Loading summary
Transcript105 lines
- [00:00]
A
Foreign.
- [00:11]
B
Hello, and welcome to the Nvidia AI Podcast. I'm your host, Noah Kravitz. As we move deeper into the age of physical AI, advanced humanoid robots are reshaping industries and enhancing daily life. Advancements in AI models and simulation tools are enabling robots to perform complex tasks with greater dexterity and efficiency. And foundation models are playing a key role in generalizing robot tasks. Here with us to pull the curtain back a little more on the rapidly developing world of humanoid robotics is Berndt Borneck. Bernd is founder and CEO of 1X Technologies, a Silicon Valley company that's dedicated to building fully autonomous humanoid robots. Bernd, welcome and thank you so much for joining the AI Podcast.
- [00:53]
A
Thank you, Noah.
- [00:54]
B
So can we start with a little bit about your journey, how you got into the field?
- [00:58]
A
Sure. So, to me, this has kind of been a lifelong journey. So as a kid, I was the kind of kid who picked everything apart to figure out what's inside. I think nothing with motors in it of our kitchen appliances survived in the early days. And I also just. I was very lucky. I grew up with a dad that loved building things. We built soapbox cars in the garage. We went all in, like, welded aluminum frames and chassis. And then I discovered computers, right? And we're back in, like, Commandor 64. And then, like, early, I think it was Intel 286, like the first one, but, like, the first kind of like, computer I had, my had only for me right in my room was like a 486.
- [01:40]
B
Okay.
- [01:40]
A
And I really got into programming, right. Mainly due to games and just being part of the mod community. Making mods for games is a big thing. One of my childhood heroes is John Carmack, and I learned writing code by reading his Quake code. And then, you know, at some point it clicks and you're just like, wait a minute, I can type code and this thing moves, right? And, like, connecting these two realities, right, the digital and the physical, is like. It's just magical.
- [02:08]
B
It's magic. Yeah.
- [02:09]
A
And, yeah, one thing led to another and very early, I think I was like 11, I decided I want to make humanoid robots. And then I follow the field. Follow the field ever since.
- [02:17]
B
Right, right.
- [02:18]
A
So I'm like, I'm the luckiest guy alive, right? I get to do my childhood dream every day in the morning I get up and like, even now I get up and there's a robot walking around in my house, and it's absolutely magical.
- [02:27]
B
That's amazing. So tell us a little bit about 1x then. What inspired I Mean, you've been talking about your journey inspired by everything you've been doing since you were a kid. But how did the company get started? And maybe you can talk a little bit about the unique approach your company brings to building humanoids.
- [02:43]
A
Sure. So I think it comes a lot back to how do you build intelligence? Right. So we talk a lot about physical intelligence, but to me, it's intelligence is intelligence. There are physical aspects and there's digital aspects. And we started out with the Internet and we still do to kind of bootstrap, because it's what we have. But if you want to get to actual AGI, you need to learn in the real world. And if you want to learn in the real world, the same rules apply as in any kind of AI model. And there's almost some poetry in it. Right. Because intelligence comes from diversity. And this is very clear in our digital AI models. You have to have information from all kinds of different dialogues, histories, whatever it is, across the Internet. If you try to create a very good language model to write poems by only reading poems, it's not going to work. You need the diversity of everything that is human knowledge and how we think. Same is true for robots. So if you have a robot that grows up, lives his life and dies in a factory work cell, moving something from A to B is not going to be very intelligent. It's not going to give you a lot of information. So it actually has to happen among people. So the first thing to realize is that for robots to be truly intelligent and also for them to have all the nuances that we appreciate throughout our life, like being careful around your pet, holding the door open for someone that's elderly and generally behaving like we want them to behave, they have to live and learn among us. So that means robots have to be safe. And that, I think, is the thing we spent the most time on in 1x from day one, 10 years ago, is how do you build a system that is as capable as a human? So it needs to be very strong, it needs to be fast, and it needs to generally be very dexterous, but still as safe as a human. And this is kind of counterintuitive because we think about robots as these big, stiff, heavy, dangerous things. Right. At least for me with a robotics background. Right. Because you're used to the industrial robots, right?
- [04:42]
B
I was going to say, I just see the videos of the robot dogs running around and I'm like, oh, they're agile there, but, yeah, they are.
- [04:47]
A
But, like, industrial robots are typically extremely dangerous. Right? It's very high energy devices. And if you get hit by a robot, it's really dangerous. We even have a word for it, right? In robotics, we call anything that's not planned for a collision. Well, turns out that humans collide, like many times every minute, Right. It's all we do. Because this world around us is so unstructured, it's incredibly hard to plan for when you will actually hit the world. So you need to build robots that, like, if we go running because we're in a hurry and we crash around the corner, then might be awkward, might be painful, but it's not dangerous. That's what we want to build. So we set out to do that. I think we've gotten very far. We have a robot now that's just 30 kilos, or about 66 pounds. It's as strong as an adult human. So it can squat or deadlift like 150 pounds. So that's roughly actually human, at least level in power to weight. But also the energy in the system is just incredibly low when moving because we have these tendon drives where we're pulling tendons loosely inspired by muscle, instead of using the more classical, heavily geared industrial approaches.
- [05:47]
B
Okay. And that helps with efficiency.
- [05:49]
A
It helps with efficiency, but most of all, it helps with being soft and compliant. It helps with trying not to nerd out too much on the math here, but I have to. So this principle is actually very simple. It is. When a robot moves its joint, then of course, the joint itself rotates, right, to move the arm. So let's say that that rotates at a certain velocity, a certain speed, and then inside you have a gear. So now let's say the gear ratio is 100:1. Then the components inside, they're also at 100 times faster than the joint. Energy in a system is the velocity squared. This is why you learn in school that a car that goes twice as fast is not twice as dangerous, it's four times as dangerous. So you can just think about it. If you see something spinning at 20,000 RPM, even if it's pretty light, and I ask you to stick your fingers inside, your entire body is going to go, no, I don't want to do that. Because you have this intuition, right? There's a lot of energy in that. And that is true. So if you look at these systems, for a typical robot, it actually has about 10 times the amount of energy due to the drivetrain. And a nice analogy is just think about a kettlebell. So it's kind of like you attach a 1015 kilogram or 30 pound kettlebell to the wrist of the robot. And. And then you go and do everything you're supposed to do. That's like the effective mass or weight just projected from the drivetrain and that's what we get rid of. So it's not just that the tendons are soft, they're not that soft. It's that you don't actually have these high gear ratios. And that means that your system is super dynamic and agile because this is how nature works. We generally have like 0.7 to 1.2 in ratio. That's like the animal kingdom, where the animals are agile, but also it's what makes us safe. Right. And we'll get back to this when we talk about how we learn. But needless to say, we need to be safe to be among people, but also we need to be safe to be able to explore. That's how humans learn. We need to be able to interact with the world and fail, and then the world should be okay, and we should be okay and we can try again. And this has been like a canonical problem in robotics because robots are stiff and high energy, so you can never fail. Of course, we fail all the time. That's how we learn. So that's one. Sorry, no, wrong answer. But that's one. That's the first one. And the second one is affordability. If you have a system that costs hundreds of thousands of dollars, you can to some extent defend that. If it's in a factory working 24 7, it's still a relatively low hourly rate, but it's not going to get you to hundreds of millions of robots in homes so that you can get to actual true intelligence. So those are the two main things we focused on early on, is just needs to be safe, needs to be very affordable, and then of course, it needs to get the job done. So it needs to be able to do the things.
- [08:34]
B
So should we talk about learning?
- [08:35]
A
We can talk a bit about learning.
- [08:36]
B
Let's talk about learning. I've been hearing about foundational models. I've been hearing about reinforcement learning. I've been seeing demos of robots that learn by, you know, a human doing a task, and then whether it's through cameras or physical sensors, the robot learns from that. Or what, in your view, are kind of the most significant advancements in robot learning? I mean, recently. But you can go back because you've been doing this for years now. But what do you think are some of the milestones that really changed the way that we develop robots?
- [09:03]
A
So if we do this from the AI side first because we already talked a bit about how we develop the robots. The field has like a history of. We started with robots, right? In AI, if you go back even to early days of OpenAI and like their Rubik's Cube and also their gym, right? And all of these agentic behaviors with rl, because you get pretty far on this and it's a very interesting way to approach intelligence, right. I would say that in general, all of the early models out of DeepMind, they're actually closer, in my opinion to AGI than a lot of the things we have today, like in the approach that they're taking.
- [09:39]
B
Okay, can you dig into that a little bit?
- [09:41]
A
Well, it's just about how we learn, right? How do you create reasoning tokens and like these long term deep reasoning tokens, right? And you can write them out or try to bootstrap them, or you can have them naturally emerge from kind of like whatever reward you're tuning with respect to how you survive in the world, right? I'd say like for humanity, the ultimate one has been basically survival, right? And I think there's still this kind of open question of exactly how do reasoning emerge naturally just from data? And there seems to be a lot more tangible proof there. But then the realization is we don't have the data, right? So where do we have the data? Well, we have the Internet. So now if we have the Internet, let's use that data. And then turns out that you can get really freaking far. It works surprisingly well. So that's really where I think the last few years have been. And some canonical, very important problems have to a large degree been solved. Right? Now what has not been solved is things like spatial reasoning. We say that these language models have a world understanding and they're kind of like a world model, but they still catastrophically fail on very simple things like just If I move one meter forward, will I collide? Right. It's 50, 50 where I guess it, right? So it doesn't understand physics that well. And in general you can clearly see that the data we have does not contain necessarily disinformation about the world.
- [11:01]
B
Right.
- [11:01]
A
What's also extremely interesting is that all of these different approaches are merging. And this kind of always happens once you go through a full cycle. And we see now in LLMs that they're using unsupervised learning, supervised learning, reinforcement learning, and like everything comes together to solve the problem. Right. It's just the question is mainly in which order do you do this? And a Very interesting. Part of this is what we call grounding, or verifiable truth for your model. So think about when you have a reasoning model. It's pretty good at math, actually starting to be amazing at math, and the same at coding, because you can verify whether or not your solution is correct, right? And as long as you can do this, you can apply reinforcement learning really well. Now, robots are kind of like the ultimate tool to verify truth, because ground truth is literally the ground, right? So, like, when you have robots in the real world, anything is verifiable. Some of it might be quite convoluted to get to an answer.
- [11:58]
B
Oh, right.
- [11:58]
A
But. But everything is verifiable. That's how we ground our reality.
- [12:02]
B
Right?
- [12:02]
A
We have the real world. And I think this is going to be where the biggest impact from humanoids is going to be. It will progress AI in general, not just physical intelligence. Because back to what I said, in my mind, they are the same, right? They're different aspects of the same. So to more directly answer your question, though, I think right now, in robotics, the way we approach this is we have reinforcement learning in the bottom, where we learn how to handle our body and how to handle our own dexterity.
- [12:30]
B
Okay?
- [12:30]
A
And this works pretty well because we can do this on relatively simple tasks, right? So if you think about walking, running, just squatting down, or sitting in a couch and getting up, or all of these kind of things, then the object interactions are pretty simple. The ground is basically roughly flat. Maybe it's not completely flat. You can model this a bit. Not too flat, but it's pretty simple compared to you're peeling a shrimp, which is very complicated with respect to how does this object deform, what's the different forces as this going on? And we're nowhere near being able to simulate that. My favorite example here actually is I also do a lot of mechanical engineering, of course, because we're making robots. And if I have something as simple as I'm going to bolt two things together, then I go and run a simulation on this on a small supercomputer overnight. And what we're simulating is like 0.1 second of a load interaction on this bolt circle, and then I still go and test it in the lab because it's probably not that accurate. So that's the actual accuracy we have of simulation. So if you want to simulate complex things, it doesn't work. But if you want to simulate relatively simple things, it works really well. And that's what we leverage currently in reinforcement learning. So that Means the robots can learn to sit, to squat, to run, to walk, to support its own body in a good manner. It can also learn some dexterity for simple objects. So how do you pick up all kinds of different shapes and handle them with in hand dexterity here? It's extremely powerful because you can do this completely autonomously with no people involved. But in the end, to do this, you now have to move to the next paradigm, which is how do you do the same thing in the real world? And that's something that I'm very interested in these days and we're working a lot on, which is learning in the real world.
- [14:09]
B
Yeah.
- [14:09]
A
And also in rl. So I call this kind of like broadly learning from failure.
- [14:14]
B
Okay.
- [14:14]
A
And what it is, is you have a model which involves a good amount of reinforcement learning in the bottom to learn how to handle its own dynamics. Then you have some expert demonstrations from humans, which is typically teleoperation data. You of course also have all of the Internet because that's how you get a better understanding of a lot of things around us. And the data is there. And then this allows you to get to where you have a task that is non zero in success. Right. So you tell the robot to do something and the robot is able to sometimes succeed.
- [14:45]
B
Right.
- [14:45]
A
And at this point you can actually close the loop and you can say like, hey, now I only need to know whether it succeeded or failed, which is a way simpler problem. And then the robot can learn. But again, this requires you to have a robot that can fail and can fail.
- [14:59]
B
Yeah, right.
- [15:00]
A
And once you can do that, you can just think about it. Right? A robot opening the fridge and it's like his hand slips on the handle, tries again, but oh, it moved a bit because the fridge is stuck in vacuum. So it had to shuffle its feet a bit to not fall over. And it grabs again and like, oh, there, I opened the fridge. So now you have two examples of what not to do and one example of what to do. And as you keep iterating on this, you can kind of like self improve the system, which is incredibly interesting because teleoperation does not scale. Teleoperation is extremely interesting to bootstrap your way into non zero chance of success. But you're not going to be teleoperating millions of robots.
- [15:34]
B
No. Right.
- [15:35]
A
So it is just a tool. I think it has a bad reputation. Everyone's always saying like, oh, is it teleoperator, is it AI? And I understand that because it's interesting to see how far have we managed to push the AI. But I actually look at it the other way around. If you can teleoperate it, we can do it with AI. That's just a question of data. And what really is exciting about this is how can you take all of that data and create this data flywheel that just keeps running? And my dream, of course. Right. Is that you could come back in a million years and the system is just better.
- [16:04]
B
Right, right.
- [16:05]
A
And we're not quite there yet, but that's where it needs to go.
- [16:08]
B
I'm going to shift gears a little bit and ask you kind of about the societal aspects of robotics and humanoid robotics. There's been, I mean, even going back a couple years when LLMs and, you know, LLM chatbots came out and everybody started talking about, well, what does this mean for work? What does this mean for my job as a writer or an accountant or a lawyer or what have you. How do you think about humanoid robots from the human perspective? Right. Because we've been talking about how the robots learn and how important it is for them to be out in society picking up those nuances. What's been your experience and how do you think about how humans react to humanoids? And specifically in the context of work, where robots are coming in and, you know, depending on the job paying, the situation, I would imagine there's a range of reactions from folks, but there's got to be a little trepidation, right?
- [16:52]
A
Yeah, yeah. So let me start with what we're actually doing. Right, yeah. So right now, 2025 for us is about consumer. And there's going to be a lot of people this year with Neos in their homes. And I can't wait to give people back 2.3 hours, roughly. That's the average of work you do at home every day to just keep your house tidy and, like all of the chores around the home, which people generally do not want to do. Right.
- [17:15]
B
Yeah.
- [17:15]
A
You're finally home from work and you'd like to spend time with your family, and now you need to do the laundry. So there we're not so far at least meeting any challenges. That's mainly people being very happy. But it points to a broader thing in society. Right. Which is what do we actually want to spend our time on? What makes us happy. And I think, first of all, we need to ensure that we can keep the standard of living we currently have and preferably grow it. Right?
- [17:40]
B
Yes.
- [17:40]
A
And we're having this enormous shortage of worker coming because we're not having enough kids. I think it's going to be a long time until we kind of run out of things to do, even if we change nothing.
- [17:52]
B
Yeah.
- [17:52]
A
But I think also it's a great opportunity for us to be able to focus on really what matters and like what makes us human. Right. So I have this personal experience on this many years ago. It was back early in college and I had a job just going from house to house for elderly and cleaning just as a side gig. And so many of the people I came to, they'd already cleaned even though they were very old and frail because this was their 20 minutes a week where they got someone to talk to them. So they would just have coffee and chocolate ready and like, hey, sit down and talk with me instead. I mean, it's obvious that that job should have been done by a machine so that people could spend time on interpersonal things. And this is just one example of how, I think how we view the value of humans will change throughout society once we start moving towards an actual true abundance of labor. Right. Because that is what we're talking about here. We're talking about the same kind of shift that we saw with electricity couple hundred years ago. Where you move from this is just never going to like humanity is never going to master energy. Right. That sounds like implausible to like you flip the light switch in the morning and you're annoyed if there's no light. Right. The same is going to happen with labor. But I think it is really going to enable us to reassess what makes us human and what makes us valuable. And I don't think it is basically reducing people in quote to robots in repeating certain labor so that we can get products and goods and services. And I think so far up through history we have never run out of human creativity. I think it's Jens Necessities. Right. It's a good way of saying it. We will find new ways to do or new things to do and how to add value to each other. And I think there will be higher value than what we have today. But always when this happens, of course there's a transition period and that transition period can be painful. And I think that's a big one that we as a society need to discuss and come up with good solutions to. Because any kind of like large technological shift will mean that people need to change.
- [19:51]
B
You need to adapt. Yeah.
- [19:52]
A
You need to adapt. Yeah. Adapt is a good word.
- [19:54]
B
When is Neo shipping? You mentioned Neo's being out in people's homes doing the labor.
- [19:59]
A
I'm going to be kind to my team and not say exactly when it's shipping, but it is shipping this year to.
- [20:04]
B
Sure. And so when you talk about. Can you talk about the things that Neo will be able to do in homes?
- [20:11]
A
Sure. So, I mean, I can go through the things that we're doing right now to be very concrete, and then I can talk about what we plan to do.
- [20:17]
B
Perfect.
- [20:18]
A
So today, what works really well is vacuuming. And you can see that if you go to ttc, to our booth, you can see the robot vacuuming. It's actually very good because you can move things around, et cetera, et cetera. It can tidy, it can do the laundry, it can fold, it can fold. Not everything. 100% yet.
- [20:36]
B
Okay, okay.
- [20:36]
A
But we're getting there. I'm pretty confident we can do that and do a good job on that. Folding is one of these very interesting things, by the way, where when you. Let me first preface this, everything I talk about now are things that we are doing partly with AI and partly with Telio. And the plan is to make it autonomous, but we're going to do that together with you guys as customers. And I want to talk about that.
- [20:56]
B
Okay.
- [20:56]
A
But this is basically what is the platform capable of doing? Right? And folding is just this very interesting thing that in classical robotics is almost impossible to solve because it's like a deformable object that every time you put the shirt down, it looks completely different. And for some reason, this just works with AI. It's actually not so hard to automate folding. And I can't actually tell you why. I'm not quite sure. But, like, that's one of the things we found out pretty early that, like, you know what, this actually works really well of the automation.
- [21:24]
B
Amazing. Yeah.
- [21:25]
A
Yeah. But it is classically seen as extremely hard.
- [21:28]
B
Right. Folding today, common cold tomorrow.
- [21:30]
A
Yeah. So, but then, of course, I'm also very excited about just like having something around the house that you can interact with and I can help with small things. I had this magical experience the other day in my home where I was just sitting there. I had. I was actually a candidate for. For a job coming over and in my house, and we were just sitting there talking and the robot is going around doing his things, right? So it's like wiping the counter, I think, at this point. And then my food comes and I just tell robot like, hey, Neil, can you get the door? That's my food. And Neo goes over and opens the door, gets the food package from the delivery guy who was kind of shocked, and then goes over and puts it on the Counter for me. Right. And you could say, like, that's not that valuable. It's kind of gimmicky. But the sum of all of these things is extremely valuable. So I'm also very excited to see how people will use it because I think this will evolve over time. And we can talk about like the canonical tasks like laundry, vacuuming, cleaning, tidying, like these kind of things, and they're all there.
- [22:30]
B
But yeah, yeah. No, I was thinking to your point earlier though, that you know, during this, when the robot's cleaning the counter and goes, gets the food, you don't have to break your conversation with the candidate. Right. You're spending that time as you were saying, well, what's really important to us, I want to connect with this person and see, are we going to work together, do we have things in common, et cetera? Yeah, that's. That's fantastic.
- [22:49]
A
And it also points to how like almost everything we do that is physical labor is social. Like whenever you do something, you are kind of also navigating a social situation. Even if I just say, hey, Neo, can you get me a Coke in the fridge? Maybe someone's in the kitchen making food. Right. So you're navigating this social situation to do this. And this is not just true in the home, it is also true all out through enterprise. Right, Right. Whether you're in like a retail or a hotel or even most factories, there's a lot of people.
- [23:15]
B
There's other people. Sure.
- [23:16]
A
Yeah. So. So this is why the home and consumer has to happen first before we go into all of these other markets.
- [23:21]
B
Right.
- [23:22]
A
But back to how this works, which I think is quite important, is that if you are one of the people who buy, and I hope you're all right, buys a Nio in 2025. We are selling almost more a journey than a destination. So it'll be. Do you want to be one of the first one to ever have this at home? It is going to be a transformational product. Right. And it's something at least me and I know many others have dream, dreamt of all their life.
- [23:44]
B
Yeah.
- [23:44]
A
And you get to be part of this journey together with us.
- [23:46]
B
Right.
- [23:47]
A
Where we develop this towards an actual abundance of labor. And not everything will work day one, but it will get better every day. It's going to be a lot of fun and we're going to treat you really well and we're going to respect your privacy and we can talk about that too. I think that's a big topic and really make this into a journey that you will never forget. And it's almost like adopt an eo. Right. And I'm not saying it's not going to be useful because it's already very useful. So it will be useful and it will reduce the amount of work you do in your home every day, but it's not going to do everything for you. Right. That's not day one. And some will be autonomous, some will be teleop, and sometimes it'll fail. And that's something you have to be okay with. And I think expectation management here is extremely important. Right. Because we cannot do this alone. We have to do this together with all of you out there and kind of collaborate. Right. On. How do we teach these robots how to behave in society?
- [24:42]
B
Are there things that the end user, and I'm thinking of the consumer setting will be able to do to sort of intentionally try to teach the robot things as opposed to, you know, just interacting with the robots, learning. But like, will the user have the ability, I don't know, almost like, I was going to say like a dev kit, but that's not quite right.
- [25:01]
A
No, I get what you mean.
- [25:01]
B
Yeah.
- [25:02]
A
Let's say to be very specific. Right. Can our customers fine tune models?
- [25:05]
B
Okay, perfect.
- [25:07]
A
Well, most likely the answer to that is yes.
- [25:09]
B
Okay.
- [25:10]
A
We have some stuff that we're very proud of on that.
- [25:12]
B
Amazing.
- [25:13]
A
We'll see when we roll it out.
- [25:14]
B
Okay.
- [25:15]
A
I think it's going to be incredibly important to enable people to use this general robot kind of almost in any manner you want. Right, sure. Because the creativity of the community is going to far outweigh like the creativity of the company.
- [25:28]
B
Right.
- [25:28]
A
And just putting it into a few homes. We've learned so much. Yeah, I can imagine maybe if we were very good, like half of our assumptions are correct, the other half is just completely wrong. And that's just the way it is. Right. And then you iterate.
- [25:41]
B
Before we wrap up, I want to ask you about the world model challenge. And maybe you can speak a little bit about the importance of world models when it comes to robotics and robots learning.
- [25:51]
A
Yes. So this one is a bit harder. So, you know, when we talk about AI, it's very easy to explain like, hey, the robot learns something through demonstration, then it keeps improving it through RL and it gets really good at getting you a Coke in the fridge. The world model is a bit more abstract, but it's very connected to how this happens. So to me, what a world model is is to a large part the ability to predict the future.
- [26:17]
B
Right.
- [26:18]
A
So given that I Know my own state, right? So I know, like, where are my fingers, where, like, I basically have the positions of my joints and everything. I have the forces I'm interacting with on the world. Like, if I'm holding something, I know how heavy it is. I know what I see, I know what I hear, I know what I feel. Now, if I take an action, what will happen? And you can picture this as like a tree of probabilities, right? All of this like simulating forward in time what is the potential things that can happen. Now, this is very interesting because it allows you to do things like validation and figure out, like, am I doing the right thing? And that's kind of like the first way you use role models, practically, is to be able to help the robot improve, because you know what might or might not succeed and whether it succeeded, but. And just validating where your model is better than the previous model. But where it gets really interesting to me is that this actually makes you able to search backwards from the goal. So if I have this perfect simulator of the world, and this simulator would simulate how to peel a shrimp, because it actually has the data of all the force interactions and the vision and even the tactile, the audio of the sound. While this happens, it has all the information, then that basically becomes a search problem for how to get to an intelligent policy. And I think this lies a lot closer to how human intelligence works. But the very interesting thing about world models is that they just scale incredibly well. So if you think about robotics in general, we don't necessarily have the model architecture yet that allows you to scale to, like, infinite scale, but a world model is just incredibly efficient in taking all of the data and turning it into something useful. And then our remaining problem is now, okay, you have this incredibly powerful tool which can basically simulate the future. How do you use this to do something very useful on the robotic side and also generally on the intelligence side? I think there's a lot of exciting work there. It's not fully figured out yet, but there's a lot of early applications which are very powerful, and I think a lot more will come in the coming years. But there's a lot of different definitions of what a world model is, right?
- [28:26]
B
Sure.
- [28:26]
A
So that's what I'm saying. To me, it is basically the ability to simulate any kind of future, given the actions you want to take and your current state. And interestingly enough, if you look at our world model, the data and compute efficiency to, for example, understand physics is just off the charts, right? Because the robot actually has the data.
- [28:45]
B
Yeah.
- [28:45]
A
And this is something you see is very much missing in what I would call maybe the closest to world models that you could see publicly today, which is the video generation models, because they are essentially a subset of a world model. But they really struggle with physics, especially when it gets complicated, like how does a human move and interact with the world, they really struggle. This, of course, is completely different. When you have a robot that actually has spatial data, it is learning in 3D, it is learning directional audio, it's learning tactile, all of these things like interaction forces.
- [29:16]
B
Yeah. Oh, it's fascinating. But you weren't kidding. You have a fun job. This is fascinating stuff. And I don't know, I of course always start to think about it and then I immediately go to the robot in my house, robot my job, and kind of think about that stuff. But thinking about just the way you're approaching these problems and that idea of searching backwards from the goal and just these different ways you have to think about the problems is fascinating. Bernd, before we let you go, for people who want to learn more about any everything you've been talking about, about the company, about the robots, is the website the best place to go?
- [29:47]
A
Yeah, I'd say so. And then follow us on X. We pretty present there, but I think we have an amazing website and it tells you a lot about the product. And then of course, if you have that opportunity, stop by our booth.
- [29:59]
B
Yeah, you have to see it in person.
- [30:00]
A
You have to see it in person. And there will be a lot of places to see NIO in the coming months, I can only imagine. And I can't wait to let everyone actually experience it because it is a product that's very hard to visualize with words and even visualize with video. It's something you have to touch.
- [30:17]
B
Excellent. Well, I can't wait to interact with Aneo. It's going to be an exciting year, to say the least. Bernd Bournick, thank you so much for taking the time to talk with us. And I very much look forward to following your progress following the launch of NE and maybe catch up with you again down the road.
- [30:33]
A
Awesome. It was fun, Sam.