明年量产 5 万台人形机器人 | 1X 创始人 Bernt Børnich 访谈

今天和我坐在一起的是 Bernt Borick，1X 的创始人兼 CEO。你们明年想出货 5 万台人形机器人。那我们能不能先请你讲讲，你们到底怎么从纯研发走到在 12 个月内真正出货这么多机器人？

Today I am sitting down with Bernt Borick, the founder and CEO of 1X. You want to ship 50,000 humanoid robots next year. So can we start off with you just walking me through how do you actually go from just R&D to actually shipping that number of robots in 12 months?

好。我们现在所在的这个 Hayward 工厂，满产产能大约是 1 万台。这基本上就是今年的目标。我们不会做到 1 万台，因为产线满产的时候已经接近年底了。但与此同时，我们也在建 San Carlos 工厂。要实现这个目标，我们必须并行建设。那个工厂满产是每年 10 万台。所以加起来是每年 11 万台，但当然大部分产能要到明年很晚才会真正上线。这里真正的卡点不只是出货 5 万台，而是确保它们不会被退回来。对。所以我们必须以合理的速度扩大产量。原型很容易。这听起来像老生常谈，但确实如此。原型很容易。量产真的很难。我们在这里解决了很多问题，才走到对扩大产量有信心的地步。但每当你迈出产量的下一个大台阶，你总会发现上一个台阶没发现的问题。这其实就是统计学，对吧？你做 100 台时发现的问题，和做 1000 台或 1 万台时发现的问题不一样，因为罕见的事情在统计上会变得显著，你的良率真的必须提上来。所以我们希望能找到一个好的平衡：花时间真正迭代产品，把经验教训反馈到整个机队里，然后在跨过这些质量门槛后，进入越来越大的产量。但目标是 5 万台，所以工作量会很大。

Yeah. So I mean the factory we're in now in Hayward is built out to a capacity fully ramped about 10,000 units. So that's really been like the goal of this year. We're not going to do 10,000 because we're going to be pretty much towards the end of the year when the line is fully ramped. But in parallel we're building out San Carlos. So to make this happen we have to do some parallel building. And that's going to be 100,000 units at full capacity per year. So 110k per year but of course most of that volume is going to come online pretty late next year. The real gating item here is really kind of not just shipping 50,000 units but ensuring that you don't get them back. Right. So we have to kind of scale our volume at a rational rate. Prototypes are easy. Like it's a cliche, but it's true. Prototypes are easy. Production is really hard. We've turned through a lot of issues here to kind of get to where we're confident in ramping volume. But whenever you take like the next big step on the volume, you always find things that you don't find on the previous one. So like it's just statistics, right? When you do a 100 units, you don't find the same things that you find when you do a 1,000 or 10,000 because things that are rare just get statistically significant and like your yield really needs to come up. So we hope to find a good balance where we take the time to really iterate on the product and like get the lessons learned and get that back into the fleet and then go into larger and larger volumes as we kind of cross these gates, quality gates. But the goal is 50,000 so it's going to be a lot of work.

你刚才说的时候我在想汽车。对于汽车，如果你只是自己测试某辆车，你开着它到处跑，就能体验到拥有这辆车的大部分情况。但对于人形机器人，人们可能让这个机器人做出的随机、极其不寻常的行为，数量比汽车要多好几个数量级。所以从 0 到 5 万辆汽车已经极其困难。做到 5 万台机器人，我觉得可能还要难一到两个数量级。你们怎么为此做规划？

I was thinking about like cars when you were talking and with a car, if you're testing some car just yourself, you're going to experience most of what it is to own a car just driving around. Whereas with like a humanoid robot, the number of, you know, random super unusual behaviors that someone might have this robot in is like multiple orders of magnitude more than a car. And so going from like 0 to 50,000 cars is incredibly difficult. Going to 50,000 robots is like I think probably like one or two orders of magnitude more difficult. How do you kind of plan for that?

是的，我觉得有利有弊，对吧？人形机器人零件很多，相当复杂，但它的零件数量跟汽车不一样。我看到很多公司把人形机器人和汽车做比较。我觉得如果你这么做，你应该退一步想想你造的是什么，因为一辆车大约有 5 万个零件。一个设计良好的人形机器人大概有 1000 个零件，而一辆车大约重 4000 磅。一个设计良好的人形机器人如果要安全，希望重量低于 70 磅。所以从这个意义上说，它其实是更简单的系统，尽管单个复杂度和运动部件数量更大。所以这方面你占了一些优势。但你说的当然也完全正确：它部署的环境约束少得多。所以你会看到更多奇怪的情况。这就是现实。我觉得很大程度上归结为预期管理。过程会颠簸，不会完美。我真的很希望能提供非常好的客户服务，确保无论发生什么，我们都善待你，真正解决这些问题，迭代改进，确保你不必反复面对同一个问题。但当然，我们也需要把一部分产量部署到一些相当结构化、变化更少的环境里，而不是让 5 万台机器人去做 5 万种不同的任务。那样是管理不了的。

Yes, I think there there's pros and cons, right? So a humanoid robot is a lot of parts. It's quite complicated, but it is not the same number of parts as a car. And I see a lot of companies kind of compare humanoid robots to cars. And I think if you do that, you should kind of like take a step back and think about what you've made because a car is like 50,000 parts. A well-designed humanoid robot is probably about 1,000 parts and a car is like 4,000 lbs. A well-designed humanoid robot is hopefully less than 70 if it's going to be safe. So it is actually a simpler system in that sense, although like the individual complexity and number of moving parts is larger. So you get some on that side. But you're of course entirely correct also that like it's a way less constrained environment that you're deploying into. So you're going to see a lot more weird. That's just the way it's going to be. I think a lot of it comes down to expectation management. Like it's going to be bumpy. It's not going to be perfect. I really hope we can provide a very good customer service and ensure that whatever happens, we're treating you well and we really figure this out, iterate on this and make sure that you don't have to face the same problem repeatedly. But of course we also need to deploy into some quite structured more like less variance type environments for parts of the volume and not have 50,000 robots into 50,000 different tasks. That would not be manageable.

有些员工把机器人放在自己家里，我想你就是其中之一，让这个机器人在这个环境里走动、体验，你比其他人领先了几年。另外，你家里可能有过多个版本的机器人。所以我想知道，从个人体验层面，把第一代机器人放在家里是什么感觉，以及它如何演进，信任是如何随着时间建立起来的。

Some employees have the robots in their own home and I think you're one of them that's just got this robot walking around this environment and like experiencing it and you're kind of like a few years ahead of everyone else. The other thing too is you probably had multiple versions of your robot in your home. And so I'd like to know what it feels like on just a personal experience level to have the first evolution of that bot in your home environment and also how that's progressed and like how the trust has, you know, been built over time.

对，对。其实挺有意思的，因为距离我第一次在家里做所谓的试点已经大约 3 年了。当时我实际上把我们的上一代机器人 Eve 放在家里，那是个轮式人形机器人，体型更大更重，但能力很强。那很有趣，我们学到了很多，它也是我们为什么必须加速推进到完全双足机器人的主要灵感之一，这样才能在家里以及其他多样化环境中完成所有这些任务。腿对于操作其实也非常有用。这是惨痛的教训。但我觉得演进过程出奇地慢。这需要很长时间，对吧？这往往是因为你发现某个东西不太行，你需要大量数据，然后你需要迭代硬件，需要重新部署硬件，然后获得新数据。而我们并没有进入那么多家庭。所以跑这个循环需要时间。但幸运的是我们开始得很早，所以有很多时间思考这些事情。但我觉得对我来说，信任方面必须跨越的重要障碍是安全，这是我们仍在非常努力解决的问题，我们有一个很大的目标，今年要以更正式的方式证明安全性。我对此非常兴奋，希望年底能分享更多。

Yeah. Yeah. So it's actually kind of interesting because it's about 3 years now since I kind of had the first like let's call it pilot at home, where I actually had our previous robot, Eve, the wheeled humanoid, which is a bit like bigger and more heavy but very capable at home and it was a lot of fun and we learned a lot and it was also one of the main inspirations for why like we had to kind of speedrun getting to like the full biped to be able to do all these tasks around the home and and also in like other diverse environments. Legs are actually very useful also for manipulation. It was a hard lesson learned. But I think the evolution over time is surprisingly slow. Like it takes a long time, right? So and this is often because you find something that doesn't really work and you need some significant data on this. Then you need to iterate on your hardware and you need your hardware to kind of redeploy and now you get new data. And we're not in that many homes. So like it takes time to kind of run that loop. But luckily we started very early so we've had a lot of time to think about these things. But I think mainly to me the kind of important barrier that has to be crossed on the trust side is safety and that's something that we're still working really hard on and we really have a big goal to kind of prove out safety this year in a more formal manner. So I'm very excited about that and hopefully we can share some more about that towards the end of the year.

你们怎么决定，可能机器人在实验室环境里能做一千种任务，但就像你说的，在把它送到有孩子在场的地方之前，你们想达到某个安全标准。从创建一项新任务到真正把它放到环境里、在自己家里测试，这个过程有多快？

How do you kind of decide on there may be, you know, a thousand tasks that the robot can do in like a lab environment, but then like you said kind of there's this safety bar that you want to hit before you ship it to, you know, your kid is around it. How fast do you go from, you know, creating a new task to actually having that in the environment and like testing it in your own home?

这个系统的一个美妙之处在于，由于它的操作方式非常接近人类，所以像人一样操作机器人其实相当直接。所以这种“专家在场”或“专家操作员”的概念，让你能在实现自动化之前很久就真正测试很多任务，这非常有用。嗯，能给你一种客户印象。

So, one of the beautiful things about the system is that since it's so close to a human in how it operates, it's actually pretty straightforward to also operate the robot as a human. So, this concept of kind of like an expert in place or like an expert operator allows you to really test a lot of tasks way before you're able to automate them, which is very useful. Um, gives you kind of like the kind of customer impression.

带着这个生活是什么感觉？硬件完成这些任务的能力如何？

How is it to live with this? How well capable is the hardware to do these tasks?

所以我认为大部分事情我们已经测试过了，现在真正的大问题是如何实现自动化。在这方面，我们采取的方法与这个领域的大多数玩家非常不同。我非常相信我们可以解决通用问题。我们不需要去单独自动化每一个任务。任务很多，复杂性很高，而且变异性也比你想的要多得多。所以解决每一个任务就像是一个很深的兔子洞，对吧？我把人形机器人更多地看作是人类数据、人类智能与机器之间的桥梁。我们设计的机器人尽可能与我们相似。不仅在移动方式上，而且在与世界互动的方式上，它们的刚度、柔顺性，所有这些方面。这样我们就能利用互联网上存在的关于我们的所有视频和其他数据，训练模型，使我们的系统在几乎所有任务上都具备非常通用的能力。虽然目前不如专门化那么强，但它能让你更快地解决完整问题。我认为这通常也是家庭所需要的。所以对此我非常兴奋。我们新成立的 1X 世界模型实验室，团队非常棒，我们本月晚些时候将启动第一批新的大型训练运行，我非常期待看到新模型的表现。现在还有点早，但已经有一些非常非常有希望的结果，我认为我们很有机会在这次训练运行中解决通用问题。

So I think largely most of the things we've already tested out, and really the big thing now is how do you automate it. And here we follow a very different approach than most of the players in the space. I'm a very big believer in we can solve the general problem. We don't need to go and automate every single task individually. There's a lot of tasks and there's a lot of complexity and also there's a lot more variability than you would think. So it's kind of a deep rabbit hole, right? To solve every task. And I see humanoids almost more like this bridge between human data and human intelligence and the machines. We've designed robots that are as similar to us as possible. Even in how not just how they move, but how they interact with the world, how stiff they are, how compliant they are, all these things. So that we can take all this data that exists of us in just video and everything else out there on the internet and train models that allows our system to be very generally capable across essentially all tasks. But less capable than if you specialize right now, but it allows you to solve the full problem way sooner. And I think that's generally what's also needed for the home. So I'm very excited about that. The new 1X world model labs that we stood up, amazing team, and we're kicking off the first new big training runs now later this month, and I'm very excited to see how the new model is going to turn out. It's a bit early yet, but there's some very, very promising results there, and I think we have a pretty good shot at solving the general problem on this training run.

不？好吧。但达到这一点有点像是一个自举问题，对吧？

No? Okay. But getting to it is kind of a bootstrap problem, right?

因为你训练一个模型，它非常擅长尝试所有这些任务，并且尝试得不错，有时成功，有时失败，这给你更多数据，你把这些数据融入训练，现在你得到更多数据。所以现在模型能力更强了。然后你就这样爬山。但这里真正重要的是拥有如此强大的模型，以至于你不需要人类在回路中生成数据。你只需要大量部署的机器人。然后你就能获得关于成功和失败的非常有价值的数据。你基本上在现实世界中闭环学习，对吧？然后你需要一个能力强的模型。你需要一个安全的机器人，因为如果不安全，当它尝试做它不太知道如何做的事情时，它会伤害自己或环境。然后你需要大量部署的机器人。所以这是另一个理由，看看我们能否在明年年底前让 5 万台机器人落地。

Because you train a model that is very capable of essentially attempting all these tasks, and it does a good job of attempting them, and sometimes succeeds, sometimes fails, and this gives you more data, and you fold this into the training, and now you get more data. So now the model is even more capable. And then you kind of hill climb like this. But the really important thing here is to have models that are so capable that you don't need a human in the loop to generate the data. You just need a large fleet of deployed robots. And then you will get very capable data of both successes and failures. And you close the loop in the real world essentially to learn, right? And then you need the capable model. You need a safe robot because if it's not safe, then it will harm itself or the environment when it's trying to do things it doesn't quite know how to do. And then you need a lot of deployed robots. So that's one more reason to see if we can get 50,000 of them out by the end of next year.

在扩大制造规模方面，从第一台到前 10 台再到前 100 台，这个过程是怎样的？这一路上最大的瓶颈是什么？

On the scaling the manufacturing side, what has the process been like going from the first one to the first 10 to the first 100, and what have been the biggest bottlenecks along that journey?

所以在某种程度上我们有点幸运，系统从第一天起就是为可制造性设计的。所以即使你把这看作一棵技术树，以及我们如何选择制造机器人，它也有点不同。通常你会用普通的电机，它们不那么强大，所以你需要非常复杂的齿轮系统，这些齿轮系统会给系统引入很多能量和其他自然界没有的人工动力学。所以现在你需要大量传感器和高速控制等所有东西来掩盖它。你有点像是在对抗物理。我们采取了非常不同的方法。我们有一些非常强大的电机，我们内部开发并制造，使我们能够做这些肌腱驱动，而且它们实际上要简单得多。不仅性能更好，而且大大简化，比如更少的零件、更少的复杂性、更少的公差、更少的传感器，以及所有可能出错的东西。所以从这个意义上说，系统是相当可制造的。当然，缺点是，由于这是一种新型系统，没有供应链。所以我们必须自己制造所有东西。这花了我们十年时间，对吧？公司现在已经 11 年了。所以我们真的必须爬这些生产过程的成熟度阶梯。但现在这已经相当成熟了，这真的是一种超能力，因为我们所有东西都是内部完成的。所以无论发生什么，我们都可以很快回去迭代，确保在上游修复它。所以我认为我们现在最重要的指标是，从我们在 CAD 中对整个系统进行重大更改，到一台新机器人下线，大约需要四周。我们花了十年时间才建立起这种势头，让我们可以在硬件上如此快速地迭代。但这让我们能够从生产线上获取反馈，并将其反馈给工程部门，在大约 4 周内迭代。然后你就可以真正调整你的制造过程。这样你就可以挑出装配中犯的任何错误，或者在质量、良率、校准等方面难以做到的事情。但没有什么秘密武器。你只需要磨过所有这些问题，最后你真的要祈祷你会用完问题。如果你修复了足够多的问题，你就会用完问题。

So to some extent we're a bit lucky that the system has been designed from day one for manufacturability. So even if you think about this as a tech tree and how we choose to build robots, it's a bit different. So while you normally would have normal motors which aren't that strong, so now you need very complex gear systems, these gear systems introduce a lot of energy and other artificial dynamics into the system that nature doesn't have. So now you need a lot of sensors and high-speed control and all things to kind of mask it. You're kind of fighting physics. We've taken a very different approach. So we have some very capable motors we develop in-house and also manufacture that allow us to do these tendon drives, and they're drastically simpler actually. So not just more performant but drastically simpler, like less parts, less complexity, less tolerances, less sensors, and all things that can go wrong. So the system is quite manufacturable in that sense. Now the downside of course is that since this is a new type of system, there's no supply chain for this. So we have to manufacture all of it ourselves. And that's taken us a decade, right? Like the company is 11 years old now. So we really had to climb the maturity ladder of these production processes. But now that that is getting quite mature, it's really a superpower because we do it all in-house here. So whenever anything happens, we can just go back and iterate very fast and make sure we fix it upstream. So I think the most important metric we have right now is we use about four weeks from we do major changes on the full system in CAD until a new robot walks off the line. And it took us a decade to build that momentum where we can iterate so fast on the hardware. But that allows us to take the feedback from the line and feed that back into engineering and iterate in about 4 weeks. So then you can really tune in your manufacturing process. So you pick off any mistakes that's done in assembly or anything that's hard to do with respect to quality, yield, calibration, all these things down the line. But there's no secret sauce. You just need to grind through all these problems, and in the end you really cross your fingers that you're going to run out of problems. If you just fix enough of them, you're going to run out of problems.

你还相信吗？

You still believe?

是的，我还相信。但既然这是唯一的方法，那么你只需要能够真正衡量自己，比如我能多快迭代，因为那是我能多快解决这些问题的速度。我们看到问题越来越少。所以似乎正朝着正确的方向前进。所以迟早我们会用完问题。

Yeah, I still believe. But like since that's kind of the only way to do it, then you just need to be able to really benchmark yourself on like how quickly can I iterate because that's how quickly I can kind of turn through these problems. And we see less and less problems. So there seems to be going in the right direction. So sooner or later we'll run out of problems.

当你开始把这些机器人放进非公司内部的人家里时，我认为明年年初的机器人会和明年年底的机器人看起来非常不同。你从一月份客户那里得到的反馈会非常不同。你知道，你可以把这些融入给十二月份的客户。你如何尽可能快地整合来自现场机器人的新信息和新反馈，然后基本上启动新的制造线来扩大规模？

When you're like starting to put these in people's homes that aren't inside of the company, I think the bot at the beginning of next year is going to look significantly different than the bot at the end of next year. And the feedback that you get from January customers are going to be very different. You know, you can incorporate that for the December people. How do you kind of as quickly as possible incorporate new information and new feedback from the robots in the field and then basically spin up new manufacturing lines in order to scale that?

现在你可以通过车队监控、数据和分析做很多事情。以前这要难得多。所以我认为这是非常重要的一部分，但没有什么能真正取代贴近客户、真正倾听客户并获取反馈的需要。

There's a lot you can do now with just like fleet monitoring and data and analytics. That was very way harder to do before. So I think that that's a very important part of it, but there's nothing really that replaces the need to just like stay close to your customer and really listen to your customer and get that feedback.

然后，当我们在这里谈论迭代时，你是在整个机器人群体上进行迭代。所以一开始我们有一个大目标，至少对于第一批产品，我们也会做大量的再制造。所以如果我们发现需要修复的问题，我们不仅为新机器人修复，也应该为现有机器人修复。

And then also, when we talk about iteration here, you iterate on the entire fleet. So we do have a big goal in the beginning, at least for the first units, that we will do a lot of remanufacturing also. So if we find things that we need to fix, then we fix it not just for new robots; we should also fix it for the existing robots.

那是不是就像字面意义上的“召回”那样？

Would that be like literally like quote unquote recall, recall back in?

嗯，但不一定是召回，对吧？大多数情况下，产品相当小且轻。所以通常你可以在现场维修。如果你发现我们需要给机器人换新的左脚，我们就在现场更换左脚。我们不一定召回机器人，但严格来说你可以称之为召回。

Well, but not necessarily recall, right? Most of it is like the product is pretty small and light. So generally you can service it in the field. And like if you figure out that we need new left feet on a robot, we replace the left feet in the field. We don't necessarily recall the robot, but technically you could call it a recall.

我在看那个手的视频时，心里想，这让我想起史蒂夫·乔布斯的一个理念：如果你看苹果电脑的内部，它非常华丽。人们会好奇，为什么会这样？因为史蒂夫想设计出内外都美观且实用的东西。我觉得看你的机器人时，也有同样的感觉：即使没有穿上“衣服”或“皮肤”，看起来也很美。你对于设计出连内部都好看的东西，有什么样的理念？

I was looking at the hand video and I was thinking to myself, this reminds me of this idea from Steve Jobs where if you look at the inside of like an Apple computer, it's gorgeous. And people were kind of wondering to themselves, like, why is this? And it's like, 'cause Steve wanted to design something that was both beautiful on the inside and the outside and also functional. And I think when I look at your robot, it has that same feel where it's like beautiful to look at even when it doesn't have like clothes on or skin on. What's your kind of like philosophy on designing something where even the internals look good?

我可能在这方面有点特别迟钝，因为我认为足够先进的工程就是艺术，而这显然是先进的工程。但我认为机器的美来自于简单，而不是复杂。所以你拿一个极其复杂的东西，设法将其简化为最简单的形式，即能够完全解决问题的最简单实现，这其中有一种真正的美。我认为你经常可以在系统中看到这一点；你可以看到，从第一性原理来看，这只是系统的最小实现，这种简单中有一种真正的美。这确实是我希望我们能交付的。而且这本身就自然而然地产生，对吧？你不需要为了内部美观而设计。如果你在如何设计系统方面做好了工作，那么这就会自动发生，因为它应该始终是解决问题的最小复杂度实现。这就是你实现规模化的方式。

I'm probably like blunted here in a special way because I think sufficiently advanced engineering is art, and this is clearly advanced engineering. But I think the beauty of the machine comes from the simplicity, not the complexity. So you take something that's extremely complex and you manage to boil it down to the simplest form, the simplest implementation that can fully solve the problem, and there's a real beauty to that. And I think often you can see this in systems; you can see that, first principles-wise, this is just the minimum implementation of the system, and there's a real beauty to that simplicity. And that's really what I hope we can deliver. And that kind of comes from itself, right? You don't need to design for it to be beautiful inside. If you've done your job on how you design the system, then this just automatically happens because it should always be the minimum complexity implementation that solves the problem. That's how you achieve scale.

你有没有过这样的感觉：看着某个东西，对自己说，它显然还没达到，因为它看起来不简单，不美观？

Do you ever have a feeling where you're looking at something and you say to yourself like it is clearly not there yet because it doesn't look simple, it doesn't look beautiful?

哦，是的。每一天，每次我看着机器人，我看到的都是所有错误的地方。所以从这种意义上说，我并不幸运。就像其他人看到某个东西会觉得这很美。我会说：不不不不，这是错的，这是错的，这是错的。我认为这可能永远都会这样。所以我不是问这个问题的合适人选。但我确实认为，每次我们能进行简化时，我都能发现美。我认为过去一年中最大的简化，例如线束，就是确保线缆的复杂性大幅降低。比如最少数量的连接器，最少数量的电线。这是我们自上而下推动的。我们实际上有电缆和连接器的预算。例如，从第一性原理的角度来看，你必须有一定数量的连接器，但并不是那么多。而每多一个连接器，那都是一个例外，真的需要辩护和批准，因为它不应该存在，我们应该有计划如何摆脱它。对于运动部件也是如此，对吧？你希望机器人中的零件数量等于运动部件的数量。如果两个零件之间没有相对运动，那为什么有两个零件？应该只有一个零件。你不可能总是做到这一点，但你应该绝对知道第一性原理的最小复杂度是什么，你离它有多远，以及你如何朝着它努力，因为最终目标应该是达到那个最小复杂度。

Oh yeah. And like every day, every time I look at a robot, all I see is everything that's wrong. So I'm not that lucky in that sense. Like everyone else sees something as like this is beautiful. I'm like no no no no like this is wrong, this is wrong, this is wrong. And I think that's probably going to be like that forever. So I'm not the right person to ask that. But I do think I find beauty in every time we can do simplifications. And I think the biggest simplification that's happened over the last year is, for example, wire harnessing and just ensuring like the complexity of the cabling really going down tremendously. Like minimum number of connectors, minimum number of wires. And this is just something we drive top down. Like we literally have like a cable and connector budget. For example, there's a certain number of connectors that you have to have from a first principles point of view, but it's not that many. And every connector that you have more than that, that's an exception that really needs to be defended and needs to be approved because it shouldn't be there, and we should have a plan for how to get rid of it. And this is the same for moving parts, right? Like you want the number of parts in your robot to be the number of moving parts. If two parts don't move with respect to each other, like why are there two parts? Like there should be one part. You can't always achieve this, but you should absolutely know what is the first principle's minimum complexity and how far away are you and how are you driving it down towards that, because the ultimate goal should be to reach that.

有没有一个具体的例子，比如非常明显的，这个应该只是一个零件，你能想到的？

Has there been any specific example where there was like a very obvious this should just be one part that you can think of?

这非常小众，但我最喜欢的是电机上的旋转部件和轴现在比以往少了一个零件。因为一些非常聪明的人想出了如何在不需两个零件的情况下完成轴的轴承组装，通常总是两个零件，因为你有点需要把它夹在中间，但实际上是一个零件。对我来说，这非常美。零件数量减半了。

This is very niche, but my favorite is actually how the rotational part and shaft on the motor is now one part less than what it's always been. Because some very smart people figured out how to do the bearing assembly of the shaft without needing two parts, which usually it's always two parts because you kind of like need to sandwich it in, but it's actually one part. To me, that's very beautiful. It's half the number of parts for that.

如果你把五年前的 Bernt 传送到今天，或者把你传送回去，你会做出任何与现在截然不同的决定吗？

If you took Bernt from five years ago and you teleported him to today, or teleported yourself back, would you make any decisions drastically differently than you have?

我会让团队保持更小、更精简更长时间。我认为你在开发过程中会一次又一次地学到的一个教训是：美丽的产品是简单的产品，而美丽的简单产品是由非常小、极具才华的团队制造的。你拥有的人越多，产品中就会引入越多的复杂性。这种类型的工程并不随人数而扩展。你可以通过让很多人从事制造流程、测试、自动化等来扩展制造。但核心产品设计，一旦有多个人，就需要接口。一旦有了接口，就会引入人为的复杂性。有一个非常实质性的优势，即你能在头脑中理解和掌握多少系统，以便你作为单个人可以在整个系统上做出决策。通常我会说跨子系统，但理想情况下没有子系统，只有系统。现在对于如此复杂的产品，这是不可行的。但你要尽可能接近这一点。

I'd keep the team way smaller and leaner way longer. I think a lesson you learn again and again and again as you do development is that the beautiful products are the simple products, and the beautiful simple products are made by very small, extremely talented teams. The more people you have, the more complexity you will get in your product. And this type of engineering just does not scale with the number of people. Like you can scale your manufacturing with having lots of people doing manufacturing processes, testing, automation, all these things. But like the core product design, once you have multiple people, you need interfaces. And once you get interfaces, you kind of introduce artificial complexity. There's a real very material advantage to how much of the system can you have in your head and understand so that you as a single person can make the decisions across the full system. Usually I would hear say like across the subsystem, but ideally there are no subsystems. It's just the system. Right now this is not doable for this complex of a product. But you want to get as close to that as possible.

当你试图将这种由上千个零件组成的产品整个概念记在脑中时，你如何尽可能快地跟上进度，并在任何给定时间在脑中拥有对这东西最准确的表征？

When you are trying to keep the entire idea of like this thousand-part product in your head, how do you kind of come up to speed as fast as possible with and have the most like accurate representation of what this thing is in your head at any given time?

我认为这归结为你是否理解了机器如何运作以及为什么运作的基本原理。如果你真正理解了基本原理，那么如果你有一个好的设计，其他一切就都是合乎逻辑的。那么复杂性在逻辑上就消失了。然后你最后看着它，你会说，实际上，我们怎么花了这么多时间？这并没有那么复杂。

I think it just comes down to have you understood the fundamental principles of how the machine operates and why. And if you really understand the fundamentals, then everything else is just logical if you have a good design. So then the complexity kind of goes logically goes away. And then you look at it in the end and you're like, actually, how do we spend this much time? Like, this isn't that complicated.

这确实应该是目标，对吧？也许其他人看了会觉得，天哪，这太复杂了。但一旦你真正了解它如何运作，魔法就消失了，对吧？它只是一台相当简单的机器。我认为这需要很多时间。所以你也需要创造一个环境，让人们留下来，获得那种直觉和对系统的深刻理解。但在工具方面你也可以做很多，这相当重要——比如你如何在机电方面构建设计工具，让你能够以可验证正确的方式迭代产品。而且这变得越来越高效，因为你有这么好的工具，尤其是在编码方面。但这一直是个挑战：找到在机电方面非常深入，但仍然足够广泛，能够开发自己的工具来解决问题的人，因为大多数在机电方面非常深入的人不一定是软件人。但迄今为止最有价值的人是那些在所有领域都非常广泛的人——最好甚至深入到材料科学和其他一切，对吧？比如，你真正掌握了多少领域，你能开始看到它们之间的关联吗？这样边界就消失了。最优秀的人通常会构建很多工具来完成你的工作。

And that should really be the goal, right? Maybe other people look at it and are like, oh my god, this is complex. But once you really know how it works, the magic goes away, right? It's just this pretty simple machine. I think that takes a lot of time. So you also need to create an environment where people stick around and get that intuition and deep understanding of the system. But you can also do a lot on the tooling side, which is quite important—like how do you build the design tools on the electromechanical side that allow you to iterate on the product in a way that's verifiably correct. And this is increasingly becoming more productive because you have such good tools, especially on the coding side. But that's always been a challenge: finding people who are very deep on electromechanical but are still general enough that they can develop their own tools to solve the problem, because most people that are very deep on the electromechanical side aren't necessarily software people. But by far the most valuable people are the people who are very broad across all of these domains—preferably even down to material science and everything else, right? Like, how many domains have you actually mastered and can you start to see the correlation between them? So the boundaries disappear. The best people generally end up building a lot of tools for how to do your job.

如果长期来看，数百万人家中都会有这些东西，理想情况下，这是人类会与之建立联系并开始爱上它的东西。我看到的一个视频里，你在拥抱 Neo。我当时想，你如何经历这个过程，试图设计出某种从根本上人类可以学会去爱或决定去爱的东西？

If long-term millions of people are going to have these things in their homes, ideally, this is something that humans form a connection with and they start to love. One of the videos that I saw, you were hugging Neo. And I was thinking to myself, how do you go through the process of trying to design something that is fundamentally something that humans can learn to love or decide to love?

是的。所以有一个显而易见的答案，比如，你在温暖和材料方面做正确的事情，设计吓人的东西很容易，然后它们看起来像科幻和现代，因为它们吓人。这有点像简单的作弊码。我们不那样做。我们尝试用温暖来设计东西，就像技术融入社会，你甚至不应该把它当作技术——那才是成功。所以这有点像硬件方面。这很难做到，但相当直接。我认为在智能方面，这非常有趣。我对世界模型和我们正在做的事情非常看好的一点是，机器人实际上也通过所有这些数据学习社交行为。一个很好的例子是，如果你让机器人——那边有一些袋子——所以你让机器人过去给你一个装满赠品的礼包。它递给你，但它不松手，对吧？然后你抓住把手，然后机器人才松手。训练数据中没有这些数据，对吧？机器人没有特定任务的数据。只是这些社交行为和如何与人互动，就从模型中涌现出来。肢体语言从模型中涌现出来。就像一切都坍缩成一个大的全能模型，理解从如何完成任务到如何社交的一切，如何模拟其他智能体，像其他人一样。如果你想高效地移动和做事，那么你需要理解和预测其他人会做什么，对吧？我们一直这样做，我认为智能的那部分非常非常有趣，但也非常非常难以做好。我认为最终这很重要，对吧？因为那是你建立信任的方式，那是你创造温暖的方式。而且我认为它也会相当个性化——人们不一定想要完全相同的东西。但我确实认为这些机器会学习并持续很长时间，它们会成为家庭的一部分和社会的一部分。看到它如何演变会非常有趣。我认为拥有这样一个终身伴侣会非常美好，它总是站在你这边，记住发生的一切，永远在你身边。我想我们都需要这个。

Yeah. So there's the obvious answer, which is like, oh, you do the right things with respect to warmth and materials, and it's easy to design things that are scary and then they look kind of science fiction and modern because they're scary. Like that's kind of the simple cheat code. We don't do that. We try to design things with warmth, like technology that just blends into society and you shouldn't even think about it as technology—like that will be the success. So that's kind of on the hardware side. And that's hard to do but it's pretty straightforward. I think on the intelligence side it's incredibly interesting. And one of the things I'm so bullish on with the world models and what we're doing there is the robot actually learns social behaviors also through all this data. And a good example is just like if you get the robot to—I mean there's some bags over there—and so you ask the robot to go over and give you a goodie bag of swag. It hands it to you and it doesn't let go, right? And then you grab the handle and then the robot lets go. And there's no data in the training data, right? The robot doesn't have data of the specific tasks. It's just these social behaviors and how to interact with people just falls out of the model. Body language falls out of the model. Like everything collapses into one big omni model that understands everything from how to do the task but also how to be social, how to simulate other agents like other people. Like if you want to move around and do things in an efficient manner, then you need to understand and predict what will other people do, right? And we do this all the time and I think that part of the intelligence is very, very interesting but also very, very hard to get right. I think ultimately that will matter a lot, right? Because that's how you create the trust, that's how you create that warmth. And I think it's also going to be quite individual—like people don't like—it is personalizable—like people don't necessarily want the exact same thing. But I do think that these machines will learn and last for a long time and they will become a part of the family and a part of society. And it's going to be very interesting to see how that evolves. I think it's going to be quite beautiful to have this companion throughout your life that is always on your side, remembers everything that's happened and is always there for you. And I think we could all need that.

我认为很少有产品会被大多数人描述为生活中令人愉悦的东西，我认为这将是其中之一，当人们想到与人形机器人和 Neos 互动的体验时，会感觉愉悦。就像事情自然发生。你并没有真正告诉它做什么，它就像魔法般地完成了你没想到的任务，你会有这种小小的惊喜和快乐。你如何思考设计一个只是令人愉悦的产品？

I think there's very few products that most people would describe as delightful to have in their lives and I think this is going to be one of them where when people think of the experience of interacting with humanoid robots and Neos, it's going to feel like delight. Like things just are happening. You didn't really tell it to do something, just like magically did some task that you weren't expecting and you have this little surprise and joy. How do you think about designing something that is just going to be like a delightful product?

我认为最终一切都在数据中，对吧？所以它变成了一个数据问题：你如何通过调整数据来创造期望的行为。我认为那里仍有很多未解决的研究。有很多工作要做。我没有解决方案。像其他人一样，我们正在努力，但我认为这将真正极大地提高使用模型的愉悦性。最近有很多讨论，对吧？你看看现在最好的 LLM，它们在个性方面变得有点糟糕，因为你必须非常小心，确保它们不做错事。我们显然会在这里遇到同样的问题。我们想确保这些机器人是人类的看护者，它们帮助我们。找到平衡点，我认为会非常困难。所以，是的，这是我们将来要大量研究的事情，我确实希望通过对机器人暴露我们希望看到在其他人身上的正确行为，这是涌现行为。

I think ultimately it's all in the data, right? So it becomes a data problem of how do you create the desired behaviors through how you tune your data. I think there's still a lot of unsolved research there. There's a lot of work to do. I don't have the solution. Like everyone else, we're working on it, but I think it's something that will really, really, really greatly improve the enjoyability of using the models. And it's been a lot of discussion lately, right? And you look at the best LLMs now, they're kind of becoming worse personality-wise because you have to be so careful to make sure they don't do anything wrong. And we're clearly going to have the same problem here. Like we want to make sure these robots are caretakers of humanity and they help us. And finding the balance there, I think, is going to be very hard. So, yeah, that's something we're going to work a lot on and I do hope that through exposing the robot to the right kind of behaviors that we would like to see in other people, this is emergent behaviors.

我刚想到，如果你在家里有一个机器人，并不是所有家庭都真的幸福。可能会有学到的行为不好。但是，你知道，记住大多数 LLM 主要是在，我不知道，Discord 论坛、Reddit 上训练的，对吧？比那更糟糕的论坛。就像到处都是。到处都是，对吧？所以就像人性的黑暗面。

It actually just occurred to me, if you have a robot in a home environment, not all homes are actually happy. And there might be learned behaviors that are not good. But, you know, remember like most of the LLMs are mainly trained on, I don't know, Discord forums, Reddit, right? Like worse forums than that. Like it's all over the place. All over the place, right? So like it's like the dark side of humanity.

然后你把它拿来，你知道有一个很棒的表情包，对吧？你有一个大肚子怪物，然后有人贴上一个笑脸说现在它准备好了。现代模型就是这样，对吧？你有这个野兽，然后你就像不不不，你要表现好。

And then you take that and you know there's this great meme, right? Where you have the big gut monster and then someone slaps a smiley on it and says now it's ready. And that's kind of what a modern model is, right? You have this beast and then you're just like no no no you're going to behave well.

我认为长期来看，随着数据越来越多，有更好的方法来做这件事。现在每个人都极度受限于数据。我们也是，对吧？但当你有了这个不可思议的工具来学习如何在我们这个世界中运作时，也许你可以在数据上更有选择性。我不知道，但也许就鼓励大家对我们的机器人好一点，你的机器人也会对你好。

And I think there are better ways to do that long term as you get more data. Right now everyone is extremely data constrained. So are we, right? But as you get this incredible tool to learn about how to operate in our world, maybe you can be more selective on your data. I don't know, but maybe just encourage everyone to be kind to our robots and your robots will be kind to you.

但这里有一个平衡，因为愿意打破规则和探索是我们学习的重要部分，对吧？这是你获得新想法和创造力的方式。也许这就是今天在模型里看不到那么多创造力的原因之一。你并没有真正鼓励和暴露那种行为。

But there is a balance here because the willingness to break rules and explore is a large part of how we learn, right? It's how you get new ideas and creativity. And maybe that's one of the reasons you don't see that much creativity in models today. You don't really encourage and expose that kind of behavior.

你怎么暴露更多这种行为？比如如果你试图驱动机器人做有创意的事情，你怎么去播种这个？

How do you expose more of it? Like if you were to try to drive the robots to do creative things, how do you seed that?

我认为玩耍非常重要，我梦想一个未来，机器人好奇又爱玩，如果它们没有别的事情做，就会去弄清楚某个东西是怎么运作的。我不知道，去沙子里蹭蹭脚，弄清楚沙子的动力学是怎么回事。我不知道。当然这纯粹是猜测，但我认为验证假设的能力——因为你确实有真实世界在这里作为基础——希望能给你更多探索的能力，因为你可以真正弄清楚你的假设是否产生了你预期的结果。

I think play is just incredibly important and I dream of a future where robots are curious and playful and if they don't have anything else to do, they'll go and figure out how something works. I don't know, go shuffle your feet in the sand and figure out how the dynamics of sand works. I don't know. Of course it's pure speculation, but I think the ability to verify your hypothesis because you literally have the real world here grounding things, hopefully will give you some more ability to explore because you can actually figure out whether or not your hypothesis results in what you thought it would.

我们开车过来的路上，我在看这个小孩，就是随便——我们在红绿灯前停下，这个小孩就在到处摸，触摸这个世界。我想那好像是某种扶手，他们在摸它，就那样到处感受，然后我突然想到，你想让你的机器人在家里做的所有任务。也许你想让它洗衣服、给你做咖啡，所有这些事。但很可能大部分时间，如果你去上班 8 小时，这东西就只是闲置在那里。但你可能可以让它通过采取行动来收集数据，就像字面意义上一个孩子体验世界，你知道，体验这个家，做事情。

When we were on the drive over here I was watching this kid just randomly—we were stopped at a stoplight and this kid was just feeling around and touching the world. I think it was like some handrail and they were touching it and they just feeling it around and then it just occurred to me there's all the tasks that you want your robot to do in your home. Maybe you want it to do the laundry, make you coffee, all this stuff. But then probably most of the time if you're at work for 8 hours, this thing is just there sitting idle. But you could potentially have it collecting data by taking actions and just literally like a child experiencing the world, you know, experiencing the home, doing things.

我认为还有一点我们还没谈到，你知道有 Neo 平台这个概念，我们把它开放给开发者和所有人来在上面构建，我认为这将极其重要，因为这是一个如此大的问题。我不认为我们能独自解决它。我们真的需要让尽可能多的人来做开发并参与这段旅程。这些机器人会接触到非常不同的环境和非常不同的数据。我认为这也很重要，因为你想要尽可能宽的分布。

And I think also we haven't talked about this yet, but you know there's the concept of the Neo platform where we open this up to developers and everyone to build on it which I think is going to be incredibly important because it's such a big problem. And I don't think we can solve it alone. We really need to enable as many people as possible to do the development and partake in the journey. And these robots will get exposed to very different environments and very different data. And I think that's also very important because you want as wide a distribution as possible.

如果你想想家庭，我认为你部分是对的，但我认为实际上随着这些系统变得普遍非常有能力，我们的标准只会提高，对吧？所以就像一切都会完美地满。你所有的衣服都会被熨平，甚至你的床单都会被熨平，它们会完美地铺在你床上，所有的玻璃杯都会整齐地摆在柜子里，到处都没有灰尘，如果你想的话你可以创造很多工作，就像如果劳动力没有成本，你可以想出很多边际上有用的事情，但反正我会做，因为为什么不呢。所以我不确定会有那么多闲置时间。我们拭目以待。我们拭目以待。

If you think about the home, I think you're partially right, but I think also actually as these systems become generally very capable, our standard will just go up, right? So like everything is going to be like perfectly full. All of your clothes will be ironed, even your sheets will be ironed and they will be like perfect on your bed and like all the glasses will be online in the cabinet and like there will be no dust anywhere and like you can create a lot of work if you want to, like if there's no cost to the labor you can come up with a lot of things that are like marginally useful but I'm going to do it anyway because like why not. So I'm not sure there will be that much idle time. We'll see. We'll see.

你知道，当你谈到完美地摆放玻璃杯时，我在想，像大多数人一样，我的衣服并不是完美折叠的。所以，如果你模仿我，它是折叠了，但它并不——你能看到褶皱，你知道吗？如果大多数人的环境是不完美的，而人类在现实世界中实际运作的方式就是不断地不完美，你怎么收集最好的现实世界完美数据？

You know, when you were talking about perfectly arranging the glasses, I was thinking like most humans, like my laundry is not perfectly folded. So, if you emulated me, it's folded, but it doesn't—you see creases, you know? If most people's environments are imperfect and the way that humans actually operate in the real world is just constantly imperfections, how do you collect like the best real world perfection data?

我不认为你会。我认为你只是学习世界是如何运作的，我们学到了如果你像你那样做，你的 T 恤上会有褶皱，有时你会遇到随机的例子，它实际上是直的，两者同样重要，因为你如何学习从一个阶段过渡到另一个阶段，它更多变成你如何确保智能在这之上涌现，这样如果我说实际上我想让我的衣服完美地平整，那么机器人就懂得如何做到。这是一条相当好的路径。这是一个探索得相当充分的问题。尤其是自动驾驶就是一个很好的例子。就像大多数司机都是糟糕的司机。

I don't think you do. I think you just learn how the world works and we learned that there will be wrinkles in your t-shirt if you do like you do and sometimes you will have the random example where it's actually straight and both of them are equally important because how you learn how to get from one transition from one stage to another and it more becomes like how do you ensure that you have the intelligence emerge over this so that if I say actually I want my laundry to be like perfectly straight then the robot understands how to do that. It's a pretty good path there. It's a pretty well explored problem. Especially self-driving is a good example of this. Like most drivers are terrible drivers.

对于驾驶，这是一个非常——你希望机器人每次都做到精确、完美的工作。

For driving, it's a very—you want the robot to do exactly, you know, perfect job every time.

是的。但如果你只是克隆人类在交通中的行为，你不会得到那个。

Yeah. But like if you just clone human behavior in traffic, you're not going to get that.

不，完全不会。

No, not at all.

所以你不能真的只是那样做，对吧？你从所有关于人们如何驾驶的数据中学习。这是非常有用的数据。你学到如果你做了某事会发生什么。你不是学习如何驾驶，你学习驾驶的后果，对吧？然后当你理解了如何操作一辆车、交通如何运作、其他车通常如何移动、行人可能如何移动，并预测所有这些事情，现在你可以学习以一种理想的方式运作，前提是有人告诉你什么是理想的行为。但我不认为模型必然应该通过克隆人类行为来明确地学习。那是我们多年前的做法，但那不是你现在做这些事情的方式。你学习世界是如何运作的，然后你基本上可以在那个空间中搜索你想要的任何最优解。

So you can't actually just do that, right? You learn from all the data about how people drive. It's very useful data. You learn what happens if you do something. You don't learn how to drive, you learn the consequence of driving, right? Then when you have that understanding of how to operate a car and how traffic works and how typically other cars move and how pedestrians might move and predicting all these things, now you can learn how to operate in a manner that is desirable given that someone tells you what's the desired behavior. But I don't think models necessarily should explicitly learn through cloning human behavior. That's how we used to do things like years ago but that's not how you do these things anymore. You learn how the world works and then you can essentially search in that space for whatever optimal solution that you would want to your problem.

如果你要往前想几年，你们并不是在为你们的机器人开发所有的模型之类的，而有某种几乎像应用商店的东西，人们可以创造东西，创造机器人可以做的新的动作或任务，然后有人就可以把它下载到他们的 Neo 上，突然一个新技能就被解锁了。你打算怎么做这个？

If you had to think forward a few years and you guys are not developing all the models or whatever for your robots and there's kind of like this almost app store where people can create things and create like new actions or tasks that a robot can do and then someone can just like download it onto their Neo and suddenly a new skill is unlocked. How do you like do that?

我认为这基本上会和今天你在大型模型中看到的一模一样。所以会有少数几个主导模型，只是因为它们有最好的数据和评估。AI 里有这样一个不算保守得很好的秘密，那就是 AI 里没有秘密。一切都只是数据和评估。为什么最好的模型是最好的模型，是因为这些公司有最好的数据。它们起步早，有庞大的用户群。它们获得了大量的数据，而且现在仍然获得大量的数据。

I think it's going to be essentially exactly the same as you see in large models today. So there will be a few dominant models just because they have the best data and evals. There's like this not so well-kept secret in AI that there are no secrets in AI. It's all just data and evals. Like why are the best models the best models is because these companies have the best data. They were early, had huge user bases. They got an enormous amount and still get like an enormous amount of data.

因为他们有那么多用户数据，就能训练出更好的模型。而有了更好的模型，又能获得更好的用户数据，因为用户更多了。这有点像自我强化的飞轮。还有一些模型可以让你做专门化。你可以下载一个开源模型，尝试对它做一些微调，让它在你的特定任务上表现更好。但只有当任务非常困难、非常小众时，这样做才比直接用他们的通用模型更合理。我不认为这次会有什么不同。我们看到同样的趋势、同样的缩放定律、同样的模式。只是机器人技术到现在为止还像是个玩具问题。大家一直在处理极小的数据集，而且主要是在机器人数据上，比如“我采集了 4,000 小时这些特定行为的数据”，然后觉得这是个很大的数据集。不，4,000 小时根本不算什么。哪怕几十万小时也不算什么。

And since they have so much user data, they are able to train better models. And when you have better models, you get better user data because you get more users. It's kind of a self-enforcing flywheel. There are other models where you can specialize. You can download an open source model. You could try to do some fine-tuning on it to get it better on your specific task. But it has to be a very hard, very niche task before this makes more sense than just using their general model. And I don't think this is going to be different. We see the same trends, the same scaling laws, the same patterns. It's just robotics has until now been kind of a toy problem. Everyone's been working on extremely small sets of data and they've been working mostly on robotics data where it's like, oh, I gathered 4,000 hours of these specific behaviors with my robot, and this is a very big data set. No, 4,000 hours is nothing. Even a few hundred thousand hours is nothing.

如果你看看现在人们训练的模型，比如给人配备传感器，对吧？然后让这件事规模化，接着他们说，哦，我们用 20 万小时的数据训练了这个基础模型，这在机器人领域以前算是巨大的。但实际上它仍然很小。通用智能大约要在数亿小时的数据上才会涌现。所以你不可能去收集那些数据。这就是为什么我深信，你需要用这些机器作为人类数据和机器之间的桥梁。我确实希望，通过这样做，我们训练的模型能操作很多机器人，而不仅仅是 Neo。

And if you see the models that people train now, like equipping people with sensors, right? And then having that happen at scale, and then they're saying like, oh, we train this foundation model at 200,000 hours of data, which is huge compared to what used to be in robotics. It's actually still tiny. General intelligence emerges at like hundreds of millions of hours of data. So you're not going to go gather that data. That's why I deeply believe that you need to use these machines as the bridge between human data and machines. And I do hope that through doing this, the models that we train can operate a lot of robots, not just Neo.

你通常可以朝那个方向走，对吧？你可以拿一个大型、复杂、能力很强的模型，它是在极其多样化的数据上训练的，然后把它蒸馏到更小的问题上，那个模型就会在更小的问题上表现得很好。你不能反过来。你不能拿一个只解决那个数据的小问题模型，把它蒸馏上去解决通用问题。那行不通。这是单向的。所以你必须拥有最通用的系统，拥有最多样化的数据，经历过最多的事情，然后你才能把那种智能蒸馏到更具体的应用上。

And you can generally go in that direction, right? You can take a large, complicated, very capable model that's trained on this enormous diversity of data, and you can distill it down on a smaller problem, and then that model will be very good at the smaller problem. You can't go the other way around. You can't take this model from this smaller problem that only solves that data and distill it up to solve a general problem. That doesn't work. It's a one-way thing. So you kind of have to have the most general system that has the most diverse data and has experienced the most things, and then you can take that intelligence and distill it down to more specific applications.

所以问题是什么时候会发生，对吧？因为我们还没到那一步。与此同时，会有很多非常好的应用适合更专门的模型，这就是我们今天看到的，对吧？如果你想叠衣服，那么实际上如果你去收集大量用我们的机器人叠衣服的数据，训练一个 VLA，那会比通用模型效果更好，但它不擅长做其他任何事情。它没有对世界如何运作的通用理解，或者说它本质上不智能。它不会表现出社交行为，也不理解如何与人互动。它只会叠衣服。这就是我们今天所处的状态，但它正在快速演变。所以我认为 2027 年会看起来非常不同。也许 2026 年还不会。我认为 2026 年专门的模型仍然会更好。但我认为 2027 年你会看到向更通用智能的转变。

So the question is when that happens, right? Because we're not there yet. And in the meantime there will be very good applications for a lot more specialized models, and that's what we're seeing today, right? If you want to fold laundry, then actually if you go and just gather a lot of data on folding laundry with our robot and train a VLA, that will work better than a general model, but it won't be good at doing anything else. It doesn't have a general understanding of how the world works or like it's essentially intelligent. It won't exhibit social behavior and understanding how to interact with people. It'll just fold the laundry. And that's kind of the regime we are in today, but it's quite quickly evolving. So I think 2027 will look very different. Maybe not 2026. I think in 2026 the specialized models will still be better. But I think in 2027 you're going to see a shift towards more like general intelligence.

所以一开始你必须在某种程度上使用远程操作来弥合边缘情况之类的？

So initially you kind of have to use teleoperation in some form or another to bridge edge cases and stuff like that?

你需要所有数据。答案很简单。你用网络数据，那会占 99% 的数据。你用模拟合成数据。你用给人配备的传感器来获取一些数据。你用第一视角视频数据。你会用带远程操作的机器人数据，也会用机器人边做边学的数据。你全都要用。我认为那种特定的数据，比如机器人失败了，然后我纠正机器人，这种干预数据在规模上并不那么重要。在小规模上极其重要。比如你想用非常少量的数据解决一个特定问题，这种数据是必不可少的。这里的术语是你的分布非常小。所以你知道如何在这个范围内，机器人就知道该做什么。如果它跑到外面，你需要一些东西告诉它如何回到分布内，这样它就知道该做什么。但如果你有一个非常强大且多样化的通用模型，那么实际上很难跑到分布外，然后你就不需要这个了。你总能找到回到你想要的地方的路，因为你仍然在分布内。这与模型的鲁棒性有关。

You need all the data. That's a simple answer. So you use the web data, which is going to be 99% of the data. You use simulation synthetic data. You use sensors that you equip people with to get some of that data. You use egocentric video data. You're going to use robot data with teleoperation, and you're going to use robot data where the robot is just learning by doing. You use all of it. I think the specific data of, oh, I have an intervention in the sense of the robot failed and now I correct the robot, not that important at scale. It's extremely important at small scale. Like if you want to solve a specific problem with a very small amount of data, this data is essential. And the technical term here would be that your distribution is very small. So you know how to be within this kind of realm here, then the robot knows what to do. If it gets outside, you need something that tells it how to get back on distribution. So it knows what to do. But if you have a general model that is very capable and diverse, then actually it's very hard to get out of distribution, and then you don't need this. You can always find your way back to where you want because you're still in distribution. It has to do with the robustness of the model.

首先你在互联网数据上训练，然后你得到一点自己的数据。但最终，为了真正获得训练超级天才机器人所需的数据，你只需要有一大堆这样的机器人在运行。比如你想想，如果明年突然有 5 万台，它们每天运行比如说 16 个小时，你每天就能得到 16 乘以 5 万的新数据。我猜当你考虑规模化的瓶颈时，数据是最大的瓶颈之一。我们就是需要更多机器人在现场。

First you train on the internet data, then you get a little bit of your own data. But eventually in order to really get the data that you need to train a super genius robot, you just have to have a whole bunch of these in operation. Like if you think about if you have 50,000 next year suddenly and they're operating let's say 16 hours a day, you're getting 16 times 50,000 every single day of new data. I assume when you're thinking about your bottlenecks for scaling, your data is one of the biggest bottlenecks. Just we need more robots in the field.

是的。而且它们需要处于非常多样化的环境中，用于非常多样化的应用。实际上，你几乎从来不是受数据量限制，而是受多样性限制。所以当人们说，哦，你需要非常大量的数据。这没错，但你实际上需要的不仅仅是大量数据。你需要极其多样化的不同体验、任务和以前没见过的东西。如果你有足够多的这些，那就会变成大量数据，因为你有那么多不同的东西。但同一件事的大量数据并不能真正帮到你，对吧？

Yeah. And they need to be in very diverse environments in very diverse applications. In reality you're almost never data bound, you're diversity bound. So when people say like, oh, you need very large amounts of data. That's true, but you don't actually need just large amounts of data. You need an extreme diversity of different experiences and tasks and things that you haven't seen before. And if you have enough of that, that becomes a lot of data because you have so many different things. But lots of data of the same thing doesn't really help you, right?

假设你明年出货 5 万台机器人。这些不同的机器人会去哪里？因为也许一开始我想的是 5 万台都去家庭，但也许你不想那样。也许你只想要 1 万台去家庭，还有 2,000 台去森林里建个小屋之类的。我不知道。我觉得你只是

Let's say you ship 50,000 robots next year. What is the breakdown of where those different robots are going to go? Because maybe initially I was thinking like 50,000 to homes, but maybe you don't want that. Maybe you only want like 10,000 to homes and 2,000 that'll just go into the woods and build a cabin or something. I don't know. I think you just

那还得再等几年。我的梦想之一是能够离开家说，嘿 Neo，我这周末要办个派对，我希望我的花园变成日式风格，然后它就会去重新布置整个花园，建一个漂亮的亭子，然后你一周后回来，它就好了。那很好。但那还得几年。

That's like a couple years more out. One of my dreams is to be able to leave my house and say like, hey Neo, I'm having a party this weekend and I would like my garden to be Japanese, and then it'll go and reconfigure the entire garden and build a nice pavilion, and then you come back a week later and it's like, oh, it's ready. That's nice. But that's going to be a couple years out.

所以今年会有几个企业级应用，我们有非常大的客户，这样我们可以把大量机器人部署到特定的垂直领域。然后还有一些会面向整体，平台上会有很多。会是各种类型的良好组合。我认为从所有这些中获取信息并吸取经验教训很重要。但我也认为，有一些应用，其中解决任务真正有用，从客户的投资回报角度来看，非常重要，因为那是推动采用的方式，对吧？你需要规模来降低成本、提高可靠性等等。所以这些的良好组合。我们还没有准备好宣布具体的企业应用和客户是什么，但正在筹备中。

So this year there will be a couple of enterprise applications where we have very large customers, so we can deploy a lot of robots into specific verticals. And then there's going to be some to the whole, and it's going to be a lot to the platform. It's going to be a good mix of all of them. I think it's important to get information from all of them and get the lessons learned from all of them. But I also think it's important to have some applications where solving the task to actually be truly useful, in terms of return on investment for your customer, is very important, because that's how you drive adoption, right? And you need scale to get the cost down and to get the reliability up and everything else. So a good mix of these. We're not quite ready to announce yet what the specific enterprise applications and customers are, but that's in the works.

你是不是也以同样的方式思考，也许你需要大量折叠数据，所以在企业环境中获取数据的一种方式可能是，嘿，Old Navy，你知道，这里有 5,000 台机器人，它们整天都在叠衣服，然后你就能得到，你知道，5,000 乘以每天 16 小时的

Do you kind of think about it also in the same way that maybe you need a bunch of like folding data and so one way to get that in an enterprise setting might be like hey Old Navy you know here's 5,000 robots and they just spend all day folding clothes and then you get you know 5,000 times 16 hours a day of

那会是一个应用。我们不会把折叠作为主要应用，所以我可以揭穿这个迷思。其他人都在做折叠，所以已经有很多折叠数据了。

That would be one application. We're not going to do folding as our main application, so I can debunk that myth. Everyone else is doing folding, so there's already a lot of folding data.

好的，所以其他人都在做折叠。那什么是没有人做但对你来说显而易见的事情？你能说吗

Okay, so everyone else is doing folding. What is the one thing that no one is doing that seems obvious to you? Is that could you say

不？有一天我会再联系你，某一天可能只是几个月后，然后告诉你，比如，这是我所思考的。

No? Someday I'll reach out to you again, some someday being just a few months, and tell you like here's what I thought about.

好的。明年我们就能知道了。

Okay. Next year we can know.

好的。我们认为我们已经弄清楚了。我们认为我们有一些事情要做，会非常重要。

Okay. We think we've figured this out. We think we have a couple things we're going to do that's going to be very material.

我记得有这样一个情况，一家石油公司收购了一家化肥公司，然后突然所有其他石油公司都开始收购化肥公司。就像第一个家伙只是说，我刚买了它，你知道，为了好玩，然后一堆其他人开始无缘无故地这么做。是不是像折叠，一些人开始做折叠，然后其他人就像，他们在做折叠，也许我们也应该做。

I remember this one situation where one like oil company bought a fertilizer company and then suddenly like all these other oil companies started buying fertilizer companies. Like the first guy was just like I just bought it, you know, for shits and gigs and then a bunch of other people started doing it for no reason. And is it like fold some guys started on folding and then everyone else like while they're working on folding maybe we should too.

不，实际上折叠是，我们也做过折叠，你可以去我们的 YouTube 频道，你可以看到我们几年前实际上用 Eve 来折叠衬衫。所以我想我们在折叠方面是早期的。它一直是一个非常困难的问题，因为它是软商品，对吧？就像你永远不知道它会是什么样子。你把它揉皱，它看起来完全不同。就像一个杯子，同一个杯子每次看起来都一样。它可能有点脏，但形状相同。衣服每次你看它都是不同的形状。所以模拟起来非常困难，因为它柔软且可变形。所以这只是机器人学中的一个大型开放问题。你怎么做折叠？在经典意义上极其困难。但恰好是那些用 AI 很容易做的事情之一。你需要很少的数据，它就能工作。我不知道，也许有人正在写一篇非常好的学术论文，可以向你解释为什么会这样。我实际上不知道。但它是做起来最简单的事情之一，但它曾经非常复杂，现在非常简单。所以，它就像最被过度使用的演示，因为它看起来你非常有能力，因为你可以折叠这些衣服，但实际上并不那么难。实际上相当简单。它是最简单的问题之一。

No, it's actually folding is kind of we also did folding like you can go on our YouTube channel like you can see us using Eve actually a few years ago to like fold shirts. So I guess we were early in folding. It's been this incredibly hard problem because it's soft goods, right? Like so it's you never know how it's going to look. Like you crumple it, it looks completely different. Like a cup the same cup looks as like the same cup every time. It might be slightly dirty, but it's the same shape. Clothing is a different shape every time you look at it. So it's incredibly hard to simulate because it's like soft and deformable. So it was just this large open problem in robotics. How do you do folding? Extremely hard to do in a classical sense. And just happens to be one of those things that are very easy to do with AI. You need very little data and it just works. And I don't know, maybe someone's working on like a very nice academic paper that can explain to you exactly why that is the case. I don't actually know. But it is one of the simplest things to do, but it's something that used to be very complex and now is very simple. So, it's like the most overused demo because it looks like you're extremely capable because you can fold these clothes, but it's actually not that hard. It's actually quite simple. It's one of the simplest problems to solve.

我喜欢这句话。创办一家初创公司就像嚼玻璃并凝视深渊。而成为创始人就是成为享受这个过程的人。你上一次经历彻底的失望是什么时候？

I love this line. Starting a startup is like chewing glass and staring into the abyss. And being a founder is being someone that enjoys that process. When was the last time that you experienced utter disappointment?

我的意思是，这每周都会发生。我的意思是，总有不顺利的事情。所以我学到的是你必须享受旅程。就像那是秘密。因为是没有尽头的，对吧？就像不是，哦，如果我只要修复这些我知道是错误的事情，那么一切都会好起来。不，不，不。然后你进入下一个级别，更多的东西是错的。如果你修复了那个，你进入下一个级别，甚至更多是错的，每次都会更糟，对吧？因为事情变得更大更复杂，所以你永远不会到达终点。所以如果你不觉得那有趣，那么你就不会持久。你必须真正享受旅程。我一直是那种宁愿有高潮然后忍受低谷，而不是只是处于中间的人。所以我感到非常幸运，我每周都能有高潮和低谷。它从不无聊。至少我可以说它从不无聊。有时我会承认我醒来时想，我为什么不直接设计一个应用？我为什么决定建造一个机器人？有更容易赚钱的方式。但

I mean, this happens every week. I mean, there's always things that don't go well. So what I've learned is you have to enjoy the journey. Like that's the secret. Like there's because there is no end, right? Like there's it's not like, oh, if I just fix these things that I know is wrong, then everything's going to be good. No, no, no. Then you get to play on the next level where more stuff is wrong. And if you fix that, you get to the next level where even more is wrong and it's just worse every time, right? Because things become bigger and more complex and like so so you're never going to get to the end of it. So if you don't think that's fun, then you're not going to last. Like you you have to really enjoy the journey. And I've always just been a fan of like I'd rather have the highs and then endure the lows than kind of like just be in the middle. So I feel very fortunate that I get to have the highs and the lows every week. It's never boring. At least I can say it's never been boring. Sometimes I'll admit I woken up and thought, why the heck didn't I just design an app? Like why did I decide to build a robot? Like there's easier ways to make money. But

但你并不是真的为了钱。

But you weren't really in it for the money.

嗯，不。我不是。那也不会那么有趣。

Well, no. I wasn't. And that wouldn't have been as fun.