Training Robots in Worlds That Don't Exist | World Labs with a16z

Last week, World Labs announced its acquisition of SceniX, bringing together two teams working on one of AI's biggest unsolved problems: how to give machines a true understanding of the physical world. Martin Casado sits down with Fei-Fei Li, co-founder and CEO of World Labs, creator of ImageNet, and pioneer of spatial intelligence, alongside Yunzhu Li, co-founder of SceniX and assistant professor at Columbia University. They discuss why World Labs acquired SceniX, how simulation can unlock the next generation of robotics, and why training robots may require a fundamentally different approach than training language models. The conversation explores real-to-sim-to-real pipelines, world models, robotics foundation models, evaluation, synthetic data, and why the future of AI depends not just on understanding language—but on understanding and interacting with the physical world. Timestamps: 00:00 - Intro 01:08 - World Labs & SceniX 06:04 - Marble & the Data Bottleneck in Robotics 07:13 - How the Two Teams Come Together 10:55 - Building a Foundation Model for Robotics 12:35 - Video Models vs Real-to-Sim-to-Real 19:38 - Why Simulation is Essential for Robot Learning 23:01 - Training, Evaluation & Real Customer Use Cases 29:17 - Humanoids, Semi-Structured Environments & the Grand Challenge 36:56 - Integration Plans & What Success Looks Like in Two Years Resources: Follow Fei-Fei Li on X: https://x.com/drfeifei Follow Yunzhu Li on X: https://x.com/YunzhuLiYZ Follow Martin Casado on X: https://x.com/martin_casado Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Fei-Fei Li?

  We are building the next frontier of AI, which is what we call spatial intelligence.

• 0:06Speaker 01

  At Scenics, 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 environments by using the data that can generate at a scalable way in our digital world.

• 0:20Fei-Fei Li?

  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.

• 0:34Speaker 01

  What we are building is a consistent world, consistent both over space, over time, over different viewpoints and over different type of interactions. My north star is I want the robotic world.

• 0:45Fei-Fei Li?

  The world we live in can be multiverse, that we create technology to allow people, builders, developers, to act within different spaces.

• 0:55Martin Casado?

  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?

• 1:06Fei-Fei Li?

  The TLDR is...

• 1:08Erik Torenberg?

  Language models transformed how AI understands words. The next frontier is teaching AI to understand and act within the physical world. Following World Labs' acquisition of Scenics, Martin Casado sits down with Fei-Fei Li and Yunzhu Li to unpack the vision behind the deal. They 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.

• 1:39Martin Casado?

  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?

• 1:49Fei-Fei Li?

  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. And 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.

• 2:28Martin Casado?

  Yeah, 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 futuristic, but now you're acquiring a robotics company. And so maybe talk a little bit about the timeliness of this and the intentions.

• 2:48Fei-Fei Li?

  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. And World Labs' 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.

• 3:34Fei-Fei Li?

  So robotics is very much that. So World Labs 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 Scenics and Scenics team to World Labs is part of our long-term vision and mission. We've always committed to that.

• 4:01Martin Casado?

  Amazing. So Yunzhu, you're the co-founder of Scenics. So maybe provide everyone with a quick overview of your background and what Scenics does?

• 4:09Speaker 01

  Yeah, so I'm Yunzhu. 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-Fei.

• 4:22Martin Casado?

  Really?

• 4:23Fei-Fei Li?

  Yes.

• 4:24Martin Casado?

  That's great.

• 4:24Fei-Fei Li?

  The world is small.

• 4:25Martin Casado?

  Yeah, it is.

• 4:27Speaker 01

  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 a real physical environment. So for Scenics, the unique opportunity we see is that there has been a lot of bottlenecks. Right now, we see faced by the developments of general-purpose robots, especially around training and also around evaluations. So at Scenics, we are developing what we call a real-to-sim-to-real pipeline.

• 4:57Martin Casado?

  Okay.

• 4:57Speaker 01

  We're wanting 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. So that is how everything started. In Scenics, we put together a very, very strong and best teams around robotics, robot learning, and also simulation and rendering, trying to build this real-to-sim-to-real stack to solve some of the key bottlenecks.

• 5:33Martin Casado?

  It's amazing that you two work together.

• 5:35Fei-Fei Li?

  Yeah, and there is a funny story here because you would think because we work together, he was my amazing postdoc, we've been talking about this, Scenics and World Labs integration for a long time. It's actually not true. They came into World Labs as a customer.

• 5:51Martin Casado?

  Really?

• 5:51Fei-Fei Li?

  When we released the first version of our generative model called Marble last winter, around November, December, Scenics just signed up.

• 6:02Martin Casado?

  No kidding, as a customer?

• 6:04Fei-Fei Li?

  Yes. And I didn't even know what it was. And then I realized this is Yunzhu's company. I called Yunzhu, I'm like, wow, this is your company? And then we realized there's so much synergy.

• 6:18Martin Casado?

  Maybe Fei-Fei just quickly describe what Marble is?

• 6:21Fei-Fei Li?

  Yeah, Marble is the code name for the base model that World Labs has 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 and/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. Really, what Scenics 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.

• 7:04Fei-Fei Li?

  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 is battling with in robotics.

• 7:25Martin Casado?

  It'd actually be great to talk about this synergy. You have put together a very, very talented team. You have put together a very talented team. And so to what extent is there overlap? To what extent is this an extension? Maybe talk a little bit about that.

• 7:35Fei-Fei Li?

  Yeah, that's actually...

• 7:36Martin Casado?

  How complimentary it is?

• 7:37Fei-Fei Li?

  It's actually, the TLDR is very complimentary and with a shared mission. So Yunzhu is one of the three technical co-founders. The other two are Changxi Zheng, 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 Sunny Hu, who is a phenomenal engineering leader who was also in a startup that was acquired by Amazon many years ago. So he worked in many different tech stacks in the computer vision field in Amazon. So when we started talking more seriously,

• 8:21Fei-Fei Li?

  I recognized that a couple of things that Scenics has from a talent point of view is extremely complimentary to World Labs. One is obviously Yunzhu'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, have the real job. So he was a full-stack researcher in robotics, from modeling to hardware. And of course, Yunzhu and his students at Scenics

• 9:06Fei-Fei Li?

  was that pool of talent World Labs 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 World Labs 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 Scenics needs. So together, these two sides come together and make it much more complete.

• 9:51Martin Casado?

  Fei-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.

• 10:07Speaker 01

  Yeah, so at the very beginning, we were deciding, okay, do we want to just keep going? But after chatting with Fei-Fei, after see all the synergies that are happening in the middle, it just makes perfect sense for the forces to join each other. So in any sense, at Scenics, what we've been doing is real-to-sim-to-real, is to do dense reconstruction of the environment. So we capture the appearance of the environment, geometry of the environments, 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

• 10:52Speaker 01

  like Marble and other like capabilities at World Labs in order to do very efficient reconstructions and modeling of the environments.

• 10:59Martin Casado?

  So can we expect a foundation model for robotics from World Labs?

• 11:05Fei-Fei Li?

  World 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, they take multimodal input, they have multimodal outputs. And what is a foundation model for robotics? It's very likely gonna involve actions. It's very likely gonna involve the output of actions in addition to the state of the world. And we're definitely not ruling this out.

• 11:41Martin Casado?

  Yeah, great.

• 11:42Speaker 01

  So for example, for the foundation models, it essentially needs to be a multimodal model. So it has to take into account from text, image, depth, and different kind of modalities. And action is a very, very important parts of that modalities. So if you think about frame actions as inputs, that essentially afford simulator. 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, given a specific goal, like 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.

• 12:28Speaker 01

  And this can also act as a backbone for you to find to into specific robotic applications to making sure it's really live up to the reliability and efficiency that expected by the clients.

• 12:39Martin Casado?

  You know, Yunshu, 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.

• 13:07Speaker 01

  Yeah, so in order to create words with the robot handler, the words, as I mentioned, is to capture the essential structure of the problem. And one of the very important and necessary like 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 world. Consistent both over space, over time, over different viewpoints and over different type of interactions. And marble, the generated words from marble is also provide an infrastructure, a components of that entire words that we believe is necessary for the robot handler. Imagine if a robot is pushing an object forwards, the object just magically disappear, which has been a problem of many of the existing like video prediction models.

• 13:50Speaker 01

  This one provides good enough signal for the robot to know like 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 like a video models. So we actually see a way where some of the infrastructure we build can provide as initial momentum. And to go into this data flywheel of going from this like a more simulation driven models into like a robot policy models, which gonna do the execution in the real environment, collecting new data, the data will come back in, where 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 times, be able to scale and become better and better as you accumulate more data.

• 14:34Martin Casado?

  You know, I've worked now with FeiFei very closely for a while and, and, 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? Like build the system, like do the thing?

• 14:58Speaker 01

  My North Star is to make robots work.

• 15:00Martin Casado?

  Amazing, yeah.

• 15:00Speaker 01

  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 FeiFei during my postdoc, we are building this kind of benchmark. We actually send out surveys asking the general public, what do you want the robots to do for them? Among the thousand tasks we collected, one third of the tasks are about cleaning. People just don't like to do those like a Tao and dirty tasks. And those are the scenarios. So we really want to making sure we have robotic solutions to deal with.

• 15:30Fei-Fei Li?

  One thing I really like about Scenics, Martin, especially continuing your question, there's a lot of robotics companies building models and all that. One thing I truly like about Scenics is, is Rindu and his co-founders have such an incredibly pragmatic approach to robotics. They, especially they come from academia, right? Sunny doesn't, but Rindu and Changxi come from academia, but their first instinct is work with design partners and customers in real industry, whether it's, is labs at industry labs or warehouses or.

• 16:12Speaker 01

  Electronics.

• 16:13Fei-Fei Li?

  Electronics, you know, assembly. That is such a refreshing, actually, a refreshing way of approaching robotics. And that, that really true made me very excited to work with them.

• 16:26Martin Casado?

  Maybe this is for Yunshu, but I just think this is, 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 World Labs has done a lot of, you kind of don't need to, because, you know, I mean, you know, 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'll always be two points in the design space?

• 16:56Speaker 01

  So 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.

• 17:08Martin Casado?

  And by the way, is there, again, 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.

• 17:17Speaker 01

  So, so let me put it this way. For example, models over the developments of all different kind 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 environment.

• 17:38Martin Casado?

  Ah, I see.

• 17:38Speaker 01

  But if you look at those locomotion robots, like quadruped robots, bipedal robots, they can walking on snows, they can walking on bushes, but you don't need to have a simulator. It can simulate all the bushes and snows like 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 Scenics and together with World Labs, we're trying to investigate like what is the level of fidelity we need to model the massive, massive worlds 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.

• 18:16Martin Casado?

  As 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.

• 18:39Speaker 01

  So they don't contradict with each other. So if you think about the simulation, simulation essentially trying to predict how the environment is going to change when you apply the actions. And this is essentially a model of the worlds. It doesn't necessarily have to be pure physics. It 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 like a data flywheel. Maybe at the very beginning, we have stronger emphasize on we have more physics to making sure we have the right consistency and right structure for us to learn the worlds, 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,

• 19:22Speaker 01

  we'll have the data that will be moving towards more towards more learning based, like modeling of the environments. So this kind of transition and also this kind of data file is really an enabling factors of both getting the best of both physics and the, and the geometry and consistency as well as all the power and magics from the data and compute.

• 19:42Fei-Fei Li?

  I want to add to this and be slightly philosophical here is there isn't a, a binary choice between simulation or no simulation. All this come in together to, 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, while you play it out, you learn how to act in it.

• 20:28Fei-Fei Li?

  Humans do this all the time. We probably don't, you know, we just, I know you were at World Cups.

• 20:34Martin Casado?

  I was at the World Cup.

• 20:35Fei-Fei Li?

  Congratulations to Spain winning. I'm sure in the planning of every game, there is simulation, whether it's digital or on the, 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, the industry of self-driving cars. Weibo has officially said they use billions of hours of simulation.

• 21:10Fei-Fei Li?

  And and actually, Waymo is more simulation-heavy than just real-world data-heavy. So these are real examples. And and as you know, Martin and Yunzhu, too, cars are the simplest kind of robots.

• 21:23Martin Casado?

  Yeah, 2D, yeah.

• 21:24Fei-Fei Li?

  Yeah, they so so clearly, simulation plays a huge role in robotic learning.

• 21:30Speaker 01

  I also want to add to that. So like there are, 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 reliably in the real environment, you need data to provide systematic coverage of all the state space and variations that robots might encounter. That's how you can learn of how is 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 kind of physical parameters to making sure you have sufficient coverage of the state space. So this is what can give the robotic systems reliability.

• 22:15Speaker 01

  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're actually collecting the data at a speed that is actually slower than human actually doing the task. But for many of our clients, human speed to them is not good enough. They want faster than human speeds. So for the robot to like move faster, it's not as simple as just drives a 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 environments. So this is what can give like our clients for them efficiency.

• 22:57Speaker 01

  So both for the reliability and efficiency, though there are some kind of like very unique like values where simulation can provide.

• 23:05Martin Casado?

  You've talked about the technology and the platform, what it does. Maybe talk about the specific use cases people use it for.

• 23:10Speaker 01

  Um hum. There are essential like two specific use cases, especially around both training and also around evaluations.

• 23:17Martin Casado?

  Okay.

• 23:17Speaker 01

  Yeah, starting from the evaluations. So evaluation is something like people often overlook in the robotics. But if you are training like robotic models, you have to know how well it works, and that is the only source of information for you to iterate.

• 23:32Martin Casado?

  Yeah. By the way, a lot of, a lot, every every AI person really understands what the 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.

• 23:45Speaker 01

  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, ninety-five percent of the time or ninety-nine point nine percent of the time? And the key criteria people use in industry is how long does it take? How long in work, workload time does it take for you to distinguish between a checkpoints that is ninety percent from a checkpoint that is ninety, ninety-two point? And if you only do that in the real environment, it's just take so long for you to do the distinguishment. 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.

• 24:30Fei-Fei Li?

  Yeah.

• 24:30Speaker 01

  So not only is like the robotic tasks very varied, very diverse.

• 24:35Martin Casado?

  Oh, yeah, because like you actually have to do the thing.

• 24:37Speaker 01

  Yeah, you have to actually do the same, right? Like bricks and glass.

• 24:40Martin Casado?

  Adams have to move through space.

• 24:41Fei-Fei Li?

  Yes.

• 24:42Speaker 01

  Exactly. But only there are the laws of physics have to be obeyed.

• 24:44Fei-Fei Li?

  And have you watched those robotics videos? Every video has like ten X, eight X, because it moves so slowly.

• 24:52Speaker 01

  Exactly. So not only is slow, it's dangerous, it's costly, but at the same time, the speed is also like multiple orders of magnitude like slower. So some of our clients actually needs this digital environment that's can be used to evaluate their robotic like 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 environments, as we have also been discussed in the blog post. So that actually give our clients very strong confidence in actually using the data, using the signal from the digital environment

• 25:33Speaker 01

  to do scalable, safe, and much faster evaluations of their robotic systems.

• 25:39Martin Casado?

  Great.

• 25:39Speaker 01

  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 making sure you have sufficient coverage of all different kind of scenarios such that you'll be able to generate like informative data for your robots to be robust. And 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 like how many robots you have, how many teleoperation device you have. There's like a whole different kind of like challenges around all the data operations around it.

• 26:24Speaker 01

  But in simulation, everything can be controllable, everything can be systematic, and everything can be understand at the, at the 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. So those kind of confidence and efficiency and scalability is something that our clients also value to use our like digital words for the training of robotic systems.

• 26:50Fei-Fei Li?

  Here's a crazy thing. Even before, as Cenix 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, from robotics, early-stage robotics companies who are developing their models all the way to downstream, very pragmatic use cases. And we're, we're seeing these, these needs.

• 27:18Martin Casado?

  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. And 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, will begin.

• 27:46Speaker 01

  So what we have been building, you can imagine, is a infrastructure, like with the softwares around these infrastructures for people to, for them build words such that robot can learn and evaluate. And this infrastructure is naturally model-agnostic and embodiment-agnostic.

• 28:05Martin Casado?

  So I just want to be very clear, just because this is actually a very subtle, I mean, for you it's obvious, but this is a very subtle point, which is, um,

• 28:12Martin Casado?

  from what you said, that's not building a robot, it's building an environment which another company can place their robot brain to navigate and to learn.

• 28:21Speaker 01

  Yeah, so for our customers right now, they have all different kind of robots. Some are using, for them, single robot arms, some are using bim manual, some are using a fixed arm, some are using like a mobile manipulators, some using grippers, some are using some more elaborate versions of the N D factors. So our platform right now is just naturally embodiment-agnostic. We can very easily integrate different kind of robotic embodiments, be able to put them into the words we generated, we digitalized, such that we will be able to give those individual robots capabilities of doing the right tasks and at the right levels of reliability and efficiency in the real environments.

• 28:57Martin Casado?

  Yeah.

• 28:57Speaker 01

  And we are also, for example, model-agnostic. So we can just using the data generated by our words to train different models, either from scratch or doing post-training of existing foundation models, like vision, language, action models, or word action models. So to us, it doesn't matter. We just want to making sure we have the infrastructure, we have all the words such that the robot can work reliably in the real environment.

• 29:21Martin Casado?

  You know, you, you, you have told me that you think a lot of the predictions around humanoids are a little bit aggressive and we're likely to see more constrained rollouts, like warehouses or whatever. Can you talk a little bit about that and like how that impacts what you're going to be tackling here at the, like, World Labs?

• 29:41Speaker 01

  So that's a very good question. So if you look at, for example, all the progressions of robotic applications in the real environments, it has always followed the trend from going from fully structured environments into semi-structured environments and then into unstructured environments.

• 29:56Martin Casado?

  Yeah.

• 29:57Speaker 01

  For fully structured environments, what do we mean is that you have knowledge and the control over all the configurations within the environments.

• 30:03Fei-Fei Li?

  Like factories.

• 30:04Speaker 01

  Like factories or for them car manufacturing lines. Those has been automated for decades.

• 30:09Martin Casado?

  Yeah, yeah, yeah.

• 30:09Speaker 01

  And then you have, for example, semi-structured environments, which you have certain controls over the environments, for example, like the Amazon warehouses or for example, like restaurants, hotels, where you have certain control over the environment to just make the task easier for your robots. But there are obviously many other like objects or for example clothes, those are the objects you don't have control.

• 30:31Martin Casado?

  Yeah.

• 30:31Speaker 01

  And then for the unstructured environments, it's like your home and my homes.

• 30:34Martin Casado?

  Yeah.

• 30:35Speaker 01

  Those is, I would say, the, the grand challenge.

• 30:38Martin Casado?

  Especially, especially my house, trust me, three dogs, five-year-old.

• 30:43Fei-Fei Li?

  Yes, dogs.

• 30:44Speaker 01

  Exactly. If you're thinking about where does the robustness coming from? Robustness coming from a sufficient coverage of the scenarios that robots might encounter. So it's so much easier and more approachable, at least like right now, to focus more on the semi-structured environments before we move on to fully unstructured environments. So we will move into that direction. It's just we want to take a more sustainable and more realistic approach towards it.

• 31:09Fei-Fei Li?

  I think your point here is that humanoids mimics human body, and evolution has optimized human body for unstructured environment. And so our fingers, our legs are not the best apparatus to do one thing. But for example, if, if, if our only goal as a species is to climb trees, we will not have this body necessarily, right? So we'll have different kind of fingers. But what humans end up having are involved, evolved into is this, this body shape that can be very general, but not necessarily best at everything.

• 31:51Fei-Fei Li?

  And that is for the survival of unstructured environment. But, but, but from a business point of view, from a pragmatic. Um, technology point of view that this unstructured environment and a generalized body is actually the hardest problem to solve. It's not necessarily even the right way to solve the problem. It's we specialize, so we take more specialized body to solve a narrower problem. But the challenge for Scenics is that to be more body agnostic, so that their infrastructure can serve different bodies and different semi-structured environments.

• 32:34Martin Casado?

  Yeah, a common lens to look at exactly this question is an economic lens, right? Which is, you compare it to like generative L L Ms, they can create a prose or a code 10,000 times faster than a human being, a bunch cheaper than a human being. So the economic case makes sense because our brains aren't very efficient at that. However, our brains and our bodies are very efficient at 3D navigation, right? You know, like moving through the world or picking things up. And so this is just a prediction question, but do you believe we'll ever be able to build robots, at least in the foreseeable future, that have the power efficiency of a human being when it comes to menial tasks?

• 33:16Martin Casado?

  So let's say just basically, you know, minimum wage or something like that. Like, how far away are we from this? Is this like five years or this is like never?

• 33:24Speaker 01

  I think it's gonna take a very long time. So if you're really thinking about like robots in the real environment, in the end, it will always be a system. So every working robot in the real environment is a system work. It's you need to be very mindful and thoughtful about how the system are coming together. The hardware, the software, the brain, even like to the details of, for example, what's the friction coefficient of your fingers? So there's a lot of things you have to consider to make these things a reality, and it will take iterations. But what I am excited about is that I have always been at the state of the art of robot learning and also trying to push the state of the art forwards.

• 34:02Martin Casado?

  Yeah.

• 34:02Speaker 01

  But the state of the art is always moving faster than I expected.

• 34:06Martin Casado?

  Yeah, yeah, yeah.

• 34:07Speaker 01

  So what's I'm focusing on and trying to investigate right now is very different from, for example, I started my PhD. So this is a speak to how fast the whole ecosystem has been evolving and all the moving pieces started coming together of building these robotic systems. But we also have to be like calibrated about our predictions. So we will see a lot of like progress, but to achieve, for example, human level efficiency and capabilities, it will take longer.

• 34:34Fei-Fei Li?

  Martin, the hardest thing in today's AI is to have the right measured optimism.

• 34:39Martin Casado?

  Right, it's totally true, yeah.

• 34:43Fei-Fei Li?

  I mean, even L L Ms does not have human brain efficiency. Human brain operates on 30 watts.

• 34:50Martin Casado?

  Yeah, that's true.

• 34:51Fei-Fei Li?

  That's so we are far from that, so.

• 34:53Martin Casado?

  But that, but I mean, performance to power it may be close, right?

• 35:00Fei-Fei Li?

  In narrow tasks like soccer.

• 35:02Martin Casado?

  Like generating an image or software engineering, that it is, right? Yeah, I think so. I don't think we're anywhere close when it comes to robotics. Does this change how you think about your, like strategically, the level of ambition that your team can go after? I mean, does it, is it changed that or is it still very much in line with what you expected to do when you started?

• 35:21Speaker 01

  It definitely changed the trajectory in a very profound manners. So we see a lot of unlock in be able to do this whole process, do the modeling of the environments, you know, much more efficient and much more scalable manners, especially in partner together with World Labs. And I also want to add to Fei Fei, if you're thinking about, for example, the current states of the language models.

• 35:42Martin Casado?

  Yeah.

• 35:42Speaker 01

  So those are models that's with incredible capabilities.

• 35:45Martin Casado?

  Yeah.

• 35:45Speaker 01

  But still, you don't just blind trust it to book your flight tickets or make your hotel reservations. You still, hopefully, there's still a person who's reading the output from those language models.

• 35:55Martin Casado?

  Yeah.

• 35:56Speaker 01

  But that is very different from how people and will be using, for example, robotic models. Because for robotic models, out of the box, the robot has to work reliably in the real environment.

• 36:06Martin Casado?

  Yeah.

• 36:06Speaker 01

  So, and we don't even have the data. We don't even have all the necessary infrastructures around those for the robots to just out of the box work reliable in the real environment. So for that reasons, be able to create this digital worlds, this scalable digital worlds where the robot can learn and evaluate within, that's it's gonna unlock so much more potentials for be able to replace all the like costly and unsafe data in the real environments with the data generated from the words for the robots to be able to do scalable learning and evaluations.

• 36:38Martin Casado?

  Yeah, I've seen, I've seen many of kind of these kind of integrations. They actually work very well at this stage when they have this much alignment, which is great. But there's always like this question of, do you integrate now into what's happening now or do you keep things quite separate and provide kind of like a long-term trajectory that will, you know, be realized, you know, in the year timeframe? How are you thinking about this, Fei Fei? Is this something that integrates right away or is this kind of a separate longer term?

• 37:06Fei-Fei Li?

  This is a great question. I think at this point, you know, Yunzhu, Changxi, Sunny, Justin, Ben, and I have been talking about this. At this point, we are going to take it thoughtfully. We're not rushing to integrate everything from code base to teams because I think Scenics does have a very, um, well-thought, and I wouldn't call it standalone completely, but fairly contained tech stack as well as their customers, as well as the, the kind of products they're building. We're gonna take time.

• 37:47Fei-Fei Li?

  We definitely will, we already have a simulation side as well as the, the potential base model, um, action condition model side, we already are starting to talk. And also, they are using Marvel as a internal customer, and so we, we will be integrating, but we're not rushing to blend the team as like full salad bowl.

• 38:10Martin Casado?

  How are you thinking about geographies with this? Will Scenics move? Is it gonna stay in the same place?

• 38:15Fei-Fei Li?

  We're gonna, Yunzhu is gonna move.

• 38:18Martin Casado?

  Oh, well, welcome. Here, I'm moving to San Francisco. Yeah, Florence to the Renaissance, perfect.

• 38:23Fei-Fei Li?

  I think we, our labs is officially becoming a bi-coastal company where the headquarters in San Francisco. I've, you know, I live in Palo Alto, I feel like I'm in a different state. We, we, but we, but I'm actually excited that we're gonna have an office in New York that can help us to attract talent on East Coast. And also, we have been talking about making sure that in both offices, we set up the robots so that we get to basically test out and, and mature our engineering stack so that we can work with robots remotely because we have to do that for, for our customers anyway.

• 39:07Martin Casado?

  So maybe just to be very concrete, Fei Fei, maybe let's just pencil out, like, what is the, the perfect success case in two years? Like, what product do you have? Who's engaging with it? How do they use it? Just, just a crisp, like, what this becomes.

• 39:20Fei-Fei Li?

  I would be very happy that Scenics team, World Labs team will have validated, um, customers in, in a small number of important vertical use cases where our system, our infrastructure has proven to be truly beneficial to their automation needs. And these customers became our lighthouse examples to scale our business.

• 39:58Martin Casado?

  Now, how early, let's say someone listening to this is running a robotics company, how, at what stage should they engage with World Labs? Is it really early on? Is it somewhere in the middle?

• 40:12Speaker 01

  So right now, for our customers, because we are building this kind of real to sim to real pipelines, like where the simulation is essentially the words, we're gonna provide the training and evaluation grounds. Some customers, they need only the real to sim part. They want to digitalize the task they care about and be able to do the evaluations of their robotic systems. Some customers, they need this real to sim to this entire pipeline, such that they will be able to like have policies like running on their hardware. So our platform is also designed in a way that is flexible, depending on what our clients needs. And at the same times, the clients we are working with are actually pretty close to the deployments, like a stage. So basically, they are working on very, very practical tasks.

• 40:57Speaker 01

  Those tasks, when replaced, when we have robotic solutions that are there, can just create value immediately. And they have like, at least like tens or hundreds of like this kind of situations they are thinking about to do the automations for. So as, like together with World Labs, we'll be able to develop reliable solutions for those scenarios. As we have already showed, we have a number of scenarios already instantiated in our blog post, and we'll be able to further our investigation to see how they can actually solve the key like, like requirements and, and also constraints faced by the real world deployment.

• 41:32Martin Casado?

  Great, I want to be very specific about this. Is it ever too late or too early to call World Labs if you're a robotics company?

• 41:38Speaker 01

  No, I, we, we want everybody to call us. We want to learn about your use case.

• 41:43Martin Casado?

  Wonderful. If you're listening, if you're listening to this and you're anywhere close to a robotics project or robotics company, please track World Labs.

• 41:50Fei-Fei Li?

  Yes, thank you. Definitely open for business.

• 41:55Speaker 01

  Yeah, we are open for business.

• 41:56Fei-Fei Li?

  Not too early.

• 41:57Martin Casado?

  All right. If you're doing robotics, call World Labs. Thank you both very much for coming.

• 42:01Fei-Fei Li?

  Thank you.