the model doesn't necessarily have to be physics only or learning only, but somewhere in the middle, which be able to capture the essential structure of the problem. But at the same time, be able to scale and become better and better as you accumulate more data. You know, I've worked now, Seifei, very closely for a while. And you've always had this North Star, which has driven this. And, you know, you've articulated variously as kind of 3D and in a number of other ways. And I'm just wondering, for you, is there also a similar philosophical North Star? Or you're more the pragmatic, like I am.

14:55Like build the system, like do the thing. My north star is to make robots work in the real environment. I'm a very practical person. I want the robot to work. One interesting thing that's actually coming from my collaborations with Phoebe during my postdoc, we are building this kind of benchmark. We actually send out surveys asking the general public what they want the robots to do for them. Among the southern tasks we collected, one third of the tasks are about clean. People just don't like to do those like dull and dirty tasks. And those are the scenarios that we really want to make sure we have robotic solutions to deal with.

15:30One thing I really like about Cynix, Martin, especially continuing your question, there's a lot of robotics companies building models and all that. One thing I truly like about Cynix is Renju and his co-founders have such an incredibly pragmatic approach to robotics. Especially they come from academia, right? Sunny doesn't, but Yunzhu and Chanxi come from academia, but their first instinct is work with design partners and customers in real industry, whether it's labs, industry labs, or warehouses, or electronics, you know, assembly. that is such a refreshing, actually, a refreshing way of approaching robotics.

16:22And that really made me very excited to work with it. Maybe this is for you and you, but I'll just be, this is personal curiosity, which is it seems to me that for robotics, you have to be pretty exact. I mean, not perfect, but pretty close. But for the creative use cases, which WorldEps has done a lot of, you kind of don't need to because, you know, I mean, even sometimes like being wrong is stylistic or intentional or whatever. And so from a technical perspective, what is the challenge here for reconciling these two things? Or do they never get reconciled? Like, there will always be two points in the design space.

16:56So they will be reconciled in the long terms, of course. And modeling of the environments doesn't have to be perfect. The model doesn't have to be perfect in robotics. And by the way, this is pure curiosity, but is there like a bit more formal way to say that? Like, what does that mean not to be perfect? It has to be pretty close. So let me put it this way. For example, models over the development of all different kinds of robotic applications has been a very important cornerstone. If you look at all the existing robotic applications, like plane, drones, Roomba, or even for quadruped robots, bipedal robots, model has been the way for them to actually work and be able to transfer from simulation to the real robots.

17:38But if you look at those locomotion robots, like quadruped robots, bipedal robots, they can be walking on snows, they can be walking on bushes, But you don't need to have a simulator. You can simulate all the bushes and snows very precisely. You need to have a simulation that captures the essential structure of the problem and do whole different kind of randomizations inside the digital environments. So that is what we're aiming for. So basically, with Cynics and together with WordLabs, we're trying to investigate what is the level of fidelity we need to model the massive world besides the robots such that we'll be able to transfer the robotic systems training the simulated environment and digital worlds back into the real scenarios.

18:16As an investor, I've heard other researchers, say like Sergey Levine, say simulation will always eventually deviate from the physical world and real world data collection is absolutely critical. And so maybe talk a little bit about like the viability of this approach where simulation is a cornerstone as opposed to some other approach. So they don't contradict with each other. So if you're thinking about the simulation, simulation is essentially trying to predict how the environment is going to change when you apply the actions. And this is essentially a model of the world that doesn't necessarily have to be pure physics.

18:55It can be a combination between both physics and also learning. We are collecting real-world data. We will be using those real-world data. It's just at different stages of this data flywheel. Maybe at the very beginning, we have stronger emphasis on we have more physics to making sure we have the right consistency and right structure for us to learn the world, for us to train the robot policies. But as we accumulate more and more data, both through data collection and also through the collaboration with our clients, we'll have the data that will be moving more towards more learning-based, like modeling of the environment.

19:28So this kind of transition and also this kind of data fly-off is really enabling factors of both getting the best of both physics and the geometry and consistency, as well as all the power and magics from the data and compute. I want to add to this and be slightly philosophical here, is there isn't a binary choice between simulation or no simulation. All this come in together to make robotics work. Think about human intelligence. We do a lot of simulation in our head. you know why there's a very important role simulation plays that real world data doesn't play which is counterfactual reasoning is that you play out events that hasn't happened or cannot happen or you don't have enough data to make it happen in real world and while you play it out you learn how to act in it humans do this all the time we probably don't you know we just i know you were at the World Cups.

20:34Congratulations to Spain winning. I'm sure in the planning of every game, there is simulation, whether it's digital or on the whiteboard or whatever. That simulation, the role simulation plays is counterfactual reasoning. And that's really important in robotics because we just do not have, cannot possibly have enough real world data for that. Here's a real-life example, the industry of self-driving cars. Weibo has officially said they use billions of hours of simulation. And actually, Weibo is more simulation-heavy than just real-world data-heavy. So these are real examples. And as you know, Martin and Yunju, too, cars are the simplest kind of robots.

21:23Yeah, 2D, yeah. Yeah, so clearly simulation plays a huge role in robotic learning. I also want to add to that. So if you put things more specific, simulation can provide two levels of benefits. The first one is reliability, and the second one is efficiency. So for reliability, if you're thinking about a robotic system working reliable in the real environments, You need data to provide systematic coverage of all the state space and the variations that robots might encounter. That's how you can learn how that is robust. So with simulation, you can do systematic randomizations and control and variations of lighting, frictions, geometries, object types, and also all different kinds of physical parameters to making sure you have sufficient coverage of the state space.

22:11So this is what can give the robotic systems reliability. And second is about efficiency. So right now, many people are doing teleoperation. And if you look at many of the teleoperation devices, imagining all the actual skeletons you are using, you are actually collecting the data at a speed that is actually slower than humans actually doing the task. But for many of our clients, human speed to them is not good enough. They want faster than human speed. So for the robot to move faster, it's not as simple as just drive the robot faster because the gravity doesn't change. But in simulation, you can do systematic speed-up of the robot's behaviors to train the robots such that it considers all the dynamics, changes of the environment.

22:53So this is what can give our clients, for them, efficiency. So both for the reliability and efficiency, there are some kind of very unique values where simulation can provide. You've talked about the technology and the platform, what it does. Maybe talk about the specific use cases people use it for. There are essential two specific use cases, especially around both training and also around evaluations. Starting from the evaluations. So evaluation is something people often overlook in the robotics. But if you are training robotic models, you have to know how well it works. And that is the only source of information for you to iterate.

23:32By the way, every AI person really understands what evals are and uses it all the time. non-AI people, it often means something a little different. So maybe it's even worth just describing specifically what you mean by evaluation. Okay. So what I mean by evaluation is you'll be able to understand for this specific checkpoint, how well does it perform? Does it perform, for example, 95 % of the time or 99.9 % of the time? And the key criteria people use in industry is how long does it take? How long in work time does it take for you to distinguish between a checkpoint that is 90 % from a checkpoint that is 92 points.

24:12And if you only do that in the real environment, that's just taking so long for you to do the distinguishments. And if you really think about also the robotic evaluations right now people are doing in the real environments, the iteration speeds is multiple orders of magnitude slower than iterations of those language models. So not only is the robotic tasks very varied, very diverse. Oh, you actually have to do the thing. Yeah, yeah, yeah. Right, right. Like atoms have to move through space. Exactly. But only the laws of physics have to be obeyed. And have you watched those robotics videos? Every video has like 10x, 8x because it moves so slowly.

24:52Exactly. So not only is it slow, it's dangerous, it's costly, but at the same time the speed is also like multiple hours of magnitude is like slower. So some of our clients actually need this digital environment that can be used to evaluate their robotic systems. And because our digital environment has proven alignments with the real world, so meaning whatever happens in the sim is also likely to happen in the real environment. If a checkpoint is working better in the simulation, it's also highly likely to also work better in the real environment, as we have also been discussed in the blog post.

25:25So that actually gives our clients very strong confidence in actually using the data, using the signal from the digital environment to do scalable, safe, and much faster evaluations of their robotic systems. Great. So that is on the evaluation. Then on the training. So on the training side, so basically, like I also mentioned, it's about controllability. So you want to control all the different possible variations of states, parameters, lighting, frictions, physical parameters, like even object geometry, object types. So you want to make sure you have sufficient coverage of all different kinds of scenarios, such that you'll be able to generate informative data for your robots to be robust.

26:05And this is just going to be so hard to do just in the real environments. Like we discussed, if you do teleoperation, the speed at which you're collecting data is slow. You're also limited by how many robots you have, how many teleoperation devices you have. There's a whole different kind of challenges around all the data operations around it. But in simulation, everything can be controllable, everything can be systematic, and everything can be understood at a level where you know exactly and making claims about exactly what distribution you have covered. To develop confidence about within the distribution, we know the robot will work.

26:40So those kind of confidence and efficiency and scalability is something that our clients also value to use our digital words for the training of robotic systems. Here's a crazy thing. Even before Cynics and we are talking, our inbound customers from Marble were already seeing this kind of demands. We just cannot serve these customers, but we are already getting a lot of phone calls from robotics, early stage robotics companies who are developing their models all the way to downstream, very pragmatic use cases. And we're seeing these needs. When people hear you're going into robotics, what they're going to envision is you're pulling out a 3D printer and you're going to be making hardware and then you're going to be programming the brain of a robot and sticking it in the robot and then you've got a robot.

27:32And I don't think that's what you guys are talking about here. So maybe talk about what, you know, where this fits in the life cycle of creating a robot and like where you will end and where the rest of the ecosystem will begin. So what we've been building, you can imagine, is an infrastructure with the softwares around these infrastructures for people to, for them, build worlds, such that robots can learn and evaluate. And this infrastructure is naturally model-agnostic and embodiment-agnostic. So I just want to be very clear, just because this is actually a very subtle, I mean, for you it's obvious, but it's a very subtle point, which is, from what you said, that's not building a robot.