仓库中的自主机器人：现场演示

等等，Rex 来了。

Hang on. Rex in the house.

哎呀。

Uh oh.

各位，我们这里有红队了。

We have red in the house, ladies and gentlemen.

好吧，好吧。

All right. All right.

很高兴见到你，Brett。

Great to see you, Brett.

嗯，把我们拉回来，这样你就能在右下角下单了。

Well, drop us bring us back in so you can order down in the bottom corner.

是的，当然。马上。

Yes, absolutely. In a minute.

你们一直在做什么？

What have you guys been doing?

嗯，我们一直在看 Pink Drive。

Well, we've been watching Pink Drive.

太棒了。你们在看机器人。

Awesome. You're watching the robot.

是的，当然。很好。

Yep. Absolutely. Good.

而且我们都会……所以是的，我们一直在看，Brett。你也是。我的意思是，你一直都能看到。所以我猜你有更多别的事情要操心，而不是盯着机器人看八个小时。比如你有真正的工作。

And we're all gonna be so... So yeah, we've been watching it, Brett. You as well. I mean, you see it all the time. So I imagine you've got more other things to worry about than watching the bot for eight hours. Like you got a real job.

我们运行，就像我们运行……我们有很多……我们在办公室很多不同的地方运行像这个用例一样的机器人。

We run like we run like we have many... we run a bot just like in this use case in many many different places in the office like this.

是的，没错。

Yeah. Right.

是的。

Yeah.

相当疯狂。

Pretty crazy.

我们做这个已经好几个月了。就像……是的。我们喜欢……我觉得它处于一个不错的状态。我的意思是，我觉得……好吧，对我们来说疯狂的是，我们到了一个系统真正稳定的阶段。也就是说，这个系统运行的是 Helix 2。它基本上是从像素到动作的全身神经网络。从细微的脚步到所有关节，都由 Helix 指挥。我们必须通过像素进行推理。所以我们必须能够以非常高的频率感知世界并理解该做什么。我们需要能够将包裹推下传送带系统，基本上每 3 秒一个包裹，这达到了人类速度的 KPI。我们必须在条形码朝下放在传送带上的情况下，保持 90% 的成功率。它们被扫描，然后最终在另一侧贴上另一个标签，用于送货卡车。然后我们不仅需要这些，如果我们想 24/7 运行，我们必须让机器人相互通信。当一个机器人在传送带上电量低时，它会通知后面的机器人来接替它的位置。所以当这一切发生时，后面的机器人正在过来，等待的机器人退出传送带，然后基本上机器人就跳进去。所以在 30 秒内，我们又重新开始移动包裹，停机时间非常短。然后机器人回去充电。充满电只需大约一个小时。然后任何时候机器人有任何问题，它们会去维护。然后我们通知车队再带一个机器人进来。所以我们的状态空间实际上相当大。

We've been doing this for like for months. Like this is just... Yeah. We like... I think it's in a good spot. I mean, I think like the... Okay, so the crazy thing for us is we're at a point where the system is really stable now. And meaning one is like this is running Helix 2. It's basically running a full body neural network from pixels all the way to actions. Everything from subtle footsteps to all the joints commanded from Helix. We have to reason through pixels. So we have to be able to see the world and understand what to do at a very fast frequency. We need to be able to push the packages down the conveyor system basically around 3 seconds a package, which is at the human speed KPIs. We have to do all this while being at like 90% successful with barcodes being face down on the conveyor system. They get scanned and then ultimately there's another label shot on top on the other side for the delivery truck. And then we need not only that, but if we want to run like 24/7, we have to have robots communicating with each other. As one robot is low on state of charge in the conveyor, it messages a robot behind it to come in and take its place. So while that's happening, the robot's coming behind it, waiting robot's backing off the conveyor, and we just basically the robot just jumps in. So within 30 seconds, we're back to moving packages again with basically very minimal downtime. The robot then goes back and charges. Takes only about an hour to full charge. And then anytime the robot has any problems, they head to maintenance. And then we message the fleet to bring another robot in. So we have like a... the state space is actually pretty large.

而且机器人完全自主地做这一切，彼此自组网，互相发送消息，基本上是在移动机器人。我的意思是，这相对简单。这是一个简单直接的用例，但机器人和周围其他机器人可能出很多不同的差错，这些都必须自主推理解决。所以，总之，是的，这太棒了。很高兴看到，我们也很期待看看能否推进到 8 小时。

And the robots are doing this all autonomously, self-networked with each other, messaging each other, and moving basically moving robots around. I mean, it's relatively straightforward. It's like a simple pretty straightforward use case, but there's a lot of different things that could go wrong with the robot and with other robots there that have to kind of be, you know, autonomously reason through. So, anyway, like yeah, this is awesome. It's great to see, and we're excited to see if we can pull forward on eight hours.

那么，Brett，我们注意到那里有三个机器人而不是两个。显然，对于充电，你只需要两个。第三个是不是某种故障保险，以防其中一个出现硬件故障之类的？

So, Brett, we noticed that there are three bots there instead of two. Obviously, for charging, you would only need two. Is the third one there kind of a fail safe in case one of them has a hardware failure or something like that?

是的，我们有……然后，你知道，最坏的情况，如果机器人有硬件甚至软件问题，需要退出，我们可以……我们办公室里有数百个机器人。它们可以呼叫车队里的机器人，它们会进来并停靠。它们可以在任何地方停靠。所以，在某种程度上，我们基本上可以连续运行好几个月，不断替换不同的机器人，机器人出问题就换，我们基本上可以把办公室里的任何机器人投入到这个用例中。

Yeah, we have... and then, you know, worst case scenario, if the bots have issues with hardware or even software and they need to roll out, we could just... we have hundreds of robots in the office. They can just call robots in the fleet and they'll just come in and dock. They can dock anywhere. And so, in some way, we could basically just run for months and months and be subbing out different robots, robots having problems, and we'd be basically able to put any robot into the use case at the office.

那么，在你们现在运行的演示中，你们会收集……顺便说一句，很高兴见到你。你们是在同时收集数据和训练，还是这纯粹是一个演示，或者有点像……

So, in this demo you're running now, do you collect... Nice to meet you by the way. Are you collecting data and training at the same time, or is this purely a demo or kind of like...

顺便说一句，在你回答之前，Gustav Anderson 是 Gustav Anderson 医生。他是瑞典的手外科医生。所以……

By the way, before you respond, Gustav Anderson is Dr. Gustav Anderson. He's a hand surgeon in Sweden. So...

他是个很棒的人，到处都有很多……

Great guy all over the horizon to do a lot of...

把你的手拆开，观众们。

Pick apart your hand audience.

是的，是的，是的，是的。

Yeah. Yeah. Yeah. Yeah.

嗯……

Um...

交给你了。

Over to you.

嗯，你们想看机器人吗？我们走。你们想看看吗？

Well, you guys want to see the robot? We go. You guys want to check it out?

太棒了。那太好了。是的。

Awesome. That would be awesome. Yeah.

回答问题。

Answer questions.

好的，等一下。

All right. Hold on.

终极 Snory 摄像头。太棒了。

Ultimate Snory Cam. Awesome.

这就像圣诞节一样。

This is like Christmas.

我知道。说真的。

I know. Seriously.

好了。

There we go.

哦，天哪。所以，他就在那里。哦，我们得到了一个和其他人不同的视角。耶。

Oh gosh. So, there he is. Oh, we got a different point of view than everybody else does. Yay.

没错，没错。

That's right. That's right.

哇，太棒了。

Wow. That's fantastic.

那么，Brett，他主要是通过强化学习训练的，还是主要……是的，就是这样。如果他……如果他说“嗨 Brett”，机器人会回应吗？

So, Brett, was he primarily RL trained or was he primarily... Yeah, there you go. If he... if he say hi Brett, will the bot respond?

我不知道……我想知道他是不是听不到我们，因为他……是的，可能太吵了。

I don't know if... I wonder if he cannot hear us because he's... Yeah, it might be too loud.

好吧。你觉得呢？

All right. What do you think?

那很酷。

That's pretty cool.

太神奇了。是的，那真的很酷。

Amazing. Yeah, that's really cool.

当然，你还有其他的。是的，没错。

And of course, you've got the others. Yeah. Right.

嘿 Brad，我是 Phil Trouy。嗯，其实有个问题，Noah 博士，我就直接问了。这个机器人是如何被训练来完成这项任务的？你能大致告诉我们他用了什么技术吗？

Hey Brad, it's Phil Trouy here. Um, there was a question actually, Dr. Noah, I'll just ask. How was this bot trained to do this task? Can you kind of tell us even just in general like what kind of techniques he used?

是的。所以我们用 Helix 2 端到端地训练了这个。实际上我们网站上有一些好的信息。我们在这个用例和其他事情上都做过。Helix 2 基本上是我们内部设计的全身神经网络策略。它是一个视觉语言动作策略，我们已经训练了它。数据集基本上是……我们现在已经收集了这些数据，完成了训练，然后这里发生的都是测试时的推理。所以机器人现在通过一个全身控制器进行推理。所以仅仅从视觉和状态，机器人就能指挥细微的脚步等动作。机器人可以，你知道，冲刺和迈步，所有这些都来自相机条件，视觉条件。所以仅仅通过环顾四周，机器人就能同时指挥所有关节如何移动身体、手指、骨盆、脚步，一切，完全端到端。所有这些 AI 工作都在机载进行，实际上是在机器人躯干中。

Yeah. So we trained this end to end with Helix 2. There's some good information actually on our site. We've done on both this use case and other stuff. Helix 2 is basically a full body neural network policy we've designed in house. It's a vision language action policy, and we've trained it. The dataset is basically... we've now collected this data, done training, and then this is all test time inference happening here. So the robot is now reasoning through a whole body controller. So from just vision and state, the robot can command things like subtle footsteps. The robot can take, you know, lunge and do footsteps, all coming from basically camera conditioning, vision conditioning. So just through looking around, the robot's been able to command all joints simultaneously on how to move the body, fingers, pelvis, footsteps, everything, fully end to end. All that AI work is happening on board, actually in the robot torso.

机器人的大脑在躯干胸部，所有 AI 推理都在本地完成，不需要网络。

The brains are in the torso chest of the robot and does all that locally on board with no network needed to do any AI inference.

对，但训练数据是怎么采集的？是有人戴着 GoPro 头戴设备做这些动作吗？因为机器人总得学这些动作，我只是好奇是通过视频、遥操作还是两者结合。

Yeah, but how is the training data captured? Was it like somebody wearing a GoPro headset and doing these motions? Because somehow the robot has to learn these motions, and I'm just wondering if that was through video or from teleop or a combination.

对。我不想具体透露我们到底用了什么配方，但我们已经收集了从遥操作数据、人类视频数据到在线视频数据的一切，并且一直在这些领域训练和部署策略。最终，我们需要像人类一样看世界、像人类一样在世界中移动。所以最终，世界上最好、最普遍、最大的数据集就是人类数据。因此，在极限情况下，Figure 需要尽可能多地从人类那里学习。我是说，我们是人形机器人，对吧？我们有这个不公平的优势。我们有五根手指、手腕、骨盆，头上有摄像头，可以被动地通过 RGB 空间看世界。我们需要在人类数据规模上进行训练。

Yeah. I don't want to specifically comment on exactly what we've done as a recipe, but we've collected everything from teleoperation data all the way through human video data to online video work, and we've been training and rolling out policies on all those areas. Ultimately, we need to be able to see the world like a human and move around the world like a human. So ultimately, the best and most ubiquitous and largest data set in the world is human data. So in the limit, Figure needs to learn from humans as much as possible. I mean, we're a humanoid, right? We have this unfair advantage. We have five fingers, wrists, pelvis, cameras on the head that see RGB space passively through the world. We need to be trained on human data at scale.

说到这个，我能问一下吗？这些 Figure 机器人手上有摄像头，对吗？这些有还是没有？

And so actually speaking of that, can I ask you because these figures have the cameras in the hands, is that correct? These ones do or not?

那你能用这个吗？你能用这个进入作弊模式，用手上的摄像头吗？现在看起来不像，但也许它们确实在用。我不知道这个具体策略是否使用了手上的摄像头，但我们在所有工作中都大量使用手上的摄像头。如果手上的摄像头正在被使用，我也不会惊讶；我只是现在一时想不起来。

So are you able to use that? Are you able to sort of go into cheat mode with that and use the cameras in the hands? It doesn't look like it right now, but maybe they are. I don't know if this policy is using cameras in the hands or not in this exact one, but we do use cameras in the hands quite a lot for all of our stuff. I wouldn't be surprised if the cameras in the hands are being used; I just actually don't know off the top of my head.

目前，每个 Figure 3 的手上都有一个摄像头，并且已经投入使用。我们看到了很多好处。有很多区域，某些头部摄像头确实看不到，比如箱子、盒子，而且在很多情况下，我们是在头部前方抓取物品，部分被遮挡。所以手上装有摄像头非常有益。

We do have a camera in every Figure 3 hand as of now; that's operational. And we have seen a lot of benefits of doing that. There are a lot of areas where certain head cameras really can't see into, like bins and boxes, and then in a lot of cases we're grabbing items in front of the head, partially occluded. So having cameras in the hands has been very beneficial.

酷，酷。

Cool. Cool.

对，我们想知道的是：手上的摄像头能不能扫描看到条形码在哪里。所以如果它跑到后面，它就能做到。好的。我们只是不知道在这个案例中它是否在这样做。

Yeah, that's what we're wondering: is the camera with the hands could scan to see where the barcode is. So if it's run behind it, it would be able to do that. Okay. We just didn't know if it was doing this case.

在这个用例中，我们不需要扫描条形码。我们只需要知道条形码在哪里。

In this use case, we don't need to scan the barcode. We just need to know where the barcode's at.

对，看到它，看到它。

Yeah. To see it. To see it.

对。给正在观看的用户解释一下，这个用例是找到包裹，确定条形码在哪里，把条形码朝下放，然后放到传送带上。传送带从下面扫描条形码。然后几秒钟后，有一个气动贴标机，基本上快速打印出新标签，并气动地把它扔到包裹的另一面，用于送货卡车等。所以有人问为什么不把条形码朝上，或者为什么不能扫描它们。我们在这个案例中的工作不是做那个；这基本上是一条小包裹分拣线。

Yeah. Just for the users watching, the use case is to find the package, figure out where the barcode's at, put the barcode down, and then put it on the conveyor. The conveyor scans the barcode from underneath. And then a few seconds later, there's a pneumatic label machine that basically prints out a new label really fast and pneumatically throws it on top of the package on the other side, and it's used for the delivery trucks and elsewhere. So some people ask why not put the barcodes up or why not be able to scan them. Our job in this case is not to do that; it's basically a small package sorting line here.

好的。好的，它也知道条形码在哪里。它看着盒子，没有看到标签，但知道标签在背面，所以它正确地翻转了它。还是它必须看到标签才能推断出来，还是它能……

Okay. Okay, where it is as well. It looks at the box and it doesn't see the label but it knows that the label is on the back so it flips it correctly. Or does it have to see the label to kind of infer it, or can it...

对，我今天看了一会儿。有很多好例子，比如一个立方体、一个方盒子，标签在底部。如果它确定在四个侧面或顶部都没有看到标签，它就会判断标签在底部，甚至不看就直接放置，而且通常相当准确。

Yeah, I was watching a little bit today. There are a lot of good times where, say, a cube, a square box, and the label will be on the bottom. If it determines it has never seen a label on any of the four sides or on top, it'll determine it's on the bottom and just not even look and place it, and it's usually fairly accurate.

对，我注意到了。那真的很突出。我注意到了。对。

Yeah, I spotted that. That really stood out. I spotted that. Yeah.

对。所以，立方体盒子很有趣。它像拿魔方一样拿着它，可以移动它。

Yeah. So, the cube boxes are funny. Like it holds it like a Rubik's cube and can move it around.

你去年没有那个，对吧？我觉得去年的盒子更偏长方形。我不记得去年有立方体。

You didn't have that last year, right? I think the boxes were more rectangular last year. I don't remember any cubes last year.

立方体也有。

Cubes as well.

所以，看起来这是你新加入的，因为同样，它似乎在那个上面有点卡顿，因为它不太确定标签在哪里。如果是扁平盒子，那就很明显了。它是一个包裹。速度快得多。

So, it looks like it's a new one that you may have thrown in because again, it seems to be stuttering a little bit on that because it's not quite sure where the label is. Where it's a flat box, it's pretty obvious. It's a package. It's much quicker.

这些其实都只是数据问题。更快的速度得益于我们在 Helix 方面做出的一些架构决策，确实让它更快了。

These are all just data problems. The faster speed has been due to some architecture decisions we've made on the Helix side to definitely make it faster.

所以你那边的情况在变好，然后你有数据，比如更好的数据、更高质量的数据、更多样化的数据流入 Helix 进行训练。我们在网站上做了一些消融实验；你可以去看看，更多数据在提高成功率的同时也在降低速度。所以我认为如果你回顾去年我们在一些物流岗位上做的 Figure 2 工作，我们在网站上有一些统计数据，显示速度和准确性都随着更多数据而提高。所以很大程度上，盒子或者你看到的任何问题，在这一点上都是数据问题。如果你想让这个用例运行得更快、成功率更高，而且偶尔它会把条形码放在错误的一侧，这些都只是数据问题。

So you have that side of things getting better, and then you have data like better data, higher quality data, more diverse data flowing into Helix for training. That's all we've done some ablations on the website; you can go look at where more data is increasing success rate and also decreasing speed. So I think if you went back to the Figure 2 work we did on some logistics posts last year, we had some stats on the site where we showed both speed and accuracy improving with more data. So largely, the box or anytime you see any issues, these are all data issues at this point. If you want this use case to run faster and have higher success rate, and every once in a while it'll have the barcode on the wrong side, these are all just data problems.

你在 Figure 2 和 Figure 3 之间升级了 Nvidia 处理器吗？

Did you upgrade the Nvidia processor between Figure 2 and Figure 3?

我不这么认为。没有。

I don't think so. No.

好的。那你期待未来升级处理器吗？

Okay. And are you looking forward to a processor upgrade presumably in the future?

嗯，我们之前谈过这个。我们想知道你是否遇到了当前推理处理器的极限。这就是我们好奇的。

Well, we talked about this earlier. We were wondering if you're running up into the limits of your current inference processor. That's what we're wondering.

对，我们喜欢……对，听着。我们希望在设备上有更多的内存和更多的算力吗？答案是肯定的。我们目前在设备上的推理主要受内存限制。

Yeah, we like... Yeah, listen. Would we love to have more memory and more flops on device? The answer is we would love that. We are largely memory constrained on device for inference today.

我们确实计划对下一代机器人进行升级，以帮助解决这个问题。但正如你所见，在设备端运行模型至关重要。这有很多优势。一是速度，控制端动态变化非常快，所以在设备端运行模型很重要。显然，能够在没有网络或高延迟网络的区域运行也极为有利。所以我认为，最终你会获得更好的性能、更多的内存和算力，模型也会更大，机器人在设备端方面会变得更聪明、更快。

We do have a planned upgrade coming on the robots for future generations to help with this. But as you can see, having on-device models running is super critical. There are a bunch of advantages. One is speed, and we have such fast dynamics happening from the control side, so having on-device models is important there. Obviously, being able to operate in an area where you don't have a network, or maybe a high-latency network, is extremely beneficial. So I think in the limit, you're getting better, more memory, more flops, and you're getting bigger models, and the robots are getting both smarter and faster with on-device stuff.

那么，你是否预见到一个未来，你的机器人能完成一百种完全不相关的任务？你是否预见到机器人会为特定任务下载特定模型，就像 Figure 的应用商店一样？

So do you see a future where you have a hundred completely unrelated tasks that your bot can do? Do you see a future where the bot will download specific models for a specific task, like an app store for Figure?

不，我认为人形机器人的美妙之处在于，你基本上可以从迁移学习中受益。你可以让一个机器人拥有一个单一的数据池，理想情况下是一个单一的神经网络，它吸收所有不同类型的使用场景数据，跨越人形硬件，你基本上就像在文本 LLM 的预训练中那样，从这个池中学习。这是你从多种不同机器人形态中得不到的——你无法从其他事物中获得迁移学习的好处。这是人形机器人最大的优势之一：你基本上可以拥有一个 AI 大脑。你可以让数据从所有这些不同的使用场景流入，从而改善整个系统。我给你讲个故事。大约一年半前，我们在 Figure 2 上做了一些工作，当时我们在运行 Helix 的第一个版本 Helix 1，我们把一些物品放进冰箱。策略的成功率大约是 55-60%，把不同的调味品、奶酪等放进冰箱，而且它只是用冰箱数据训练的。我记得我们把与冰箱无关的数据量增加了两倍，比如打开柜子、抽屉，桌面上放东西，然后用更多样化的数据（冰箱加其他）重新训练模型。第二天，我在一个有限的评估中看到同一个模型在冰箱任务上做到了 90%。所以，我们从仅用冰箱数据的 60% 成功率，到用更多非冰箱数据重新训练同一个模型，甚至没有增加冰箱数据，数字就上升了。那时我们想，天哪，拥有极其多样化的数据确实有好处。这很有道理，因为我们在其他领域处理袋子的经验越多，在冰箱里处理奶酪袋也会更好。所以这是一个巨大的优势。这意味着，随着你在不同使用场景中引入更多样化的数据，机器人会变得更聪明、更好。我们在 Figure 3 上看到了这一点，在 Figure 2 上也看到了。这是一个巨大的现象。对于人形机器人来说，这就像一个奇迹，你基本上有一个单一平台，可以出去做几乎人类能做的任何事情。随着更多数据的引入，整个机队都在集体进步。所以，与其像应用商店那样有专门的模型，你基本上是把所有这些数据集中到一个模型中进行训练，然后训练一个大型离线模型，然后通过 OTA 更新每个机器人的权重。所以不公平的优势在于，一个人形机器人在某件事上变得擅长，比如物流，机队中的每个机器人都会在物流方面达到同样的性能水平。

No, I think the beauty of having a humanoid robot is you can basically benefit from transfer learning. You can have a robot that has one single pool of data, ideally one single neural network that is pulling in all this very diverse types of use case work across humanoid hardware, and you're basically learning from this as a pool, like you do in pre-training for text LLMs. This is what you don't get from having multiple different robot form factors—you don't get the benefit of transfer learning from other things. This is one of the biggest benefits to humanoid: you can basically have one AI brain. You can have data flowing in from all these different use cases, and it's bettering the entire system. And I'll tell you a story. We were doing some work on Figure 2 about a year and a half ago, where we were running Helix, our first version, Helix 1, and we were putting some items into a fridge. The policy was generating about 55-60% success rate of putting things in the fridge—different condiments, cheese, other things—and it was just trained on fridge data. I remember us tripling the amount of data that had nothing to do with a fridge. It was like opening cabinets and drawers, things on a tabletop, and training the model again with more diverse data that was fridge plus other things. And that same model, I watched in a limited eval the next day, basically do 90% on the fridge work. So we went from 60% success rate with fridge-only data, and then we trained that same model again with more data that wasn't fridge, not even increasing the fridge data, and the numbers went up. That's when we were like, holy cow, there's a real benefit to having extremely diverse data. It makes sense because the more we handle bags in other areas will also help us handle the cheese bag better as we move to the fridge. So this is a huge benefit. This means as you pull more diverse data in across different use cases, the robot gets smarter and better. We saw this on Figure 3, we've seen it on Figure 2. This is a massive phenomenon. It's like a miracle for humanoids where you basically have a single platform that can go out and do almost anything a human can. And you're getting better collectively across that fleet as more data pulls in. So instead of having specialized models like an app store, you're basically pulling all this data into one centralized model into training, then you're training a big offline model, and then you're basically OTA updating all the weights on every single robot. So the unfair advantage is like a humanoid that gets good at one thing, say logistics, every robot in the fleet will get better at logistics to the same level of performance.

这与人类不同。你知道，我看着我的孩子们，他们都必须独立学习，而且他们想要独立学习，他们不想被帮助，而且他们都按照自己的节奏。人形机器人就不是这样。整个系统将会有集体学习，通过一个中央大脑集中，在整个机队网络中共享。

Which is unlike humans. You know, I'm watching my kids; they all have to learn independently, and they want to learn independently, they don't want help, and they're all at their own pace. That's not the same for humanoids. There'll be a collective learning through the whole system, centralized through a central brain that gets shared across the whole network of the fleet.

所以我有几个问题，我想在座的人轮流问。Brett，我确信你的时间有限，所以我想给每个人一个机会，尤其是 Gustav 关于手的问题，因为他是我们的手部专家，还有 Devong，欢迎来到 Humanoid Hub。想一个你想问 Brett 的问题。但我先问你两个问题。首先，Helix AI 非常出色。你们最近推出的 S2 层是否带来了更细粒度的角色？你如何看待这个 AI 栈演变成更直观、基于特征的进化？第二，你向我们透露了一些关于 Figure 04 的信息。在代际之间，你认为硬件和 AI 栈哪个进化得更快？SL 芯片是什么？你期望在代际中看到哪些功能和能力的增量跳跃？然后我还有另一个关于中国供应链的问题。

So I have a couple of questions, and I want to go around the panel. Brett, I'm sure your time is limited, so I want to give everybody an opportunity, especially Gustav for the hand because he's our good hand doctor, and Devong, welcome to Humanoid Hub. Think of a question that you want to ask Brett. But I have two for you. First of all, Helix AI is phenomenal. Is the two S2 layer that you guys introduced recently that brings in the more granular role? How do you see that AI stack evolving into a more intuitive, feature-based evolution? And B, you've teased us with a few bits and pieces of information about Figure 04. Between generations, what do you see evolving faster, the hardware or the AI stack? And what's the SL chips? And what's the sort of incremental jump that you're expecting to get in features and capabilities in generations? And then I have another one about the China supply chains.

我想我们昨天刚刚完成了 Figure 4 的关键设计评审。这是我们进入设计锁定并开始出货前的最后一个主要设计关口。我要说的是——我不能谈论 Figure 4 的任何产品发布，但我会告诉你，我们实验室里有一个地方，你可以来看每一个版本，比如 Figure 1，我是说 Scott 你看过，Figure 2，Figure 3，你可以看到从 Figure 1 到 Figure 2 的巨大能力飞跃。与 Figure 2 相比，Figure 1 看起来就像一个宿舍原型。

I think we just yesterday finished our critical design review for Figure 4. So it's our last major design gate as we enter design lock and ship parts out. And I will say this—I can't talk about any product launch for Figure 4, but I'll tell you, we have a spot in our lab you can come to and see every version, say Figure 1, I mean Scott you've seen this, Figure 2, Figure 3, and you can see this huge capability jump from Figure 1 to Figure 2. Figure 1 looks like a dorm room prototype compared to Figure 2.

Figure 2 不错，有点重，然后 Figure 3 变得更苗条，质量更小，传感器更好，手也更好。你看到每一代模型之间都有这样的进步。Figure 4 将是我们历史上版本之间最大的一次飞跃。一年前我不会相信这一点，因为当你发布 Figure 3 时，在最后的设计关口，你会觉得里面有你想要的一切，你会非常高兴。你欣喜若狂。然后，在过去的一年里，随着我们设计 Figure 3，我们走出去，学到了东西，我们也在 Helix 上取得了进展。我们退后一步，能够思考如果我们要从头再来一次，它会是什么样子。我认为 Figure 4 将是这个领域的第一个 iPhone 时刻。

Figure 2 was nice, a little heavy, and then Figure 3 got slimmer, less mass, better sensors, better hands. You saw this step up every single time between models. Figure 4 will be the largest step up we've ever made between versions in history. I wouldn't have believed this a year ago, because when you ship Figure 3, at the last design gate, you think you have everything you'd want in there, and you're so happy about it. You're over the moon. Then, in the last year, as we've designed Figure 3, we've gotten out, we've learned, and we've also progressed on Helix. We've taken a step back and been able to think about if we had to redo this from scratch again, what it would look like. I think Figure 4 will be the first iPhone one moment for the space.

它在各个方面都如此戏剧性地不同。与我们之前的版本相比，它作为机器人已经面目全非。

It is so dramatically different on every aspect. It's unrecognizable as a robot from the previous versions we have.

我认为这是第一个真正的架构版本，将扩展到数百万台机器人，我们学到了很多。如果我要把所有的知识压缩成几句话，我们为 Helix 3 设计了 Figure 4，而 Helix 3 是为数据而设计的。整个机器人都是围绕数据设计的，这是我们迄今为止将机器人大规模推向世界的最大限制因素。我不知道这是否有帮助，但 Figure 3 可以说是世界上最好的，也许是最好的人形机器人，遥遥领先。Figure 4 简直完全超凡脱俗。我们不再向任何人展示。我们基本上什么都不展示。它在这里的一个秘密房间里，整个公司只有少数人真正了解它。但天哪，它太疯狂了。我很兴奋。要让它准备好还有大量的工作要做，但天哪，它将会是……抱歉，我的回答有点冗长。

I think it's the first version of a true architecture that will scale into the millions of robots, and we've learned a lot. If I had to compress all my knowledge into a couple of sentences, we've designed Figure 4 for Helix 3, and Helix 3 was designed for data. The whole robot was designed around data, which is our largest limiting factor by far for getting robots at mass scale to the world. I don't know if that helps, but Figure 3 is arguably one of the best, maybe the best humanoid in the world, by a long shot. Figure 4 is just completely out of this world. We don't show anybody anymore. We don't show basically anything. It's in a secret room here, and only a few folks in the whole company really understand it well. But my god, it is crazy. I'm excited. There's a lot of work left to get it ready to go, but man, it is going to be... Sorry, I'm a little long-winded in my answer.

我们在笑，因为我们刚才在谈论 iPhone。

We're laughing because we were talking about iPhones.

是的，我们刚才就在谈论那个。

Yeah, we were just talking about that.

我们刚才就在谈论。那么，Brett，首先，我必须做一个观察。这些年来我一直远远地看着你，我钦佩你如何处理每一个试图将你与竞争对手（如特斯拉 Optimus）进行比较的问题。我非常钦佩你专注于眼前的事情，而不是贬低竞争对手。我非常尊重这一点和你的态度，我认为这影响到了你的整个团队。话虽如此，美国每个人形机器人公司以及每个非中国人形机器人公司都深深依赖中国的供应链。我很好奇你如何看待整个回流过程，以及如何使自己免受未来可能影响供应链的地缘政治风险的影响，还有你在多大程度上考虑垂直整合，将所有东西都引入内部？

We were just talking. So Brett, first of all, I have to make an observation. I've watched you over the years from afar, and I've admired how you've handled every question that seeks to compare you to the competition, like Tesla Optimus. I have a lot of admiration for the way that you've focused on what's right in front of you and not really spoken down about the competition. So a lot of respect for that and your attitude, and I think that trickles down to your entire team. Having said that, every humanoid bot company in America and every non-Chinese humanoid bot company is deeply reliant on Chinese supply chains. I'm curious how you're thinking about the whole process of onshoring and insulating yourself from geopolitical risks that may impact your supply chains in the future, and how much are you thinking about vertically bringing everything in-house?

所以我想也许对于听众中的朋友们，我们几乎从零开始设计整个机器人的所有东西。我们设计电机、转子定子，我们设计齿轮箱。我们也在现场制造齿轮箱，这太疯狂了。我们制造传感器，我们设计传感器。机器人里有超过 100 个 PCB；我们都在内部完成电气工程方面的工作。相机是设计的；我们从 Figure 3 开始就在这里设计相机。所以所有的手，所有东西，整个堆栈都是巨大的复杂性。我们显然不实际制造每一个零件，但我们几乎设计所有东西。然后其中一些我们实际制造单个零件，然后我们……

So I think maybe for folks in the audience listening, we design almost everything in the whole robot from scratch. We design the motors, the rotor stator, we design the gearboxes. We make gearboxes here on site too, which is crazy. We make sensors, we design sensors. There are over 100 PCBs in the robot; we do all that work internally on the electrical engineering side. Cameras are designed; we design the cameras here from Figure 3. So all the hands, everything, it's a tremendous amount of complexity all the way down the stack. We don't actually manufacture every single part obviously, but we design almost everything. And then some of that we actually manufacture individual parts, and then we...

你必须去中国采购这些零件或其他零件，这让你感到沮丧吗？

Does that frustrate you that you have to go to China for these parts or other parts?

我想我需要知道确切的数字，但我的预测是，也许到下一个季度，我认为我们的供应链方面不会有任何东西来自中国。

I think I need to know the exact numbers, but my forecast is maybe by next quarter, I don't think we'll have anything coming out of China on the supply chain side.

哇。真的吗？

Wow. Really?

在过去的一年里，我们已经将大部分供应链转移到了中国以外，以应对关税、地缘政治风险等等。所以现在我们很少有单个零件的制造来自中国，非常低，低到从这个角度来看真的不再有风险了。

We've moved most of our supply chain outside of China in the last year, for tariffs, geopolitical risks, everything else. So we have very little manufacturing of individual parts coming out of China these days, very low, low enough where there's no risk really anymore from that perspective.

哇。

Wow.

我认为去年世界上每个大公司都在这样做。所以我认为目前真的没有哪个大集团在那里有巨大的敞口。可能有一些，但无论如何，我们已经齐心协力在过去一年里大幅降低了这种风险。

I think every major company in the world has been doing that last year. So I don't think there's really a major group with huge exposure there at this point. There may be some, but anyway, we've made a concerted effort to derisk that significantly in the last year.

好的。我要给 Devong 和 Gustav 一个机会问一两个问题。你想先来吗？

All right. I'm going to give Devong and Gustav a chance to ask a question or two. You want to come first?

当然。嘿，谢谢，恭喜 Brett。演示看起来非常令人印象深刻，速度的提升在这一个上非常明显。我的问题是：人形机器人使用的 AI 模型非常渴望数据，我假设你正在从不同的数据模态中收集有用的数据。那么当你展望一年后，你认为你仍然会缺乏一些重要数据，还是会受到算力的限制？

Sure. Hey, thanks and congrats Brett. The demo looks super impressive, and the speed improvements have been noticeable with this one. My question was: the AI models that humanoids use are very data hungry, and I'm assuming from different data modalities you are working on gathering useful data. So when you look at one year out, do you think you would still be data scarce for some of the important data, or would you be compute constrained?

这是个好问题。有很多限制，对吧？我们需要制造很多机器人；那里有一个限制。我们需要大量的算力；那里有一个限制。我们需要训练和推理算力。然后我们需要大量的数据。我们现在处于这些限制之中，这真是一个绝佳的位置。

It's a good question. There are a lot of constraints, right? We need to make a lot of robots; there's a constraint there. We need a lot of compute; there's a constraint there. We need both training and inference compute. And then we need a lot of data. This is a hell of a place to be in where we have these constraints now.

我们不是在问人形机器人能不能行、能不能做有用的工作、能不能连续运行五小时或八小时甚至更久，这些都是值得思考的好问题。我来回答这个问题：当然，目前我们最大的约束是数据。如果我们有正确的数据，打个响指，我认为通用机器人今天就能解决。我们就能做人类在世界上能做的大部分事情。然后你可能受能源、算力和制造约束，而且是远远地受约束。可能按顺序是数据约束、算力约束、制造约束，这样想比较对。长期来看，你又受算力约束，因为市场上有那么多机器人在你前面跑。

We're not asking like will humanoids work and can they do useful work and can they run for five or eight hours or longer. These are great things to be thinking through. I think I'll answer the question: certainly right now, our biggest constraint is data. If we had the right data, we could snap our fingers and I think general robotics would be solved today. We'd be able to do most things a human can in the world. Then you're probably energy, compute, and manufacturing constrained, by a long shot. Maybe in the orders of data constrained, then compute constrained, then manufacturing constrained. That's the right way to think about this. Long term, you're compute constrained again because there are so many units out in the market that you're running in front with.

所以现在，我们需要解决、我们关心的是，打造一个像人类一样有常识的通用机器。如果我们只想展示我们现在做的，然后把它推出去，让几千、几万甚至几十万台机器人上路，我们能行。我们知道怎么做，只需要去执行。但如果你想解决那种像穿着西装的人一样的体验，也就是像其他人一样，那我们就受数据约束，而且是数量级的约束。我们需要大量正确的数据。这必须在讨论算力约束之前解决。

So right now, what we need to solve, what we care about, is solving for a general purpose machine that has common sense like a human. If we want to just show you what we're doing now and ship that thing, get thousands, tens of thousands, or hundred thousands of robots out, we can do this. We know the playbook; we just got to go execute. But if you want to solve for the experience you feel with a human in a suit, meaning other humans, we're data constrained by orders of magnitude. We just need so much of the right data out there. That needs to be solved before we talk about how we're compute constrained.

在我们内部，我们不受算力约束。我们和英伟达关系很好，他们给我们供货。我们刚推出了一个新的 B200 集群，几周前宣布的，这个月上线了。我们正在训练、预训练一些我们有史以来最大的模型。上周，我们开始预训练一个比以往大好几倍的模型。这些大模型现在需要时间在非常大的集群上训练，对我们来说是非常大的集群。所以现在，我们不是指望打个响指就能得到更多算力。显然，有更多算力会很好，但我们现在有足够的算力来训练为 Helix 设计的模型，这很好。随着时间的推移，能获得更多算力会更好。我们正在计划，而且我们有合适的合作伙伴。所以我觉得我们没问题。我们得去搞定这件事，伙计。我们得去解决通用机器人。我们得打造我们这次通话都希望实现的东西：一个我能和它说话、能做任何事的穿着紧身衣的人。

Internally here, we're not compute constrained. We have a great relationship with Nvidia, and they've supplied us. We just launched a new cluster of B200s that we announced a few weeks back, which went live this month. We're training, pre-training some of the largest models we've ever done. Last week, we started pre-training one of the largest models we've ever trained by several factors. These are large models now that take time to train on very large clusters, very large for us. So right now, we're not hoping we'd snap our fingers and get more compute. Obviously, it'd be great to have more compute, but we have all the compute we need to train the models we're designing for Helix now, which is good. Over time, having access to more compute would be great. We're planning on that, and we have the right partners in place to do this. So I think we're okay. We got to go knock this thing down, man. We got to go solve general robotics. We got to build what we're all on this call hoping will happen: a human in a body suit that I can talk to and just do anything.

是啊。说到这个，我知道 Gustav 有几个关于手部的技术问题想问你。当然，你愿意分享多少就分享多少，但嘿，我们什么时候能自己弄一个？

Yeah. Speaking of which, I know Gustav has a couple of technical questions about the hands for you. As much as you're willing to share, of course, but hey, when can we get one for ourselves?

是啊，我的意思是，有不同型号，对吧？你可以找其他团队，他们第一天就卖机器人给你。也许他们能用，也许不能。那不是我们在这里要做的事。我们要推出的是苹果级品质的产品。我们要推出一个能用的产品，我们不想把一个不能用东西送到你家。我们不想把这个卖给商业客户，结果不能用。这东西必须真正好用。所以我们不想推垃圾。我们不会对这个决定草率。Figure 会努力推出一个令人难以置信的产品。

Yeah, I mean, there are different models, right? You can take other groups that will just sell you robots day one. Maybe they work fine, maybe they don't. That's not really what we're here to do. We're here to ship an Apple-style quality product. We want to ship a product that works, and we don't want to ship something that isn't going to work to your house. We don't want to ship this to commercial customers that doesn't work. This stuff needs to really work well. So we don't want to ship crap. We're not going to be impetuous about this decision. Figure is going to try to ship an unbelievable product.

而且这是一个非常难的领域。

And this is such a hard space to be in.

我得问你，因为我们聊过那个 Figure 在你家和孩子们在一起的演示。我们在 Rise of Scott 和你这里聊过很多次了，那真是令人印象深刻。当锅就在孩子们旁边时，你是什么感觉？显然你觉得足够安全才这么做。我是说，你绝不会冒险，对吧？但它有没有让你看到未来、看到潜力？

I have to ask you because we've talked about that demo of a Figure in your house with your kids. We've talked about it so many times here on the Rise of Scott and yourself, and that was super impressive. What were your feelings when the pot was right there next to your kids? Obviously you felt that it was safe enough to do that. I mean, you wouldn't ever take a risk, right? But did it give you a kind of a view of the future, the potential?

哦，当然。我是说，听着，我每天都在这里。我一周七天都在。我每晚都待到午夜。我经常看到这些机器人。我看到机器人的次数比看到家人还多。我工作很多，你知道。所以这些就像……我不知道。我一直和它们在一起。我甚至想不起上次机器人出严重故障是什么时候。我在办公室和机器人在一起时，没遇到过这种情况。几年前这几乎每天甚至每小时都会发生，现在我们已经到了机器人能正常运作的地步。做测试很好。比如在我家，它运行自主策略，我记得是在把衣服放进洗衣机。它做得很好。非常自豪。但标准在这里，我们在这里。我们需要设计一个能每天全天做日常事情的东西。就像我们现在做的，但在你家、在任何人家里，全天候。那是我们要去的方向。所以我们觉得今天离那还很远。我们还没到。所以我们渴望到达那里，这就是我们刚才说的数据问题。所以感觉很好。我们非常兴奋能让机器人在那里做那些事，并且觉得很幸运能研究这个问题。但我们要解决这个问题。如果我们不解决，会很尴尬。我们准备好了。我们有资金。我们有很棒的团队。我们拼命工作。我们得去解决通用机器。这就是我们在这里每周工作 80 到 100 小时的原因，就是怎么解决那个问题。

Oh yeah, of course. I mean, listen guys, I'm around here every day. I'm here seven days a week. I'm here every night till midnight. I see these bots all the time. I see my bots more than I see my family. I'm at work a lot, you know. So these are just like... I don't know. I'm around them all the time. I can't even imagine the last time we had a bad failure with a robot. I haven't been around a robot at an office where we had this. It used to happen every day or almost every hour years ago, and now we're in a place where the robots live. It's good to do testing. Like it was in my house, running autonomous policy doing, I think in this case, putting laundry in the washing machine. It was doing great. Super proud. But the bar's here, we're here. We need to design something that can do everyday things all day every day. Like stuff we're doing now, but do it in your home, in anybody's home, 24/7. That's where we want to head. So we feel like we're far away from that today. We're not there yet. So we're dying to get there, and that's a data problem we just talked about. So it feels great. We're super pumped to have robots there doing that, and feel very fortunate to work on the problem. But we want to solve this problem. It's going to be embarrassing if we don't. We're ready. We got the cash. We got a great team. We're working our asses off. We got to go solve for a general purpose machine. That's where we're pushing 80 to 100 hour weeks here, just how do we go solve that problem?

我的意思是，这很明显，因为你愿意接下这个挑战，而且你足够自信，愿意把他们的缺点和一切都说出来，这让我非常敬佩，也非常感激。因为整个观众和全球社区都在从中学习很多，我们也需要让人们有所准备，不仅仅是让他们知道，还要让他们在心理上为这个新时代做好准备。所以，我真的无法形容这有多棒、多厉害。现在我把话筒交给 Gustav，他作为手外科医生，肯定有几个技术问题要问。

I mean it's clearly evident because I mean the fact that you were willing to pick up the gauntlet and you were confident enough the fact that you're willing to just laid out their warts and all is just mad respect for that and so appreciative of that because I mean the entire audience and the entire community around the world is learning so much about this and we also need to sensitize people and not only just make them aware but just get them you know mentally ready for this new era. So, I mean it's it's it's I I cannot just tell you how how crazy good and you know awesome this is. So, I'll hand it over to Gustav now who I'm sure has a couple of technical questions being the hand surgeon.

嗯，我不知道是不是技术性的，但我的意思是，你的身体已经在做有用的工作了。为什么还需要重新设计手？这不是已经够好了吗？我是说

Well, I don't know if they're technical but I mean so your body is doing useful work. Why would you need to redesign the hand? Aren't isn't this good enough? I mean

是的。嗯，我们在这里做的手比机器人还多。嗯，你知道，我们一开始，我一开始认为右手应该是基于肌腱的手，和人类非常相似，大部分执行器都在前臂，通过肌腱进行机械驱动。我们为 Figure 1 设计并交付了这款手，它现在在我们的设计工作室里。我当时为它感到非常自豪。这只手太不可思议了，自由度非常高。我们为手腕、所有手指和前臂定制了所有执行器，基本上整个肌腱系统都是我们定制的。我个人在场见证了它的诞生，然后和它相处了大约一个月，最后我终止了这个项目。那是在 2022 年到 2023 年之间。从那以后，我们已经研发了五六代新手，有些你们见过，有些可能没见过。肌腱方案绝对是错误的方向，它是一个局部最优解，从长远来看，对于人形机器人来说不是正确的工程设计。

Yeah. Um we we have done more hands than robots here. Um you know we started as like a t we I I started like thinking that the the right hand was like a tendon based hand. Uh very very similar to um to how humans are today where most of your actuators are actually in your forearm and they're being you know um mechanically driven from like tendons. Uh we designed and shipped that for figure one. It's in our it's in our uh design studio. Um and I I was just so proud of this. The hand was like so incredible. Very high degree of freedom. We customized all the actuators for the wrist and all the fingers and the forearm. We like custom designed all the all like basically the whole tendon system. Uh I personally was there when we brought it up. Uh and I spent about a month with it and I ultimately killed the program. And this was in 2022 2023. And um and since then we've now like worked on five or six new generations of hands that some of you seen, some of you haven't maybe seen. Um tendance is for sure the wrong way to go with hands. It's u it's a local maximum and it uh it it is not the right engineering design long term for humanoids.

我亲身学到了这一点，你知道，这是我犯过的一个工程错误。我的意思是，我们学到了很多，所以我不知道这是不是错误，但嗯，从那以后，我们在过去四年里设计出了越来越好的手。我们大约一个月前预告了一款新手，它的运动范围和关键性能指标基本上达到了人类的水平。这太疯狂了。我不知道，世界上没有任何一只手能接近它，完全没有。在系统的可靠性、运动学、速度和扭矩等关键指标方面，都是如此。我们还在机器人内部配备了一些最先进的触觉传感器。这只机器手简直不可思议，热管理也做得非常出色。嗯，是的，我们稍后显然会详细讨论这些。我认为，要实现机器人的通用性，一只与人类手相媲美的机器手是至关重要的。

And I learned that firsthand and as like you know was an engineering mistake I've made. I mean, we learned a lot, so I don't know if I'm a mistake, but like uh um and since then, we've now designed kind of better and better hands last like four years. We have a new generation hand that we teased about a month ago that has like basically equivalent level of human of human range of motion and uh like KPIs as a human. And it is just it's bananas. I don't know. There's not a single hand in the world that's even close. Not even close. Like in terms of like the the reliability of that system, the the kinematics like um the the KPIs in your own speed and torque. Um we have like like some of the greatest sensors in inside the robot for for like for tactile. Uh the rob the robot hand is just absolutely insane. Thermals are insane in the hand. Uh yeah, and we we'll talk more about this obviously as we as we go on. Um I think to achieve general purposeness in the robot the a human hand a human hand at par a robot hand at parody with a human hand is is uh is critical

而且这似乎也是目前大多数人在努力攻克的地方。我的意思是，你仅通过纯粹的屈伸就能实现很多功能，但在演示中，你有点……

And it seems like that's where most people are struggling at the moment as well and I mean you manage a lot with just having the pure flexion extension but in the demo you kind of the

我的意思是，让我给你举个例子

I mean let me give you like

你觉得你真的需要它吗？

do you really need it you think

我们认为需要。比如，我小时候，我妈妈总是这样把袜子叠在一起，而我们 Figure 3 上的当前手做不到这一点。

We we think you do. Like I think if you you know when I when I grew up my mom my mom would always like fold socks like this inside of each other and um our current hand on figure three can't do that.

它没有足够的自由度。

It doesn't have the degrees of freedom.

嗯，我们新一代的手可以做到，将随 Figure 4 推出。如果我们回到之前讨论的约束条件，如果约束是数据，我们希望从人类数据中大规模学习，而我们有能完成这些操作的人类数据，但机器人却做不到，那么这些数据我们既无法从中学习，也会污染我们的数据集。

Um our new generation hand can that we'll put out with figure 4. And if our goal is to if we go if we go back to like talking about the earlier conversation around like what is the constraint if constraints data and we want to learn from human uh data at scale and we have human data that can do that stuff and we're not able to do that in the robot. This is data that will uh a we're not able to learn from it and b it will pollute our data set.

是的。

Yeah.

所以，我认为要达到类人智能，拥有与人类手相媲美的机器手是基础。

So, I would argue that in order to get to humanlike intelligence, having a hand at par, a robot hand at parody with human hands is fundamental.

是的，我完全同意。我的意思是，我最期待的就是拥有一个手与我们相当的人形机器人。我认为我们正在接近这个目标，但那里，好吧，你可能不想透露，但你是如何解决把所有东西塞进手和散热问题的？是所有东西都得自己发明，还是有现成的产品可以做到？

Yeah, I think she Yeah, I totally agree. I mean, it's it's what I'm looking most forward to is having the the humanoid bot with a hand that is equivalent to ours. And I think we're getting there, but there, okay, you may might not want to tell, but how do you solve for the like packing everything into the hand and the heat? Is it all did you have to kind of make up come up with all yourself or are there products out there that can kind of do that? Today,

我们完全是自己设计这只手的。嗯，我刚开完 Figure 4 的手部评审会，大概有一百页幻灯片，涉及热管理、电气传感、运动学、驱动、结构等等。这是很多很多人长期努力的结果，是我职业生涯中见过的最好的工程工作之一。所以这是累积多年的工作，仅仅为了这只手，很难解释，你知道，我们如何解决散热问题，我们研究了每一毫米、每一克，我们为这只手设定了非常严格的关键性能指标，并且全部达成了。所以，现在我们的机器手拥有的执行器比身体其他部分加起来还多。这是机器人上最复杂的部分，很难在几分钟内解释清楚我们如何设计人形机器人。我昨天刚开了一个十小时的会议，专门讨论这个细节。

We we designed this whole hand and end ourselves. Um, so like I just left the hand review for figure four and it was like maybe a hundred slides on thermals, electrical sensing, kinematics, uh, actuation like uh, structure uh, all of it. And it was just um, many many many people working over a long period of time doing some of the best engineering work I've ever seen in my whole career. So uh this is like cumulatively like years of work that we did just for this hand that is just too hard to explain like in you know how do we sell the thermals here or there it's just it's just uh you know we worked on every millimeter every gram uh we worked towards like very hard requirements in terms of KPIs for this hand and we like we nailed them all. So like uh it just like the robot hand itself now is you know the robot hands for us have more actuators now than the rest of the body. It's just it's uh it's the most complex thing we have on the robot and um you know it's like it's hard to explain how we design a humanoid robot here in like few minutes right it just uh I you know I left a 10-hour meeting yesterday just talking about this in detail

而且你，是的。所以我们只是

and you Yeah. So we only

我们就是这么做的。我们花了四年时间，几十名工程师一直在解决这个问题，最终达到了现在的成果。看起来就像制表工艺，零件非常小，超级紧凑，没有多余空间，而且会发热。里面有很多电机，大量电子布线和线缆，部件在弯曲，试图松动，还有灰尘、碎屑和水等东西试图进入。所以这是一个非常重大的系统工程问题，我们必须攻克。

we do this. It's like uh we spent like dozens of engineers for four years have been working on this problem and it's like uh culminated to like you know this like very it looks like it looks like um it's like watchmaking man. These these parts are small super compact. There's no space. It's very it gets hot. It's um you know there's a lot of motors there. It's just this it's like a lot of electronics wiring. Things are bending trying to come loose. uh stuff trying to get in there like dirt and debris and water and so it's just it's a very very significant system engineering problem that we had to go through.

Brett，你认为你会达到匹配或赶上进化的程度吗，还是你需要，甚至能超越进化？我的意思是，在功能上，你不需要被人类形态所限制，对吧。

Brent, do you think you'll ever reach a point of matching or catching up with evolution or do you need to or can you even you know go past evolution? I mean you don't need to be limited by the human form factor in terms of functionality. Right.

我非常相信与那种“我们会超越人类形态、达到超越人类表现”的观点相反的看法。我个人就是不相信。简单来说：你可以把世界想象成今天有 80 亿人类被扔到这里，然后我们努力围绕它建造了一个世界，以便我们能与之协作。我们有住所、门和可用的工具。我们为我们的生物特性建造了世界。我们没得选身体构造，但我们确实选择了我们建造世界的依据。我们利用我们生物学上的样子，去建造了一个我们能互动的世界。我们生活在物理世界中的人类操作系统里，我们建造那个系统是为了每天都能与之互动。没有比人类更通用的机器了，恰恰就是因为那个原因：我们就是为它而设计的。如果你采取更激进的观点，如果我太矮或太高，或者如果我只有一条胳膊而不是两条，你能做的事情就不如一个正常人类多。这是因为一个正常人类，比如五英尺多高，已经围绕自己设计了世界，以便我们每天都能与之互动。所以没有变得更好的余地。我们就是专门为一个有十根手指、大约 5 英尺 2 到 5 英尺 5 英寸、能处理世界上一切事情的人类设计的。世界就是这样建成的。它针对那个标准进行了完美优化，虽然形态上有些差异，但主要是针对普通人类。如果你太矮，比如两英尺高，你够不到橱柜。如果你太高，比如 20 英尺，显然也无法很好地适应世界。所以我们已经把整个世界都优化得只适合人类。它完美地适合人形机器人，就像一个机械人类。你可以把它想象成一把完美匹配锁孔的钥匙。

I'm a big believer in the opposite of the school of thought that says we'll just go beyond the human form and get to beyond human performance. I just don't believe that personally. A simple thing is: you could think of the world as eight billion humans dropped here today, and then we worked to build a world around it so we can work with it. We get shelter, doors, and tools we can use. We built the world for our biology. We didn't get to choose our body composition, but we did get to choose what we built the world of. We used the way we look biologically to build a world we can interact with. We live in a human operating system in the physical world, and we built that so we can interact with it every single day. There is no greater general-purpose machine than a human, just because of that very reason: we've designed it just for us. If you take a more aggressive view, if I was too short or too tall, or if I had one arm instead of two, you're not able to do as many things as a normal human can. It's because a normal human, you know, 5-foot-something, has designed the world around themselves so we can interact with it every day. So there is no getting better. We've designed it specifically for a ten-fingered human that's roughly 5'2" to 5'5" that can do everything in the world. That's the way the world was built. It was optimized perfectly for that, across a little bit different form factors here and there, but mostly an average human. If you're too short, two feet tall, you can't reach the cupboards or cabinets. If you're too tall, like 20 feet, you obviously won't work well in the world either. So we've optimized the whole world just for humans. It's perfectly built for a humanoid, like a mechanical human. It's a perfect key to a keyhole, you could think of.

对吧？

Right?

这就是为什么人形机器人能创造出神奇的东西，因为你可以用一个 AI 模型构建一个硬件，它能像海绵一样吸收所有数据。

And that's why a humanoid will build something magical, because you can build one hardware with one AI model that can take all the data in like a sponge.

嗯。

Yeah.

随着更多数据进入那块海绵，它会变得更好。

It can get better as more data comes into that sponge.

嗯。

Yeah.

而且硬件不需要为此改变。这就是为什么人形……

And the hardware doesn't need to change for that. And that's why human...

Brett，我能插一句吗？因为这其实和我的问题相关。我其实想和你聊聊你们几天前发布的那个整理卧室的视频，两个机器人协作的那个。其实 Scott 和我几天前就那个视频做过一期节目，我觉得最迷人的地方在于，很明显，通过点头和视觉交流，而不只是 Wi-Fi 心灵感应式的通信，你们是在为智能体构建一个世界，而不是为机器人。就像机器人必须与人类互动，它们必须彼此互动，它们可能还得与其他形式的智能互动，比如大型语言模型之类的。系统 2 思维，你们在做的那种，跳出框架的东西。所以我很好奇你怎么看，因为我觉得那真的……对，就是这样。我觉得特别迷人。尤其是点头的动作，我觉得非常非常了不起。然后，这跟这个有点关系，但我很好奇：你们这个场景里光线很完美。你们会不会故意把光线调得不完美，让这些机器人不得不在更自然的环境里工作，比如光线忽明忽暗之类的？

Brett, can I jump in on this? Because this was actually related to my question. I actually wanted to talk to you about the bedroom-making thing that you guys released a couple of days ago, where the two bots are working together. That is actually something Scott and I did a video on a couple of days ago, and I thought the fascinating part about this was it's clear that with the head nods and the visual communication, not just like Wi-Fi telepathic communication, you're kind of building out a world for intelligences, not for the robot. It's like the robots have to interact with humans, they have to interact with each other, they may have to interact with other forms of intelligence like large language models or something. System two thinking, the stuff that you're doing, you know, kind of outside the box. So I'm curious what your thought about that is, because I find it really... Yeah, there we go. I find it really fascinating. The head nods in particular, I thought were really, really remarkable. And then, this is related to this in a slight way, but I was curious: you've got perfect lighting in this scene. Do you guys mess around with the lighting some to try to get imperfect lighting so that these guys have to work in more natural environments where you might have slashing lights and things like that?

关于卧室视频，有两点。我认为人形机器人的一点是，你基本上是在用智能来创造产出和经济价值。在这个案例里，是打扫房间、整理床铺。在今天直播的案例里，是搬运包裹。但你可以把人形机器人看作一个基本上就是把智能转化为有用产出的机器人。而这是我每天在家里都想要的东西。是的，如果光照变化很大，我们确实看到了策略退化。这就是数据……也许有办法可以通过训练来克服。还有其他技术可以用来改变训练集，加入不同的光照条件，做一些调整。这也只是另一个答案：我们只是需要在系统里加入更多数据，比如世界各地不同光照条件的地方等等。所以，是的，我觉得点头很酷。在这个案例里，他们需要在某个时刻拉紧床单，这样一个人拉的时候另一个人还没准备好。他们用点头作为一种交流的手势，这真的很酷，作为一种彼此沟通的方式。我们认为……

Two things about the bedroom. I think one thing about a humanoid is that you're basically using intelligence to create output and economic value. In this case, it's cleaning up a room, making a bed. In the case of the live stream today, it's moving packages around. But you can think of the humanoid as basically a robot that is just turning intelligence into useful output. And this is something I want in my house every single day. Yeah, if lighting does change significantly, we've seen policy degradation for sure. This is where data is just... maybe there's a way you can train through it. There are other techniques you can use to change the training sets to have different lighting conditions and do some alterations there. It's also just another answer: we just need more data in the system, like places with varying lighting conditions across the world and all this. So yeah, I think the head nod was cool. In this case, they needed to tension the bed at a certain time so one's not pulling while the robot's not ready. And they use a head nod as a gesture of communication that's happening, which is really cool, as a way to kind of communicate with each other. We think...

但它们需要吗？我的意思是，在我看来，点头更像是给人类看的，作为一种信号，而不是因为它们真的需要给出物理上的点头或任何指示，因为它们彼此是连接的，对吧？就像蜂巢思维一样。

But do they need to? I mean, it kind of looks to me as if the head nod is more for humans, as a signal, rather than because they don't really need to give a physical nod or any indication because they're connected to each other, right? It's like a hive mind.

嗯，在这个案例里，机器人之间没有明确的通信。它们完全通过视觉和点头来协调动作。所以它们就是这样交流的。还有其他方式可以做到这一点，但这就是机器人在这里表现的方式。

Well, in this case, there's no explicit messaging between the robots. They're coordinating their actions fully visually with the head nods. So that's how they're communicating. There are other ways to do this, but that's how the robots are performing here.

但是，是的，我不知道。我觉得看到机器人能用 AI 策略做这样的事情很酷，而且能展示多个机器人彼此互动也很酷。我们现在就在直播里做这件事。我们会让机器人在轮班期间轮流工作，彼此交流，互相发送消息，这非常酷。

But yeah, I don't know. I think it's cool to see robots be able to do things like this with an AI policy, and it's cool to be able to show multiple robots engaging with each other. And we're doing this on live stream right now. We'll have robots taking turns throughout the shifts, communicating with each other, and messaging each other, which is very cool.

那么，我猜这也是为了能和人类交流，因为我知道你想卖给我两个。我其实明白你想卖给我三个机器人。但我只买得起一个。

Now, I assume it's also so it can communicate with a human, because I know you want to sell me two. I understand actually you want to sell me three bots. But I can only afford one.

所以，能看到对面有个真人会很有趣，看看他们是怎么做到的，因为我猜你这样就够了。现在它能读懂我的想法了。

So, it'd be interesting to actually see a human on the other side to see how they're able to, because I assume you that's just enough. So, it'll read my head now.

是的。我们现在花了很多时间在这上面，几乎就像如何与人类进行有效的语音到语音交流。实际上，今天的 Figure 机器人正在使用我们设计的 HARK 语音模型来实现这一点，这真的很酷。所以如果你进来和这里的机器人说话，它们用的就是 HARK 语音模型。

Yeah. We spend a lot of time on it right now, almost like how to do effective speech-to-speech with humans. Actually, the Figure robots today are using the HARK voice model that we designed to do this, which is really cool. So if you come in here and talk to the robots here, they're using the HARK voice model.

酷。

Cool.

然后你看，这就是机器人。天哪，那东西……

And then here you go. There's the robot. Man, that thing is...

第一个回来了。

First one is back.

所以关于这个客户我有两个问题。第一，你说成功率是 90%。所以那是你必须达到的，而且我猜那是客户设定的。

So I've got two questions about this customer. One, you said like 90% is the rate. So that's what you have to do, and that's something I assume the customer set.

是的，我们基本上需要达到大约 90% 的条码扫描成功率。我们在内部跟踪这个数据。这不在直播流上，比如现在处理的 5759 个包裹中，有多少比例被扫描了。我们在内部有数据，大概是 90% 多。我们必须达到 90% 的成功率，而且每个包裹大约需要三秒。这些是大概的关键绩效指标。

Yeah, we basically need to do roughly 90% success rate on barcode scans. We track this internally. It's not on the live stream, like out of the 5,759 or whatever packages it is now, what percentage of those got scanned. We have it internally, it's like 90-something percent. We have to do 90% success rate there, and we have to do roughly three seconds a package. Those are the rough KPIs.

另一件事是传送带上没有护栏。那是因为客户就是这么设置的吗？

And the other thing is there's no rails on the conveyor. Is that because that's the customer setup?

这是客户的设置。

This is the customer setup.

哦，好的。因为我们想知道为什么客户不把护栏放上去，只是为了防止包裹掉下来？

Oh, okay. Because we're wondering why isn't the customer putting rails on there just to keep packages from falling off?

我们正在努力复现我们在真实站点看到的确切用例。所以一一对应，在这里设计不像客户站点的系统对我们没有帮助。所以我们做的所有事情，都尽量让它尽可能接近。而在这一点上，关键就是如何让机器人达到无故障正常运行的状态。

We are trying to reproduce the exact use cases we see at real sites. So one-for-one, it doesn't help us designing systems here that don't look like the customer sites. So all of our stuff we do, we try to get it as close as possible. And at this point, it's just like how do we get the robot to a point where it's running nominally without any failures.

对。关于 Vulcan 去维修的最后一点——你为什么不叫它……我有两个名字，你必须用其中一个。我想你可以叫它 Bots，或者叫它 Bash，Bash 是机器人疾病专科医院。

Right. And the final comment about Vulcan when it goes off to repairs—why don't you call it... I've got two names that you have to use one of them. I think you can either call it the Bots, or you can call it Bash, where Bash is the bot ailment specialty hospital.

天哪。就像领导这个项目的 Morris 命名了它……

My god. Like Morris, who led the project, named it...

不，Vulcan 没问题，但它去的地方——我的意思是实际地点应该叫 Bots。

No, Vulcan is fine, but where it goes—I mean the actual location should be off to the Bots.

是的，Bots 可能不错。我会把它放到那边的点子罐里。

Yeah, Bots might be good. I'll put it in the idea jar over here.

好的。那个和 Bash——就这两个。

Okay. That and Bash—those are the two.

你现在设计新手时，担心数据迁移吗？你有旧版手的所有数据。我想他们是为整个机器人考虑的，但专注于手，你觉得直接迁移会是个大问题吗？你必须获得一个全新的数据集来练习，还是你觉得用升级版的手继续会很容易？

Were you worried about the data transfer when you're designing a new hand now and you have all this data with the older version of the hand? And I guess they go for the whole bot, but focusing on hand, do you think it would be a big problem to kind of just transfer it? Do you have to get a whole new data set to practice with, or do you think it would be quite easy to just go on with a higher-off hand?

我实际上认为会容易得多，因为我们试图大规模地从人类数据中学习，而我们当前的手没有与人类相同的运动学映射。手指弯曲的方式和手指中连杆的设置方式现在并不理想。所以我实际上认为，从我们在内部所做的工作来看，几乎在每个指标上，新手都会更好。比如，在可靠性方面，即使它会非常像人类，我在内部见过它经受打击——我们制造了它们——可靠性简直疯狂。从 AI 学习的角度来看，我认为这将是我们做过的最好的决定之一。但如果在四年前尝试这个，我不认为我们能成功。我们让顶尖的电机专家参与，我们让顶尖的传感器专家参与。过去一年里，组织里的顶尖人才都参与其中，才达到现在的水平。这就像建造涡轮喷气发动机——这是一个复杂的系统，很难。未来还会有更多困难的事情来让这个东西工作。但我实际上认为我们在数据方面会没问题。而且我们确实计划将 Figure 4 扩展到前所未有的水平，甚至超过 Figure 3。所以从这个角度来看，它将拥有更多的数据，无论是在机器人方面，而且我认为它将能够从 Helix 和训练中做得更好。

I actually think it'll be much easier because we're trying to learn from human data at scale, and our current hand doesn't have the same kinematic mapping as a human. The way the fingers bend and the way the linkages are set up in the fingers now are not ideal for this. So I actually think, from the work we've done internally, on almost every metric the hand's gonna be better. Like, in terms of reliability, even though it'll be crazy humanlike, I've seen it take a beating internally—we've built them—and the reliability is just crazy. And from an AI learning perspective, I think it'll be one of the best decisions we've ever done. But going to this from, like, four years ago, if we tried this, I don't think we'd be able to be successful. We put our top motor specialists on it, we took our top sensor person on it. It's had the top people in the organization come through there over the last year to get where we're at now. It's like building a turbojet engine—it's a complex system and it's hard. There'll be a lot more hard stuff in the future to get this thing to work. But I actually think we'll be okay on the data side. And we do plan to scale Figure 4 to unprecedented levels versus even Figure 3. So from that perspective, it will have a mountain more of data, both on the robot side and even better, I think it'll be able to do things better coming from Helix and from training.

那没问题。我们会把它们带上来。我们会带上一堆。我们会去试试，然后就会知道。但我今年就会知道，但我很有信心它会很棒。

That's okay. We'll bring them up. We'll bring up a bunch of them. We'll go try them, and we'll know. But I'll know this year, but I feel pretty confident it's going to be great.

嘿，Brett，你很快会给 Figure 添加语音输出吗？然后我真正想问所有机器人专家的问题是：我们什么时候能达到这样的程度——把全新的 Figure 带到一个它从未见过的工厂工作站？没有人类展示过这个新工作站，人类向机器人描述需要做什么，甚至可能快速演示一下，就像你训练人类那样。你觉得我们什么时候能达到那个理想状态？

Hey Brett, are you going to be adding speech output to Figure anytime soon? And then my real question that I've always wanted to ask all roboticists is: when are we going to get to the point of taking a brand new Figure to, let's say, a factory workstation that it hasn't seen before? No human has shown anything about this new factory workstation, and the human describes to the robot what needs to be done, and maybe even demos something quickly, like you would a human, like how you would train a human. When do you think we'll get to that nirvana?

哦，是的，我回答完这个问题就得走了——我在这里有个会议，我会再联系你。但我想说几件事：第一，我们正在使用——我去年夏天成立了一个新的 AI 实验室叫 HARK。我们有一个大约 70 人的团队，正在研究新型 AI、下一代 AI 模型和一些硬件。我们花了很多时间的一个领域是实时语音到语音。所以机器人——现在 Figure 的每个机器人都搭载了 HARK 语音模型，你可以在办公室拦住任何机器人跟它说话，这很棒。然后关于你的第二个问题，这是机器人技术的圣杯，兄弟。我们希望能够把机器人带到任何地方，做任何我们想做的事。这是 Figure 的头号目标：我们如何解决通用机器人的问题？我们最早能在今年演示这样的东西，但我不确定我们能否实现。

Oh yeah, so I have to leave after this question—I just have a meeting here, I'll get back to you. But I think a few things: one is we're using—I started a new AI lab called HARK last summer. We have a team of about 70, working on new kinds of AI, next-generation AI models, and some hardware. One of the areas we're spending a lot of time in is real-time speech-to-speech. So the robots—every robot at Figure now has the HARK voice model on it, and you're able to stop any robot and talk to it here at the office, which is great. And then on your second question, this is the holy grail of robotics, man. We want to be able to take a robot anywhere and do whatever we want with it. This is the number one goal at Figure: how do we solve for a general-purpose machine and solve general robotics? The soonest we'll be able to demonstrate something like this would be this year, and I don't know if we'll hit it.

那太疯狂了。

That is crazy.

也许吧，也许不会。

Maybe, maybe not.

我们内部相信，我们有一个计划，如果执行得当，就能做到类似的事情。现在，可能这个计划很糟糕，我们做了却行不通。也可能我们执行这个计划，它成功了，但只是在很小的规模上成功。随着进展我们会了解更多。上周，我们刚刚在我们内部搭建的新 B200 集群上，启动了下一代 Helix 3 模型的预训练运行，用来测试这个计划，这个集群本月刚上线。

We believe internally we have a plan that if we execute on it, we can do something like that. Now it could be that the plan is crap and we do this and it doesn't work. Or it could be we do this plan and it works, and it just works at such a small scale. We'll know more as we go. We just launched the pre-training run for our next generation Helix 3 model to test this last week on the new cluster of B200s we just set up here internally that launched this month.

所以我觉得我们会在未来几周评估那个模型。

So I think we'll evaluate that model in the coming weeks.

我认为到那时它还不能做这样的事，但这是我们觉得应该能够做到这类事情的模型架构。所以希望在未来几年内，我们肯定希望它能做到。现在还很模糊，因为从来没有人达到过能做这件事的程度。所以我们有一个计划，我们认为如果执行得好，就有最大的机会完成这件事。如果你看看这个计划，在公司内部听到介绍，你会觉得，天哪，这个计划应该能解决这个问题。

I don't think at that point it'll be able to do things like this, but it is the model architecture that we think should be able to do things like this. So hopefully over the next coming years, for sure, we hope that you'd be able to do this. It's just foggy right now because nobody's ever gotten to a point to be able to do this. So we have a plan now that we think if we execute well on, we'll have the best chance possible to get this done. If you look at the plan and get pitched internally, you're like, damn, this plan should be able to solve this.

好的。

Okay.

但这需要海量数据，需要海量算力，我们还需要耐心去做正确的消融实验、正确的模型策略评估等等。但这是我们的头号目标。如果你进来问，Brett，我们在做什么？我会说，我们在努力解决通用机器人技术。做你刚才说的事情，是清单上的首要任务。可能是在工厂里，可能是在你家里，能够整理床铺，但是在一个从未见过的家里，比如我们当天早上才订的随机 Airbnb，我们把机器人放进去，让它去做某件事，它就能做到。这些就是我们想解决的评估场景。我们正在尽可能快地解决。希望在未来一到两到三年内，我们能实现这个目标。如果我们今年能解决，我们正在尽最大努力尽快解决。但一旦我们解决了，我会告诉你。

But it needs a ton of data. It needs a ton of compute, and we need to see patience to do the right ablations in the right model policy evaluations and stuff. But this is our number one goal. If you came in here and said, Brett, what are we doing? I'm like, we're trying to solve general robotics. Doing things like you said is at the top of the list there to be able to do. It could be in a factory, it could be in your home, could be able to do that bed tidying, but in an unseen home like a random Airbnb we just booked that morning and we drop the robot into and we ask it to go do something and it does it. These are the kind of evaluation sets we want to be able to solve. We're trying to solve as fast as possible. Hopefully in the next one to two to three years we can be able to do this. If we can solve it this year, we're working as hard as possible to try to solve it as fast as possible. But as soon as we solve it, I will tell you.

即使明年发生，我们也会震惊不已。

Even if it happens next year, our minds will be blown.

是的。我觉得如果这十年内能实现，那将是世界前所未见的最伟大的事情。

Yeah. I think if it happens within this decade, I think this will be the greatest thing the world's ever seen.

好的。那么，请再给我们几分钟，因为我还有两个问题，我相信你会喜欢的。第一个是关于长期来看，在月球和火星上进行机器人劳动力部署。你们在考虑吗？这是你们计划的一部分吗？让这些机器人适应地外部署，工程挑战有多大？还有周边产品。

Okay. So, please, if you can just spare a couple more minutes, because I have some two questions which I'm sure you'll love. First being long-term horizon off-world deployment of bots for labor on the moon on Mars. Are you thinking about it? Is that part of your plan? How big a challenge is it to adapt these bots to off-world deployments? And the swag.

绝对喜欢周边产品。

Absolutely love the swag.

我们想要 Figure Three 的周边。是的，我们想要 Figure 的，不管是什么。如果你还有时间回答最后一个问题，来自 Dang 的一个快速问题。那太好了。那么，先说说太空部署吧。

We want the Figure Three merch. Yeah, we want the Figure that's whatever that is. And if you have time for one last question, one quick last question from Dang, please. That would be wonderful. So, space deployment, please. First.

好的。嗯，目前机器人 AI 中存在预训练与微调的权衡。如果预训练数据太多或太少，都得不到泛化能力。但之后你又花太多时间在微调上。那么现在这个平衡是怎样的？公司内部有没有目标，让微调尽可能少？

Okay. Um, so there is a pre-training versus fine-tuning trade-off in robotics AI right now. If you train pre-training with too much data or too little data, you don't get generalization. But then you spend too much time on fine-tuning. So what does that balance look like right now? And is there a goal within the company to get fine-tuning as less as possible?

在太空方面，在我们进入预训练问题之前。是的，我们很想去太空。这是我们从一开始就有的总体规划的一部分。我们现在正在进行一些相关对话。我希望我们的机器人肯定会进入太空，希望越快越好。关于训练的第二个问题，我得走了，抱歉，我 2:30 有个会。我们把所有时间都花在预训练上。让机器人获得真正泛化能力的方式就在预训练中。那是在这个阶段我们学习世界的语义和这种理解。所以，我想说，请耐心等待，伙计。我们正在努力。每天我都会看到新的实验。有些成功，有些失败。我们正在攻克。我们想要你想要的。我们想要一个人类或机器人可以做任何事的世界。Figure 想要打造电影《我，机器人》那样的世界。

On the space side, before we move into the pre-training question. Yeah, we would love to be in space. It's a part of our master plan from when we started. And we're having some of those conversations now. I would like our robots will most certainly be in space, hopefully as soon as possible. On the second question on training, and I have to leave, I'm sorry, I just have a 2:30. We are spending all of our time on pre-training. The way we get to real generalization from the robot is in pre-training. That's the stage that we learn the semantics of the world and this understanding. So yeah, I would say just hang tight, man. We're working on it. Every day I see new experiments. Some work, some don't. And we're working through it. We want what you want. We want a world where a human or robot can just do everything. Figure wants to build iRobot the movie.

是的。

Yeah.

在现实生活中。

In real life.

我们在演示中看到了。

We saw it in the demo.

是的。机器人到处走动。它们在室内，在室外，在家里，在劳动力队伍中，数以十亿计。我们想实现那个。为了在 2026 年实现，Figure 需要解决泛化问题。就是 Phil 谈到的那些，我们在这里谈到的那些。这是绝对的头号目标。不是制造业，不是炸弹成本，不是供应链，不是任何这些东西。是泛化。这是每个人都应该思考的问题。问题是我们如何解决，而我认为 Figure 会第一个做到。我们现在非常努力。

Yeah. Robots are walking around everywhere. They're inside. They're outside. They're at home. They're in the workforce. They're in the billions. We want to do that. In order to do that in 2026, Figure needs to solve generalization. Stuff that Phil talks about, stuff we talking about here. That's the number one goal by a mile. It's not manufacturing. It's not bomb cost, it's not supply chain, it's not any of these things. It's generalization. That's what everybody should be figuring out. The question is how do we solve this, and Figure I think will be the first to get there. We're working really hard right now.

但听着，各位，我得走了。我有个会。谢谢你们邀请我。

But listen guys, I got to run. I got a meeting. Thanks for having me.

谢谢。

Thank you.

也总是很愉快。

And it's always a pleasure.

谢谢。感谢你的时间，恭喜。太棒了。

Thank you. Thank you for your time and congratulations. It's been awesome.

希望你能再来。再见，各位。拜拜。

Hope to have you back. See you guys. Bye.

拜拜。

Bye.