Brett Adcock 谈 Helix 2.5、机器人技术的圣杯与 Figure 4

感谢你加入我们。这对你来说是个重大时刻。你把它称为机器人技术的圣杯——能够泛化，在没见过的地方干活。不过我刚才问的问题是：现在当人们说通用 AI 时，指的是任何任务、任何时间，一次训练之后就能做。而据我理解，你说的其实是在没见过的地方做多种任务。你能解释一下你说的泛化圣杯是什么意思吗？

Thank you for joining us. This is a huge moment for you. You called it the holy grail for robotics — being able to generalize, doing work in unseen places. The question I was just asking though is: right now, when they say generalized AI, that means any task, any time, after one training. What you're saying, though, as I understand it, is doing work in unseen places but multiple tasks. Can you explain what you're saying about the holy grail for generalization?

我想说的是，我们想打造的产品，是让机器人出现在你家里，仅凭语音，也许再加上一些记忆条件，就能像一款出色的产品那样干活。只要叫它洗衣服、洗碗，机器人就能端到端地自主、安全地执行这些任务。要达到那一步需要几样东西。第一，你需要一个作为通用机器的人形机器人。轮子之类的东西行不通。第二，你需要在 AI 这一侧解决全身自主的问题。除了从底层到顶层全用神经网络和深度学习，没有别的办法能做成这件事。第三，你需要能够泛化到没见过的地方，因为你要进入的每一栋房子，里面的电器、地板布局和家具都不一样。对机器人来说，一切都几乎是全新的、分布外的。所以为了打造尽可能最高质量的产品，我们必须能够进入从未见过的地方，做真正有用的工作。它不能只是站在那儿。我们要能洗衣服、洗碗、整理家务、卸下杂货、做饭、遛狗、倒垃圾。

I mean, the product that we want to build is to have a robot show up at your house and be able to, just through speech and maybe some memory conditioning, be able to just do work like a great product. Just ask it to do laundry, dishes, and end to end the robot can execute these autonomously, safely. Getting there is going to require a few things. One is you'll need a humanoid robot as a general purpose machine. Wheels, other things won't work well. Second, you'll need to figure out how to solve for whole body autonomy on the AI side. There's no other way to make this work besides neural nets and deep learning all the way down the stack. And the third is you need to be able to generalize to unseen places, because every item in this house you're going to be in will have different appliances and floor layouts and furniture. Everything will be almost like new and out of distribution for the robot. So we need to be able to, in order to create the best quality product possible, go into places we've never seen before and do real useful work. It can't just stand around. We need to be able to do laundry, dishes, tidy the home, unload groceries, cook food, walk the dog, take out your trash.

所以我们一直在非常努力地构建一整套垂直整合的深度学习栈，同时非常激进地推进人形硬件，朝着我们认为的理想产品前进。现在有了 Helix 2.5，我们能够把机器人带到这里的场外，放到租来的房子里，展示我们开始具备泛化能力。现在还是早期阶段，所以并不完美。也许我来讲讲我的论点——Scott 上周来过办公室，我跟他聊了一些——我认为人形机器人基本上有四章。

So we've been working really hard on building an entire vertically integrated deep learning stack, working on the humanoid hardware very aggressively towards what we think is the ideal product. And now with Helix 2.5, we're able to take robots offsite here to home rentals and show that we're able to start generalizing. It's like the early days here, so it's not perfect. Maybe I'll give you my thesis — Scott was here last week at the office, I was talking to him a bit about it — but I think there are largely four chapters for humanoid robots.

第一章是能够打造出真正出色的人形硬件。硬件必须能达到你需要的速度和扭矩、续航、灵巧度、活动范围、质量、成本——所有这些都必须配合得非常好。你必须造出真正出色的硬件。在此基础上，你需要弄清楚如何构建正确的 AI 架构和配方，让它基本上能根据某种指令在人形机器人上运行。指令可以是视频指令并加以条件，也可以是记忆，或者最终通过语音。然后我们需要能够做到从像素到扭矩。我们需要能够接收输入——我在周围环境中看到的东西——并输出长时程的自主行为。这个配方可以展示几个小时，也许几天，然后你就能让机器人自主运行，无需人类介入。那是第二章。

Chapter one is being able to build really great humanoid hardware. Hardware has to be able to get the speed and torques you need, runtime, dexterity, range of motion, mass, cost — all of those have to line up really well. You have to build just great hardware. And from there, you need to figure out how to build the right AI architecture and recipe to be able to run on a humanoid basically from some sort of command. It could be a video command and conditioned, or it could be memory, or ultimately through speech. And then we need to be able to go basically pixels to torques. We need to be able to just take inputs — what I'm seeing in my surroundings — and be able to output long horizon autonomy. And that recipe can be shown over a few hours or maybe a few days, and you'd be able to run robots autonomously without a human in the loop. That's chapter two.

所以第一章和第二章，是把人形硬件和 AI 自主性都做到非常好，完全端到端。那本书的那一章——前两章——不是花钱的章节。我的意思是，世界上没有任何一笔钱能让你到达那里。如果你从街上找个人，给他 500 亿美元，我认为他很可能永远都很难做到。在我看来，这类似于轨道火箭。你不能直接花钱雇人把这件事做成。这是一个非常长周期的深度工程任务，你基本上需要一支完全疯狂的工程团队，日夜工作、痴迷于此。而且很容易把机器人做得又重又大。很容易在某个地方做几分钟的自主，但要在很长一段时间里、在非常困难、需要真正有用的地方做到这一点——那真的很难。而这其中很多都涉及移动你的手、搬包裹、搬货物、洗衣服之类的，还有导航。这真的是一个全身的体验。所以我认为这就是人形机器人的前两章。

So chapter one and two is getting the humanoid hardware and the AI autonomy running really well, fully end to end. That chapter of that book — the first two chapters — are not money chapters. There's no amount of money — when I mean not money chapters, there's no amount of money in the world that can get you there. If you found somebody off the street, gave them $50 billion, I think they would likely have a very difficult time ever getting there. It's similar to orbital rockets in my mind. You just can't go off and pay people to get this done. It's just a very long duration deep engineering task that you basically need a completely nuts engineering team that works day and night and is obsessed. And it's easy to make the robot very heavy and very big. It's easy to do autonomy for a couple minutes somewhere, but to be able to do it over some long period of time in very difficult places that need to be useful — that's really hard. And a lot of that involves moving your hands, moving packages or moving goods or laundry or whatever it would be, navigating around. It's really a whole body experience. So I think those are the first two chapters of humanoid robots.

我认为接下来的两章真的不一样了。我认为再下一章是构建人形智能，我觉得这可以画成模型中的损失下降，或者最好的衡量方式是机器人在世界上能成功做更多人类能做的事——意味着机器人能做更多事情，更多人会想要它们。你基本上想提升整个栈中的智能，而在我看来，这真的取决于数据和算力。如果你有一个深度学习架构，你很大程度上受制于深度学习的物理规律，也就是数据、算力和模型规模。幸运的是，这实际上是书中花钱就能买到的一章，只要你有正确的架构。你基本上只要投入数据和算力，就能输出行为——希望是成功的行为——也就是机器人能做什么。

I think the next two chapters really change. I think the next chapter after that is building humanoid intelligence, and I think that could be plotted as losses coming down in the models, or best yet, as the robots being able to do more things successfully in the world that humans can — meaning robots can do more things more people would want them. You basically want to increase intelligence in the stack, and that is really dependent, in my mind, on data and compute. If you have a deep learning architecture, you're largely bound by the physics of deep learning, which are largely data, compute, and model sizes. And thankfully this is actually a chapter of the book that money can buy you, if you have the right architectures. You can basically just put data and compute in and you can output behaviors — hopefully successful behaviors — of what the robot can do.

然后这本书的第四章是制造。你不能把制造放在那三章中的任何一章之前。一旦你有了智能，一旦你有了数据和算力，一旦你在增加数据和算力、提升智能，你基本上就在提升机器人推向市场时的成功特性和行为。意思是机器人现在能在我家里做更多事情。它能更快、以更高成功率做这些事。你基本上越来越接近一个大众市场产品。到那时，你需要几乎与智能以同样的斜率并行地扩大机器人的规模。这也是一个花钱就能解决的问题。钱、钱、钱——你基本上就是等。它非常难工程化，但它是靠资本来扩张的。

And then the fourth chapter of the book is manufacturing. You can't put manufacturing ahead of any of those three chapters. Once you have intelligence, once you have data and compute, once you're increasing data and compute and increasing intelligence, you're increasing fundamentally, I think, the success characteristics and the behaviors of what the robot can do to market. Meaning the robot now can do more things in my home. It can do those things faster and with a higher success rate. You're basically getting more and more of a mass market product. At that point you need to be scaling up robots almost in parallel at the same slope as intelligence. That is also a money-can-buy problem. Money, money, money — you wait, basically. It's very difficult to engineer, but it is something that is scaled with capital.

所以前半部分章节是花钱买不到的。你需要世界上最疯狂的工程团队。而之后的下一个大组，是花钱就能买到的两章。

And so the first half of the chapter is a money-cannot-buy. You need the craziest engineering team in the world. And the next big group after that are two chapters of money-can-buy these things.

我认为在 Figure，让我们兴奋的是，我们感觉自己可能正处在 Scaling（规模扩张）智能的第一章——我们如何基本上把更多数据和算力投入到这个问题上，并展示机器人能够做更多事情。Helix 2.5 是我们第一次公开描述这一点，我们展示了随着更多数据进入系统，机器人在物理上变得更有能力。我们可以观察到这一点。我们每周在大量出租房屋中运行评估。我们有一个 Zoom 链接，你可以在内部观看所有这些评估。我想 Scott 我们给你看过。

And I think what gets us excited here at Figure is we feel like we're in maybe the first chapter of scaling intelligence—how do we basically put more data and compute at this problem and show the robots able to do more things. And Helix 2.5 was our first way to characterize this publicly, where we showed as more data is entering the system, the robot's physically getting more capable. We can watch it. We run evaluations across tons and tons of rental homes per week. We have a Zoom link you can watch internally of all of them. I think Scott we showed you.

是的。我看到了，那绝对很酷。我不知道你是否会让它上线，或者它能不能——那里面好像有人们的地址。

Yeah. I saw that was absolutely cool. I didn't know if you're going to make that go live or not or whether it can—like there's like people's addresses in there.

哦是的，很酷。

Oh yeah, it was cool.

它覆盖整个湾区，你一整天都在看这些。我们在做消融实验和评估，你可以真正看到数据较少的模型表现不佳，数据较多的模型表现更好。我们可以量化这一切，我们可以为一切定价。所以我想也许对 Herbert 总结一下，我认为总的来说我们对这个感到兴奋，因为如果我们处在一个可以为机器人像大语言模型那样扩展的时期，我认为那可能是很长一段时间以来最激动人心的时刻之一。我认为描述整个四章过程很重要，因为在某种程度上，在做这个之前制造大量机器人是非常不健康的，而且不会成功。如果你的机器人不智能，无论你的制造能力如何，你基本上是在制造死尸。

It's across the Bay Area and like all day you're just watching these. We're doing ablations and eval on evals of all this and you can literally see the models that had less data performing not great and the models that have more data performing better. And we can quantify all this and we can price everything out. So I guess maybe to Herbert to summarize, I think that's in a whole we're excited about this because if we're in a period of the world where we can scale similar to large language models for robotics, I think that might be one of the most exciting times in a long time. And I think it's important to characterize that whole four chapter process because in some way, manufacturing a lot of robots before doing this is very unhealthy and it won't work. You're basically making dead bodies no matter your manufacturing capacity if your robots aren't intelligent.

是的。

Yeah.

但如果机器人变得智能，你能制造就很重要。所以我们正在与 Baku 合作。然后在智能方面，我们现在已经启动了 index。我们正在全球超过 100 个国家收集数据。现在，本周我们的应用已经有 107,000 名活跃用户。我们已经上传——我们每秒上传大约 50 分钟的数据。

And but it does matter if the robots get intelligent that you can manufacture. So we're working on that with Baku. And then on the intelligence side, we've now spun up index. We are collecting worldwide across over 100 countries. Now, this week we've already had 107,000 active users on our app. We've uploaded—we're uploading about 50 minutes of data per second.

哇。

Wow.

本周，

This week,

嗯，上周你告诉我是 36，现在你达到了 50。

um last week you told me it was 36 and you're up to 50 now.

上周 36。本周我们达到了 50。

36 last week. We're up to 50 this week.

哇。

Wow.

嗯，一个月前我告诉团队，我说，你知道，嗯，就像把所有安全旋钮都拿掉，把所有旋钮都拿掉，他们之前从预算角度人为地限制了这个。我说把它拿掉，让它尽可能高地运行。所以，我认为我们正在扩展数据，已经为此工作了大半年。这是一个艰难的项目。我们必须亲自出去收集这些数据。嗯，所以那个项目正在加速，随着我们把更多数据放入系统，它变得更好，数据在世界各地变得更加多样化。这实际上是我们网站上的实时视图。嗯，我想这可能是一个旧视频，但你也可以去网站看看。这是一个实时视图，每个点亮的像素都是上传到我们服务器的数据，我们最终会清理并准备好用于训练。嗯，所以我们现在进来的数据，希望在 30 天内我们就能用它训练。嗯，所以这是从数据角度说的一点，第二是从算力角度，我们几周前刚刚宣布与 inscale 达成协议，基本上带来多达 100,000 个 Nvidia Bear Rubins，VR200s，市场上最好的 GPU，用于训练。嗯，我们已经承诺了 35 亿，我们现在有希望将其增长到 60 亿，嗯，是的，这就是协议。我认为我们处在一个非常不可思议的位置，我们有一个非常好的架构，机器人从根本上就是惊人的，我们正在内部疯狂地扩展数据和算力，所以明年我们真的很兴奋看到我们能把它推到多远。嗯，我觉得如果我们能打个响指，今天就能拥有所有正确的数据和算力，我们就能解决通用机器人问题。嗯，但这是一个很大的声明。我们需要，你知道，所以我们正在收集这个，我们正在训练这个，嗯，是的，我总体上很兴奋。Scott 上周在这里。我们聊了更多，但嗯

Uh, I told the guys a month ago, I was like, you know, um, like there was just take all the take all the safety knob like take all the knobs off of like they were like, you know, artificially constraining this from a budget perspective. I was just take take that out of there, let it rip as high as possible. So, we're I I think like we are scaling data beyond like working on it for larger part of a year. It's been a hard project. We have to physically go out and collect this data. Uh so we have that project ramping and we're as we're um as we're like putting more data in the system, it's getting better and the data is getting more diverse um across the world. This is actually a live view from our website. Um I think this might be an old video, but you can go actually go to website too. This is a live view of uh every pixel being lit up is uh data being uploaded into our server that we will ultimately clean uh and uh and get ready for training. Uh so we will like this data coming in now within 30 days we'll be like hopefully training on it. Um so that's that's one from a data perspective and two is from a compute perspective we just announced a couple weeks ago a deal with uh inscale to basically uh bring about up to 100,000 Nvidia Bear Rubins the VR200s like the best GPUs in the market for uh for training. uh and we've committed to 3.5 billion that we have like uh we have line of sight now to hopefully uh growing that to six six billion and um yeah this is the deal like I think um so we I think we're in a really incredible spot we have like a really good architecture robots are just fundamentally amazing and we're like scaling data on compute like crazy internally so the next the next year we're really excited to see how far we can really push it um I I feel I feel like if we could snap our fingers and have all the right data and compute at our fingertips today, we would have general robotics solved. And um but that's a big that's a big that's a big statement. We need um you know, so we're like we're we're collecting this and we're training this and um yeah, I'm just overall excited. Scott was here last week. We were kind of chatting through it a bit bit more, but um

是的。

yeah.

是的。是的。

Yeah. Yeah.

嗯，我有几个问题，可能大家都想知道，你刚才也谈到了。机器人出现在你家。所以我假设当你去这些 Airbnb 时，机器人出来了，你把它放在那里，你需要把它带到某个房间。所以，你引导它到某个特定的房间吗？它什么时候知道开始整理床？什么时候知道开始整理毛巾？它只是注意到有什么不对劲吗？因为当我看视频时，我注意到，有人会过来弄乱床，机器人已经站在那里，然后某个时刻它决定行动，它做了我几乎称之为“cobble reload”的动作，它像拿起手，稍微抽动一下，然后才决定行动。所以触发机制是什么？我假设你现在还没有所有你想要的东西，但你在设想它会更多一些。

Um I've got I think a couple questions that probably everyone wants to know and that you you kind of talked about. The robot shows up at your house. So, I'm assuming when you went out these Airbnbs, the robot came out, you had it there, and you needed to get it into a certain room. So, did you guide it over to a particular room, and when does it know to start to do the bed? When does it know to start to do the towels? Is it just kind of notice that something wrong? Because when I've been watching the video, I notice, well, someone will come up there and mess up the bed and the robot is already standing there and then at one point it just decides and it does what I almost call like the cobble reload where it like takes the hand and twitches a little bit before it goes ahead and and decides to go. So, what is the triggering mechanism and and I assume it's you don't have everything in there that you want right now, but you're envisioning it being a little bit more.

是的，我认为在高层次上，我们在进行非常有结构的实验。我认为对我们重要的是能够有条不紊地了解，随着数据进来，模型发生了什么，机器人的评估发生了什么。我们真的想理解这一点，这样当我们出去用更多数据训练下一代模型时，它能很好地优化以继续扩展。嗯，然后我们想基本上做一定数量的消融实验，跨策略理解随着我们扩展数据，我们是否看到更多成功，我们需要真正构建这个。嗯，我认为让机器人随意行动做任何事情真的很难，你知道我们如何比较基线。嗯，这里我们运行了非常有结构的评估，跨三个主要任务

Yeah, I think like at a high level, we we're out running like really structured experiments. I think what's important for us is being able to be be able to methodically know um as data is coming in uh what's happening to the model and what's happening to the valuation of the robot. We really want to understand that so when we go out and train the next generation of like the next version of the model with more data um it's it's optimized well to keep scaling. Uh and then we wanted to basically do uh a certain amount of ablations across the um the the policies understand as we're scaling data are we seeing more success and we needed to really structure this uh I think let just letting robots loose and doing anything is really difficult to like you know how do we like compare bas uh here we ran like really structured evaluations across like three main tasks

嗯，在家庭中。所以我们非常勤勉，嗯，你知道，确保我们非常准确地记录了这一点。对我们来说，成功就是端到端完成整个任务，比如整理客厅，有不同物体。我们想要清理每一个物体。

uh in the home. So we were pretty diligent about um you know uh making sure that we recorded this really accurately. Success for us was like completing the whole task end to end in case of like tidying up the living room with like uh different objects. We wanted like we wanted to clear every single object.

我们做得相当细致，其实我们在博客上写了很多。我们有一整个附录讲我们是怎么评估的，还有各种定义，团队把这些细节整理得非常好。如果你有时间，我建议你去看看。我记得大概有五页左右的内容。他们真的把一切都写得很好。所以是的，这是我们第一次尝试展示：我们在机器人领域几乎处于一种数据匮乏的状态，而随着数据不断进来，机器人正在变得更好。随着我们投入更多数据、在家里展示更多行为，这只会从这里继续增长。

We were pretty detailed, and we actually wrote a lot about this on our blog. We have an entire appendix about how we evaluated this, and definitions that the team did a really good job detailing out. If you have time, I would go check it out. I think it was like five or so pages of work. They just did a really good job writing everything up. So yeah, this is our first attempt to try to show we're in almost like a data-starved regime for robotics, and as data is coming in, the robots are getting better. This will only increase from here as we're putting more data in and as we're showing more behaviors throughout the home.

今年年初我有个挺疯狂的野心：今年最好的情况是什么？就是能把机器人放进一个家庭里，做长时程的自主作业。这次发布我们还没到那一步。但我们正在努力，而且现在感觉这件事是尽可能可实现的。那就太好了。我们能不能把机器人放进任何我们选定的家庭，并且希望在一些长时程工作上取得成功？所以我们正在攻克这个。同时我们也在做商业交付，手头事情很多。但总的来说，这是一次不错的发布，我们希望回头看的时候会说，这只是 20 章、30 章里的第一章。所以我们只需要把它 Scaling（规模扩张）起来。

I kind of had a wild ambition early in the year of like, what was the best case scenario for this year? Being able to put a robot in a home and doing long-horizon autonomy. We're not there with this release. But we are working hard, and it feels as tractable as it possibly could now to be able to do that. So that'd be great. Can we put a robot into any home we choose and hopefully be successful at some of the long-horizon work? So we're working through that. And also we're shipping commercially, and there's a lot of stuff on our plate. But yeah, overall it's been a good release, and we hopefully want to look back and be like, this is the first chapter of 20 or 30 chapters. So we just got to go scale it up.

好。那你们现在还待在那几个家庭里吗？你们是待几天，然后换一个、再换一个？再换一个？

Sure. So are you still on those homes, and are you staying there like for a couple days and moving to another one and another one? Another one?

是的，那些都是租的房子。我们每周都会换掉它们。所以我们就是在不断消耗一个个家庭，跑评估和消融实验。这一切其实一开始有点偶然。我们开始预训练更大的模型。其中一个实验就是，我们去那些从没去过的随机家庭，看看机器人能不能成功做点什么。这一直很有挑战。过去两年我们做这项工作，就是行不通。然后我们开始把机器人带到这些随机房子里，在圣何塞、湾区、东湾、湾区和半岛。几个月前，机器人开始做一些智能的事情，尽管它从没进过那个房子，也从没见过那些物体。

Yeah, those are rentals. We're out of them every week. So we're just chewing through homes and running evals and ablations. And this kind of all started as a little bit of a fluke. We started pre-training larger models. And one of the experiments was like, let's go out to random homes we've never been to and see if the robot can do anything successful. This has always been challenging. The last two years we were doing this work and this just wasn't working. And we started taking robots out to these random houses in San Jose, the Bay, East Bay, Bay Area, Peninsula. And the robots started, a few months ago, doing intelligent things where it's never been in the house before. It's never seen those objects.

我们就在想，天哪，如果我们把这里的数据翻两倍、四倍、八倍、十倍，会发生什么？这会带来什么？而随着我们往系统里投入更多数据，它就是变得更好。

And we're like, man, what happens when we're like two, four, eight-xing, ten-xing the data here? What is this going to do? And it just got better as we put more data into the system.

这跟 Index 是同一个项目吗？你们其实雇了人——顺便说一句，你们上线了这个 App，你们付给人们很多钱。你们付很多钱让人去做一个任务、把它录下来，他们戴上头显，然后做那个动作。还是说这是同一件事？

Is that a different project than Index, where you guys actually hired — you guys launched this app, you are paying people a lot of money, by the way. You're paying people a lot of money to be able to just do a task, record it, they put on a headset, and then they do that. Or is it the same thing?

是的，所以 Index 是我们的——把 Index 想成我们的数据采集计划，用来培育一个大型的——我们的对应物是什么——好吧，让我退一步讲。对于大语言模型，它们有一个非常好的基础先验，来自它们从互联网上理解到的所有知识，比如互联网上有两到三百万亿 token，经过筛选降到 20、30、40 万亿 token 用于预训练。这给了它一个很好的语义基础，用于下一个词预测。对于机器人，互联网上没有这样的东西。没有给机器人用的 YouTube。这不存在。

Yeah, so Index is our — think of Index as our data collection initiative for growing kind of like a large — what's our equivalent — okay, let me take a step back. With large language models, they have a really good base prior from all the knowledge they've understood from the internet, like two to three hundred trillion tokens on the internet that are debuted down to 20, 30, 40 trillion tokens for pre-training. That gives it this great semantic base for next-token prediction. For robotics, there is no version of that on the internet. There's no YouTube for robotics. This doesn't exist.

而且因为我们有一点这个——有一点巧合，就是人形机器人和人类非常非常相似。这是核心基础之一。我加入的时候，你们就说，房间都是为人类布置的，完全合适，硬件非常适合这件事。但数据更适合。有什么比互联网上所有文本还大？是人类数据。世界上有八十多亿人在做各种事情。我们怎么理解这些物理，并在全球范围内、在人类所做事情的大量分布中去挖掘——各种疯狂的事。他们在割草坪。他们在铲猫砂，他们在卸杂货，他们在做晚饭，他们在换机油，他们在修轮胎，就是一切。而我们怎么——这么多熵——我们怎么开始把这一切收集起来，然后带进来？

And since we have a little bit of this — a little bit of a glitch where the humanoid is very, very similar to the human. This is one of the core fundamentals. When I was joining, you guys were like, the rooms are all set up for humans, it's perfectly right, the hardware is really well suited for this. But the data suited even better. What's larger than all the text on the internet? It's human data. Eight-plus billion humans out in the world doing everything. How do we understand this physics and mine this globally across a lot of the distribution of what humans are doing — crazy stuff. They're mowing the lawn. They're scooping kitty litter, they're unloading groceries, they're making dinner, they're doing an oil change, they're fixing a tire, just everything. And how do we — all this entropy — how do we start collecting all of it and then bringing it in?

所以 Index 是我们互联网规模的预训练计划，几乎就是：我们怎么出去把世界上正在发生的所有对物理的理解收集起来，带进来，清洗它，在上面训练，然后把它迁移到机器人里。这一直是终极目标。那是北极星。我认为这是解决通用机器人唯一的办法。我不认为还有别的路。没有仿真。你没法这样仿真真实生活。能够把这些数据带进来，并最终把它用作模型的基础先验，将是必需的。

So Index is our internet-scale pre-training initiative, almost like, how do we go out and get all the world's understanding of physics that's happening, bring it in, clean it, train on it, and transfer that into the robot. And that's been the ultimate goal. That's the north star. I think that's the only way to solve general robotics. I don't think there's any other way. There's no sim. You can't sim real life like this. Being able to bring this data in and ultimately use it as the base prior here for the model is just going to be necessary.

所以我们在过去大约一年里启动了这个计划。过去六个月——四、五、六个月——我们真的投入很重。那张图就像一根完整的曲棍球杆。我们大概有两到三个月，这些数据完全没有增长，我每天早上都在站会上。我就想，天哪，这太痛苦了。我们就是搞不定。而且不只是采集。你需要合适的传感器套件，你需要把观测-动作空间和机器人匹配得非常好。理想情况下，你需要清洗它。有欺诈。还有一堆——我们一开始是出去买的。全球大概有五六十家创业公司在做这个，但就是不行。我们出去买了一堆东西，结果就是垃圾。

So we've spun up this initiative over the last roughly year. We've really been invested into it heavily the last six months — four, five, six months. The chart is like a full hockey stick. We had probably two to three months where we weren't growing at all on any of this data, and I was in standup every morning. I was like, god, this is so painful. We just couldn't get this to work. And it's not just even collecting it. It's like you need the right sensor suite, and you need to match the observation-action space really well to the robot. Ideally, you need to clean it. There's fraud. There's just a bunch of — we initially went out to go buy it. There's about 50, 60 startups that do this worldwide, and it just didn't work. We went out and bought a bunch of stuff and it was just crap.

所以我们大概五六个月前决定自己来做，然后大约三个月前我们撞上了一个巨大的拐点，数据每周环比都在猛涨。就像一条直线——像一根曲棍球杆图。我们还没有见顶。所以这就是我们现在的位置。这就是我们用于预训练的数据输入。而我认为让 Helix 2.5 真正成为可能的，是我们有了这个数据源。我们有大量算力。所以这是个很好的例子——这是 Index 的更新。我们已经超过了——我想现在也许——我觉得可能接近翻倍了。

So we ended up taking it on ourselves about five, six months ago, and then about three months ago we hit a massive inflection, and the data has just been ripping every week over week. It's just like a straight — it's like a hockey stick chart. We haven't leveled off. So that's where we're at now. That's our data input for pre-training. And so what I think made Helix 2.5 really possible is we had this data source. We have a large amount of compute. So this is a good example with — this is an Index update. We've had over — I think maybe now maybe like — I think maybe it's close to doubled.

我们现在每秒收集 50 分钟，而不是 30 分钟。我们已经通过系统上传了将近 2300 万或 2400 万个视频。我们的支付运行率远高于此。我预计我们会在短时间内为此花费数亿美元，如果事情按我们想的那样发展，长期来看会达到数十亿美元。这将是一件大事。这是模型的输入。训练算力是我们真正试图扩展的另一个因素。我们想要输出的是效果良好的机器人行为。我们看到的是，随着我们扩大这个基础人类预训练，机器人正在处理它们从未见过的任务。这些出租房屋不在我们的训练数据中。这些被子和枕头也不在那里。我们不在那些房子里。我们租了它们，把机器人放在那里，它就能工作。这就是你想从产品中看到的。你不希望任何人在你家里远程操作或任何那种废话。你真正想要的是真正的自主性。它还没有解决，但斜率看起来不错，我们将尽可能努力地推动它到极限。

We're collecting 50 minutes now instead of 30 minutes every second. We've had almost 23 or 24 million video uploads through the system. Our run rate for payments is way higher than that. I project that we will spend hundreds of millions on this in a short period of time, and then billions over time if things work out the way we think. This will be a big thing. This is the input to the model. Training compute is another factor we're really trying to scale. What we want to come out is robot behaviors that work well. What we're seeing is that as we grow this base human pre-training, the robots are working on tasks they've never seen before. These house rentals weren't in our training. These comforters and pillows weren't in there. We weren't in those houses. We rented them, put the robot there, and it was able to do work. That's what you want to see out of your product. You don't want anybody teleoperating in your home or any of that nonsense. You really want real autonomy. It's not solved yet, but the slope looks good, and we're going to push it to the limits as hard as we can.

我认为斜率是最让我印象深刻的事情，对吧？就像理解，好吧，随着数据乘以 X 量，我们看到的准确率或成功率的增加是 Y 量。我知道有很多人过去四五年一直在关注 AI，现在他们开始更多地了解机器人技术，而猜测是，好吧，这在 LLM 上效果很好，它也能很好地转化到机器人模型上吗？我想知道，LLM 的 Scaling（规模扩张）与你们在机器人技术中看到的有任何核心区别吗？或者有什么让你们感到惊讶的事情吗？比如，随着数据增加了这么多，你提到了曲棍球棒式增长，对吧？它在一段时间内是平的，然后突然有一个大的跳跃。在那个缩放定律中，有什么在机器人产品方面你们发现的有趣的事情吗？

I think the slope was the thing that stood out to me the most, right? Like understanding, okay, as data is multiplying by X amount, the increase in accuracy or success rate that we're seeing is increasing Y amount. And I know there's a lot of people who have followed AI for the last four or five years and then now they're starting to learn more about robotics and the speculation, right, is like, okay, well, this worked well with LLMs, will it also translate over well to robotics models? And I guess are there any core differences between kind of how LLM scaling has come about versus what you guys are seeing on robotics or has there been anything surprising to you guys as it's like oh you know as data has increased by this amount you know you mentioned the hockey stick growth right it was kind of flat for a while and then all of a sudden there was a big jump within that scaling law is there anything that stands out as interesting that you guys have found on the robot product side.

是的，我认为最有趣的事情是我们正处于一条 Scaling（规模扩张）曲线上。这是我们真正兴奋的地方。我认为这就是我们过去几年一直缺失的：一种可预测的方式来扩展行为。我们知道如何找到一个用例，对机器人进行后训练并执行它。Figure 在这方面非常擅长，也许是世界上最好的。但这不会让你出货 1 亿或 10 亿台机器人。你真正需要能够从世界上发生的所有这些不同经验中学习。所以我们的第一反应是：我们能否处于我们在 LLM 中看到的相同的缩放曲线上？我们正处于一条 Scaling（规模扩张）曲线上。如果你看看 LM 的缩放曲线，比如对数损失随数据和算力的变化，我们今天可能只是其中很小的一部分。所以我们处于数据和算力的早期 GPT-1 阶段。我们只有一点点，但在这一点点中，它看起来很棒。我会说 10 投 10 中。我们实际上在缩放数据上画了一条线。它在我们的博客中。我们有一个对数模型损失和斜率，一条预测线，以及一条我们从模型实际结果中看到的线，它 literally 是一条完美的线。我们当时觉得，这看起来太好了。我们在争论，如果它有点不同就好了。如果你能放大很远，你可以看到一些细微的差异，但在这些水平上，真的很难看到。到目前为止一切顺利。我们需要达到现在的 100 倍甚至 1000 倍。我们必须在机器人技术中前所未有的水平上前进。好消息是我们有资金。我们觉得我们现在有资金来做这件事，我们觉得我们有建立在数据上的管道，我们现在有算力，而且明年会更大，显著更大。所以我们现在确实有了这本关于扩展智能的书中的这一章。这只是一个钱的问题。如果你没有钱，你就做不到。你玩不了。没有神奇的数据集，比如十万或一百万小时的数据会神奇地出现在机器人的 AGI（通用人工智能）中。这不会发生。所以这需要花费数百亿美元，我认为，随着时间的推移。这将需要数百亿，也许数千亿，甚至数万亿美元的营运资金，在机器人方面用于制造和资本支出以及零件的营运资金。这将需要巨大的努力。所以我们现在只是处于这个的早期章节。

Yeah, I think the most interesting thing is that we are on a scaling curve. That's what we're really excited about. I think that's what we feel we've been missing here for the last few years: a predictable way to scale behaviors. We know how to figure out a use case and post-train the robot for it and execute it. Figure is really good at this, maybe the best in the world. But that's not going to ship you 100 million or a billion robots. You really need to be able to learn from all these different experiences in the world that's happening. So our first reaction is: can we be on the same scaling curves that we've seen with LLMs? We are on a scaling curve. If you look at LM scaling curves over losses, like logarithmic losses over data and compute, we're probably a really tiny fraction of that today. So we're in the early GPT-1 phase of data and compute. We just have a little bit of it, and within that little bit, it's looking great. I would say 10 for 10. We actually have a line drawn over scaling data. It's in our blog. We have a logarithmic model loss and a slope, a predicted line, and a line that actually we saw from actual results of the model, and it's literally a perfect line. We were like, it just looks too good. We were debating, this would be nice if it was a little bit different. If you could zoom in really far, you can see that there's some subtle differences, but at these levels, it's really difficult to see. So far so good. We need to 100 or maybe a thousand times where we're at now. We just got to go at levels unforeseen for robotics before. The good news is we're capitalized. We feel we're capitalized to do this now, and we feel we have the pipelines built on data and we have the compute now that's getting larger and even into next year which significantly larger. So we do now have this chapter of the book of scaling intelligence. It's just a money thing. If you don't have money, you can't do it. You can't play. There is no magic data set of a hundred thousand or a million hours of data that will magically appear for AGI for robots. It's not going to happen. So this is going to take tens of billions of dollars, I think, over time. It's going to take tens of billions, maybe hundreds of billions, maybe trillions of working capital on the robot side for manufacturing and capex and working capital of the parts. It's going to take just enormous effort. So we're just in the early chapters of this now.

我能向刚加入直播的观众更新一下吗？我们有 Figure 的 CEO Brett Adcock 在这里，他宣布的，公司宣布了机器人技术的圣杯是能够泛化，意味着能够在未见过的地点工作。他们租了 30 个家庭，已经在收集所有这些视频。他们看到他们的 Felix 2.5，他们作为机器人大脑的 AI，正在没有任何新训练的情况下执行任务。你只要出现在任何家庭，它就能做到。你说更难的问题是，一个人形机器人能否进入一个它从未见过的家庭，立即用整个身体自主开始工作，你说这就是你看到的。我想问你有多少任务，因为我喜欢视频的这一部分，因为你基本上展示了各种各样的任务，它只是继续下去。这些是你捕捉的视频。这是你训练它的视频。然后机器人现在能够只是... 我们在这里谈论多少任务？

Can I update the folks who just joined the stream? We have the CEO of Figure, Brett Adcock here, and what he's announced, the company has announced the holy grail for robotics is being able to generalize, meaning to be able to do work in unseen places. They rented 30 homes and they're already collecting all these videos. And they have seen that their Felix 2.5, their AI that as the brain for the bot, they're doing tasks without any new training. You just show up in any home and it can do it. And you're saying that the harder question is can a humanoid just enter a home it's never seen, immediately get to work with its whole body on its own, and you're saying that is what you're seeing. I wanted to ask you how many tasks, because I love this part of the videos because you basically show all sorts of tasks and it just goes on and on. And these are the videos you captured. This is the videos that you've trained it on. And then the bot is now able to just... How many tasks are we talking about here?

我们收集了多少任务？以及它能做多少。

How many tasks were we like collecting? And that it can do.

不，不，不是说你收集数据，而是它现在能像你说的那样泛化，走进去就能做。

No, no, not that you're collecting but that it can now generalize like you said walk in do it.

是的。这个实验非常严谨。我想明确一点：这里的实验聚焦于数据缩放曲线。我们想选取一组特定的任务。我们选了三个。

Yeah. This experiment was really disciplined. I mean, I want to be clear: the experiment here was focused on data scaling curves. We wanted to be able to take a certain set of tasks. We chose three.

好的。

Okay.

来展示随着数据增多，机器人变得更好。我们量化的方式是：它收拾客厅所有玩具的成功率是多少，它从头到尾铺床的成功率是多少，它叠所有毛巾的成功率是多少？然后我们想看看随着数据进来，更大的数据进来，这些指标变好了，百分比提高了，确实如此，这很棒。这就是这个实验的目的。然后我们还兴奋的是，我们能够拿这个基础预训练，扩展它，然后能够展示我们能在你家里涌现出哪些行为。就像你在这里展示的视频，人类在做大量疯狂的事情。我在看那个。我当时想，不只是那个，还有你刚展示的索引训练，就像有人在推割草机。就像，你知道，这就像这是索引，但没错，就是这一个。就像有人在烤饼干，或者有人在吃东西或做饭，有人拿着纸杯蛋糕，有人用卷尺。就像你看到一种疯狂的多样性。我当时想，你必须有成百上千万个任务。

To show as data got more data came in, the robot got better. And the way we quantified that is like what percentage of time did it pick up all the toys in living room and what percentage of time did it make the bed end to end and did it fold all the towels? And then we wanted to see as data was coming in, larger data was coming in, those got better, like the percentages increased, and they did, which was great. That was the point of this experiment. And then what we are also excited about is we're excited about being able to take this base pre-training, grow it, and then be able to show what behaviors we can emerge from this in your home. As you showed to the video here, humans are doing like a crazy amount of stuff. I was watching that. I was like, not just that, but like the index training that you just showed, like there was somebody mowing like a lawn mower. There was like, you know, this is like this is index, but yeah, this one right here. It's like somebody's making cookies or somebody eating or making food, somebody holding cupcakes, somebody using tape measure. Like this is just like you see like a wild diversity. I was you must be like hundreds or millions of tasks.

这里数量巨大。所以，但你希望你的机器人能够做所有这些。你知道，所有我们想要的工作。我们想在那里翻汉堡。我们想让它做饭、洗衣服，像做物流。我觉得这些是你将来会希望人形机器人能做的事情。好消息是，人类的硬件非常适合这个。就像，它只是一个人，它可以进入人类空间。它可以做人类的事情。它可以抓取人类物体。就像，你知道，人类的硬件只会变得更好。它只会更像人。所以，我不知道，只是，是的，我想这里的重点是，我们正在看到最早的缩放定律，随着数据进来，系统获得更好的性能。

Here just enormous amount. So, but you want your robot to be able to do all of that. You know, all work we wanted. We wanted to flip burgers there. We wanted it to make food, do your laundry, like do logistics. Like, I think these are like things that you're going to want humanoids to be able to do. And the good news is that human hardware is perfectly suited for this. Like, it's just a human and it can go into human spaces. It can do human things. It can grab human objects. Like, you know, the human the hardware will only get better. It's only more humanlike. So, I don't know just yeah, I think like I think the point here is like we're seeing the first early scaling laws where as data is coming in the systems getting like systems getting better performance.

好的，说到这个，我有两个问题。第一，你拿回所有租房的押金了吗？

Okay, speaking of which, I got two questions. One, did you get your security deposits back for all the rentals?

你会继续租它们。

You're gonna keep renting them.

是的。

Yeah.

是的。我实际上问了我的团队，我们昨晚在这里待到很晚。我说：“人们知道我们要带机器人进来吗？”他们说：“是的，每个人都会被告知。”

Yeah. I actually asked my team, we were here late last night. I was like, "Do people know we're bringing robots in?" And they're like, "Yeah, everybody gets briefed on it."

但这一定很奇怪。清洁人员进来。就像，“等一下。”就像这个地方一尘不染。怎么回事？或者你们真的在用吗？人类真的在用吗。

But it must be weird. The cleaning crew comes in. It's like, "Wait a minute." Like the place is spotless. What's going on? Or are you actually using it? The humans actually using it.

不，不，不。就像我们不喜欢睡在那里，但就像我们用它来评估机器人。

No, no, no. Like we don't like sleep there, but like we're using it for like evaluations for the robots.

是的，但你们不做饭或类似的事情。你们没有制造正常的混乱。这不像你们真的住在那里。

Yeah, but you're not cooking or anything like that. You're not making like a normal mess. It's not like you've really lived in there.

嗯，我不知道。我的意思是，我们试图做其他事情，我的意思是，我们就像我们在那些家里有很多事情，比如烧烤之类的，

Well, I don't know. I mean, we're trying to do other I mean, we're like we got a lot of stuff going on those homes like grilling and stuff here,

但但但至少你离开时状态良好，能拿回钱。另一件事是你谈到了你谈到了硬件变得更好

But but but at least you're leaving in the state that you do get your money. The other thing is you talked about the you talked about the hardware getting better

上周我在那里时，我们走了走，你带我看了钢铁侠测试区在哪里。在钢铁侠测试区后面，有一个烟熏玻璃房间。有什么你想谈谈的吗？有什么你想说的吗？因为我们知道你完成了设计锁定。

And when I was there last week, we walked around, you showed me kind of where like the Iron Man testing area is. And behind the Iron Man testing area, there was kind of a smoke glass room there. Anything you want to talk about that? Anything you want to say? Because we know that you've done the design lock.

是的。不，

Yeah. No,

Figure 4，但没有更新。啊，糟糕，糟糕，糟糕，糟糕。

Figure four, but no update. Ah, rats, rats, rats, rats.

是的。就像不，不，我的意思是没有 Figure 4 的东西。我们就像，听着，我会说几件事。我们就像，你知道，当我开始这个时，我们知道人类硬件要达到真正完美需要很多时间。就像，我们仍然把 Figure 3 看作 iPhone 1 之前。它就像翻盖手机。它就像诺基亚。而且你知道，人类心脏比手机更复杂。我们需要更长时间才能像手机线那样饱和。所以，我们，你知道，我们很快造出了 Figure 1，就像直接推出门。我们造 Figure 2 更像架构完整，Figure 3 是我们的第一个量产机器人。所以我们很兴奋能，你知道，让它变得更好。在 Figure 3 发货后，我嗯。是的。是的，没错。看，所以这是 Figure 1。你知道，Figure 1 很酷。它实际上在这个房间里。我还没有

Yeah. Like no no I mean no figure four stuff. We're like, listen, I will say a few things. We got like you know we when I started this we knew like the human hardware to get to really perfection would take a lot of time. Like it was like we're, you know, we still view like figure three as like pre iPhone one. It's like a flip phone. It's like a Nokia. And you know, Human Heart was like more complicated than a phone. It's going to take us like longer to saturate like the phone line did. So we, you know, we built figure one really fast, like just get it out the door. We built figure two to be like more like architecturally complete and figure three is our first like production robot. So we're excited to like get, you know, make that even better. After figure three shipment, I um Yeah. Yeah, exactly. Look, so this is figure one. You know, figure one's badass. It's actually in this room. I haven't

华丽。它像个骑士。

Gorgeous. It's like a knight.

就像我们这里有它。

It's like we have it here.

我想要一个。你说过你要送人。请给我一个。

I want one. You said you were going to give it away. Please give me one.

那些是 我们没说 Figure 1。我们只有几个 Figure 这里有一个。

Those are We didn't say figure ones. We only have a few figure Here's one right here.

嗯

Um

我喜欢它。

I love it.

我有 我们有几个，但我有一个在我家里。但就像我们，我们喜欢 Figure 1，就像，你知道，我们基本上在创业一年内就做到了。所以很疯狂。这里的 Figure 3。我的意思是，我完成了 Figure 3，我当时想我在这里没想法了。我当时想那是我最好的机器人，而且，有趣的是我们运行 Figure 3 很多次，我们已经运行了一年多。而且当你在未来把这些机器人放在一起时，就像你有 Figure 1、2 和 3，我们把它们放在那边的墙上。嗯，Figure 4 将是我们有史以来最大的进步。你想，如果你把它画成图表，Figure 1、Figure 2 变得更好，Figure 3 变得更好。就像 3 和 4 之间的差距将是我们有史以来最大的差距。它嗯，我现在真的对 Figure 5 没想法了。就像我真的不知道我们到底能做什么，因为 Figure 4 只是 Figure 4 是人形机器人的 iPhone 1 时刻。确实是。

I have We have a couple of, but I have one move in my house. But like we uh we like figure ones like was just like, you know, we basically walked that within a year of of uh of starting the company. So it was crazy. this figure three here. I I mean I got done with figure three and I was like I'm out of ideas here. I was like that was my like that was my best robot and um man the funny thing is we ran figure 3 a lot and we're running over a year now. And um when you're when you're going to put these robots next to each other here in the future like when you have like figure one, two, and three, we have them over here on the wall. Um, figure four will be the largest step up in per in like uh of we've ever made. You think like if you like chart it like figure one, figure two got better, figure three got better. Like the the gap between three and four would be the largest gap we've ever had. It is um I I am truly out of ideas after right now for figure five. Like I don't really know what the hell that we could do because figure four is just figure four is an iPhone one moment for humanoids. It is.

是的。但但但你没注意到每次你以为没想法了，你造出下一个东西。当你造出下一个东西时，突然那些想法就来了。

Yeah. But but but don't you notice every time you think you're out of it, you build the next thing. And as you build the next thing, suddenly those ideas come.

我现在对未来该做什么有几个更多想法，我们没时间在 Figure 4 上做。但就像，老兄，Figure 4 将是 嗯 你会看着 Figure 4，也许不知道它就像它是 Figure 系列。它只是疯狂。我们下了几个大赌注，它们都成功了，你知道，而且嗯是的。我只想说我们不会展示给任何人。

I have a few more ideas now what to do for the future that we didn't have time to do for figure four. But like, dude, figure four will be um you'll look at figure four and maybe not know it's like it's a figure lineup. It's it's just nuts. We took a couple big bets and they all worked, you know, and um yeah. And I just want to say like we're not showing anybody.

不给你看，Scott，还有——别这样。

Not showing you Scott and — come on.

大家都知道我还没见过它。我唯一看到的就是那些字母。造机器人很难。我们正在攻克所有难题，处理各种事情。等我们准备好了就会分享。我跟你说——这话听起来会很夸张。我甚至不确定我该不该说。它会让 Figure 3 看起来像原始人。太疯狂了。我认为 Figure 3 是机器人市场上迄今为止最好的人形机器人。我觉得没有任何东西能接近它。没有任何东西能与它竞争。这就是为什么我们的自主性做得这么好。人形硬件就是最好的。Figure 4 就是——我的意思是，你在这里有一个团队，你知道，我和团队已经埋头苦干了四年，还是同一个团队。他们就是很厉害。然后我们把所有的一切都投入到了 Figure 4 中。所以，不管怎样，将来某一天我会很兴奋地向你们展示。它将会是我们设计过的最具突破性的东西。

So everyone knows I have not seen it. Only thing I saw was the letters. Uh it's hard to build a robot. We're working through all the hard problems, working through the stuff. We'll share when we're ready. I'm just telling you — this is gonna sound terrible. I don't even know if I want to say it. It's going to make Figure 3 look like a caveman. It is insane. And I think Figure 3 is the best humanoid in the robot market by far. I don't think there's anything even close to this. There's nothing that even can compete with it. It's why we do autonomy so well. The humanoid hardware is just the best there is. Figure 4 is just — I mean, you've had a team here, you know, me and the team have been slaving away for four years building stuff, same team. They're just good. And then we just put everything we had into Figure 4. So anyway, I'll be excited one point to show you guys here in the future. And it's just going to be the most groundbreaking thing we've ever designed.

Figure 3 现在的产量是多少？

What's your run rate for Figure 3 right now?

你说的产量是什么意思？

What do you mean by run rate?

你们造了多少？每周造多少？

How many have you made? How many are being made per week?

我们上个月造了 1 万台。团队把它涂成了金色。我们在 Baku 前面拍了张照片。事情进展得很顺利。我们现在有很多机器人。有一件很有趣的事——你知道，我们就是有太多机器人了。两年前我们还有展示给你的 Figure 1。我们这里只有几台 Figure 1，每个人都——整个团队都像看宝贝一样看着它们。你得拿到机器人，得有机器人时间，每个人都需要机器人。就连启动和测试人员也需要机器人，嵌入式软件、固件、操作系统、中间件——但你得去我们的 AI 和软件团队那里才能让这东西运转起来。所以他们有优先权。但现在我们的机器人比工程师还多，所以每个人都有机会接触机器人。就像到处都是机器人，甚至校园里也到处都是。太疯狂了。这很棒，它们能工作，正在工作。所以就是很棒。我们在现场有制造部门，所以我们可以真正——我非常喜欢它离工程部门很近。工程和制造之间有一个持续的反馈循环。造它的人就坐在做它的人旁边，他们可以——每当有问题，我们打电话，我给他们打电话，我们走到隔壁就把它修好了。在某个时候这会变得非常好，意味着整个面向制造的设计和设计到制造的循环，我们正在变得更好。在某个时候我们显然会离开园区去制造，在园区外进行更大规模的生产，但在这里有这条线让我们真正迭代并真正擅长制造，是非常健康的。

We had the number 10,000 made, I think last month. The team painted it gold. We had a picture out front of Baku. And things are going great. We have a lot of robots now. One thing that's really funny is — you know, we just have so many robots. Two years ago we had the Figure 1 we're showing you. We had a couple Figure 1s in here and everybody — the whole team was looking at them like a prize. You got to get the robots and get robot time, and everybody needs a robot. Even bring-up and testing folks need robots, embedded software, firmware, operating system, middleware — but you kind of need to go to our AI and software team to work through it, to make this thing work. So they got priority. But now we have more robots than engineers, so everybody gets the ability to be on the robot. It's just like there's robots everywhere, and I mean even on campus they're just everywhere. So it's crazy. It's great and they work and they're working. So it's just great. We have manufacturing here on site so we can really — I really like it being close to engineering. There's a constant feedback loop between engineering and manufacturing. The folks that built it are next to the folks that make it, and they can just — whenever there's a problem, we call, I call them, we get next door and we fix it. And at some point that will get really good, meaning the whole design for manufacturing and design to manufacturing loop, and we're getting better at it. At some point we'll obviously leave the campus for manufacturing and have higher volumes outside of campus, but it's been really healthy to have this line here that we can really iterate through and really get good at manufacturing.

嗯，Brett，我们知道你时间非常有限。非常感谢你能来。在聊到产品线的时候，我还有一个最后的问题。你之前在想，现在机器人太多了，Figure 2 该怎么处理？有一条推文疯传。施瓦辛格来了，对吧，带着可能是最经典的机器人台词。是的。就像我们该怎么处理 Figure 2？他就说：“你应该把它们熔掉。”这简直让互联网炸了。我妈妈把那条推文发给了我。

Well, Brett, we know you have very limited time. Really appreciate you hopping on. I had one last question as we're talking through the lineage. There was a tweet that went very viral when you were thinking about, you know, now you have too many robots. What do you do with the Figure 2s? Schwarzenegger came in, right, with maybe the most iconic robotics line. Yeah. Like what should we do with the Figure 2s? He just goes, "You should melt them." Which like broke the internet like that. My mom sent me that.

是的。你想好怎么处理 Figure 2 了吗？

Yeah. Have you figured out what you're going to do with the Figure 2s yet?

是的。嗯，是的。我妈妈说：“你在网上被碾压了。”我说：“妈，你在说什么？施瓦辛格。那可是施瓦辛格。你在说什么？”他的浏览量和点赞数是我的 20 倍。她说：“Bo。”呃，是的。挺有趣的。所以我私信了——我们聊天的时候我给阿诺德发了消息。目前我只能说这么多。

Yeah. Um, yeah. My mom's like, "You got ratioed out there." I was like, "Mom, what are you talking about? Schwarzenegger. It's Schwarzenegger. What are you talking about?" Had like 20 times more views and likes than mine. She's like, "Bo." Uh, yeah. It's funny. So I DM'd — I messaged Arnold when we talked. And that's all I'll say for now.

太棒了。是的。是的。我是说，那家伙是最棒的。就像我——你知道，我们都知道终结者是什么样，但我从小看着长大的。太棒了。这真是很棒的科幻，而且他是个非常棒的人。我们对他有了更多了解。所以是的，很快会有更多消息，伙计。

Love it. Yeah. Yeah. I mean, that guy's the best. Like I just like — you know, we all know how the Terminator is, but I grew up watching. It was great. This is just such a great sci-fi and he's a very awesome person. We've gotten to know him better. So yeah, more soon, man.

他是好机器人。他变成了好机器人。对。

He was the good robot. He became the good robot. Right.

他是个很棒的代表。

He's a great representative.

如果你——有趣的是，阿诺德是唯一一个既当过终结者又拯救我们免于终结者的人。所以也许他是世界上最好的人选，来确保我们——你知道，让我们朝着正确的方向前进。

If you — it's funny enough, Arnold's like the only guy that's like been a Terminator and saved us from Terminators. So maybe like the best person in the world to make sure we like put the ro— you know, set us up in the right direction here.

好的。嗯，你们，Brett，你们简直火力全开。每周似乎都有某个重大公告震惊所有人。那个 index 的东西，你们已经做了一年。它出来了。你们比大多数人有更多的数据。我们不知道，但你现在展示出来了。你说这至少是泛化的第一步，或者——你可以进入任何环境，学习这三个任务，然后你展示你实际上在教它无数其他任务。然后你有你的 bot Q，你很快要推出 Figure 4。你的园区里有一千个机器人，而你的员工不到一千人，对吧？所以机器人更多。它们也有员工徽章，对吧？它们实际上会刷卡进出，不是吗？它们会穿过门。

Okay. Well, you guys, Brett, you guys are just on fire. Every single week there seems to be some big announcement that shocks everybody. The index thing, you guys have been working on it for a year. Comes out. You've got a lot more data than most people. We don't know, but you're showing it now. You're saying that this is a first step anyway to generalize or it's — you can go into any environment and learn these three tasks and then you show that you're actually teaching it like countless of other tasks. Then you've got your bot Q, you've got Figure 4 coming out soon. You got a thousand robots in your campus and there's like what you have less than a thousand employees, right? So there's more robots. They have employee badges too, right? They actually do badge in badge out, don't they? They walk through doors.

还有更多。就像几个月前，大概三个月前，我们在园区里跨过了员工与人形机器人的比例。

There's more. We like it was like a few months ago, like maybe three months ago, we crossed like the employee to humanoid ratio on campus.

在我的生活中就像这样——被机器人碾压了。

Back in my life like this — ratioed by the bots.

碾压。

Ratio.

被机器人碾压了。所以什么——是的。好的。嗯，你们简直太厉害了。干得好。然后我们看看接下来会发生什么。

Ratio by the bots. So what — yeah. Okay. Well, you guys are just killing it. So great job. And then we'll see what comes next.

谢谢你们邀请我。是的。还有 Kevin。酷。很酷的衬衫，伙计。这是我最喜欢的。

Thanks for having me on. Yeah. And Kevin. Cool. Cool shirt, man. It's my favorite.

嘿，这也是我最喜欢的衬衫。

Hey, this is my favorite shirt, too.

是的，很棒。那是 Figure 的短袖衬衫。是的。酷。谢谢你们邀请我，伙计们。再见。

Yeah, it's great. That's a Figure short shirt. Yeah. Cool. Thanks for having me on, guys. See you.