机器人演示与现实价值之间的差距

好了，各位。欢迎收听 TwiML AI 播客的另一期节目。我是主持人 Sam Cherington。今天，我们邀请到了 Nikita Rudin。Nikita 是 Flexion Robotics 的联合创始人兼 CEO。在开始之前，请务必花点时间点击订阅按钮，无论你在哪里收听今天的节目。Nikita，欢迎来到播客。

All right, everyone. Welcome to another episode of the TwiML AI podcast. I am your host, Sam Cherington. Today, I'm joined by Nikita Rudin. Nikita is co-founder and CEO of Flexion Robotics. Before we get going, be sure to take a moment to hit that subscribe button wherever you're listening to today's show. Nikita, welcome to the podcast.

谢谢。非常高兴来到这里。

Thank you. Very excited to be here.

我很高兴你能来参加节目，也期待深入探讨我们今天的主题——也就是当前机器人技术的现状与实现技术愿景所需达到的目标之间的差距。你在这个领域已经工作了相当长一段时间。你在苏黎世联邦理工学院获得了博士学位，并在英伟达工作过一段时间。不如分享一下你的博士研究及其重点？

I'm excited to have you on the show and I'm looking forward to digging into our topic for the conversation, which is really digging into the gap between where we are today with robotics and where we need to be to fulfill the vision of the technology. You've been working in this space for quite a while. You did your PhD at ETH Zurich and spent some time at Nvidia. Why don't you share a little bit about your PhD and the focus of your research?

我刚开始时，我们尝试使用仿真结合强化学习来教腿足机器人做一些非常简单的事情，比如在平地上行走。当机器人能走几步时，那已经是巨大的成功了。核心重点是减少实现这一目标所需的训练时间。

So when I started, we were trying to use simulation with reinforcement learning to teach legged robots very simple things like just walking on flat ground. And when the robot could take a few steps, that was already a big success. The core focus was to reduce the training time needed to achieve that.

你说腿足，是指四足机器人吗？

And when you say legged, like a quadruped?

没错。就像四足机器狗。我们用的不是波士顿动力的 Spot，而是 ANYmal。ANYmal 是一家瑞士初创公司，从我们实验室孵化出来的。它和 Spot 非常相似，不过是红色的，瑞士制造。是的，我们当时确实在努力减少实现目标所需的训练时间。在我开始之前，针对这种四足机器人的强化学习已经有了一些成果，但需要数周的计算才能取得任何进展。利用 GPU 和大规模并行仿真器，我们成功将时间缩短到了几分钟。实际上，我们在某个会议上有过一次现场演示：我抱着机器人，笔记本电脑实时运行训练，每 15 秒就会把最新的策略发送给机器人，你可以亲眼看到它从摔倒到迈出第一步，然后三四分钟后就能在舞台上走动了。这是一个非常酷的视觉演示，让每个人都能看到学习过程是如何发生的。从那以后，我的博士研究就是利用类似技术提升机器人的敏捷性。仍然是训练仿真中的神经网络，然后迁移到现实世界，但输入变得更复杂，任务也更复杂。到最后，我们可以去瑞士的一个搜救设施。想象一下倒塌的建筑、大量的泥浆、苔藓、碎石、大石头，这些地形即使对人类来说也很难穿越。我们只需告诉机器人从 A 点走到 B 点，它就会利用整个身体——用膝盖爬上大石头，跳过缝隙——完全自主、端到端地完成，依靠图像和机器人自身状态来规划下一步动作。

Exactly. Like a quadruped legged robot dog. We were not using Boston Dynamics' Spot. We were using ANYmal. ANYmal is a Swiss startup that was a spin-off from our lab. It's very similar to Spot, but it's red and made in Switzerland. Yeah, we were really trying to reduce the training time needed to achieve that. So before I started, there were some results of reinforcement learning for such quadrupeds, but it would take weeks of computation to achieve anything. Using GPUs and massively parallel simulators, we managed to reduce that to just a few minutes. So actually we had a demo on stage at some conference where we were running training live on a laptop while I was holding the robot, and every 15 seconds the laptop would send the latest policy to the robot, and you could literally see how it went from just falling over to taking a first step, and then after three or four minutes it would be able to walk around the stage. That was a pretty cool visual demo for everyone to see exactly how the learning process happens. And from there, my PhD was pushing the agility of that robot using similar techniques. So it was still training neural networks in simulation and then transferring them to the real world, but the inputs got more complicated, the tasks got more complicated. So by the end, we could go to a search and rescue facility here in Switzerland. So you have to imagine collapsed buildings, a lot of mud, moss, gravel, big rocks, terrain that is very hard to navigate even for a human. And we would just tell the robot to go from point A to point B, and it would use its whole body. So it would use the knees to climb on top of big rocks and then jump over gaps, all autonomously, all end to end, using images and the state of the robot to plan its next actions.

在讲述这个在搜救场景中部署机器人的故事时，我想象着那些我见过的类似演示：机器狗在行走，也许在开门，虽然你的演示可能不包括开门，但我见过类似的演示，机器狗爬山、穿越碎石。我认为这些演示在很多情况下试图传达一个观点：“嘿，任务完成，我们搞定了。”请谈谈你通过那个演示所取得的成果，与在真实搜救场景中部署这种机器狗还需要做些什么之间的差距。

In telling that story about deploying this robot in a search and rescue context, envisioning the demo I've seen similar things: the robot dog is going, maybe opening some doors, maybe that wasn't part of your demo, but I've seen similar demos of the robot dog climbing hills and crossing rubble. I think those demos attempt in many cases to land the idea that 'hey, flag in the ground, we're done here.' Talk a little bit about the distance between what you were able to accomplish with that demo and what you think needs to be done to deploy one of these robot dogs in a real search and rescue scenario, for example.

这是个好问题，我和英伟达及苏黎世联邦理工学院的同事们争论过很多次。核心问题是：运动能力是否已经解决了？五年前我们就争论过，当时机器人只能在平地上勉强行走，有人说“运动能力已经解决了，我们不需要再关注它了。”我的观点一直是：不，直到机器人能真正到达人类能去的任何地方，而且你根本不需要考虑它能不能做到或有多可靠，那时我才会认为运动能力解决了。所以目前的情况是，任何“盲”操作都可以非常鲁棒。“盲”意味着机器人不感知地形，只是被动反应。这让训练变得容易得多，因为你只需向它扔各种东西，它总是试图保持稳定。尤其是对于四足机器人，保持稳定非常容易。例如，你甚至可以让它上楼梯：它会撞到第一级台阶，意识到有东西，然后爬上去。有了感知，你希望它真正做出反应。你不希望它撞到第一级台阶，而是希望它更小心地放置脚。这更难了。今天，在 2025 年底，我认为我们可以做到：我们可以有相当好的策略，基于感知输入来规划动作序列，因为我们在仿真中训练一切。这意味着我们需要更仔细地模拟这些传感器，投入大量精力创建能被摄像头看到的复杂地形，以及各种噪声模型来模拟这些图像的干扰和缺陷。

It's a great question, and I had this debate so many times even with my colleagues at NVIDIA and ETH. The general question is: is locomotion solved or not? And we already had this debate five years ago when the robot could just barely walk on flat ground, and people were saying 'yeah, locomotion is solved, we don't need to focus on it anymore.' My take was always: no, until the robot can really go anywhere a human can go, and you don't even need to think about whether it can do it or how reliable it is, then my take is that locomotion is not solved. So where we are today is that anything that is blind can be very, very robust. Blind meaning that the robot does not perceive its terrain, so it's just reacting. That makes the training much easier because you just have to throw a lot of things at it; it always tries to be stable. Especially for a quadruped, it's very easy to remain stable. For example, you can even walk upstairs: it will just hit the first step, realize there is something, and then climb up doing it. With perception, now you want it to actually react. You don't want it to hit the first step; you want it to place the feet much more carefully. That's harder. Today, at the end of 2025, I'd say we can do that: we can have fairly good policies that plan the sequence of actions based on perceptive inputs, because we're training everything in simulation. This means that we need to be much more careful in how we simulate those sensors as well, to put a lot of effort into creating complicated terrains that can be seen by the cameras, and also all sorts of noise models to simulate all the disturbances and defects of those images.

所以，我听到你说的是，添加那些你认为会有帮助的额外信息实际上让事情变得更困难，因为它引入了大量噪声。而之前机器人可以跌跌撞撞地穿过地形，现在它试图结合视觉输入来规划，但结果……当它尝试这样做时，实际发生了什么？是卡顿，还是根本不起作用？从模型角度来看，只是难以训练吗？到底发生了什么？

So what I'm hearing you say is that the addition of the additional information that you would think would help actually makes it more difficult because it introduces a lot of noise. Whereas previously the robot could kind of stumble through the terrain, now it's trying to incorporate this visual input to plan, but it ends up... What actually happens when it is trying to do this? Is it stuttering, or does it just not work? Is it just hard to train from a model perspective? Like what happens?

我的意思是，最终，如果你把所有事情都做对了，最终的行为会更好。但这是一个很大的“如果”。我们通常称之为“仿真到现实差距”。如果你在仿真中训练并在现实中部署，情况是不一样的。在仿真中效果很好的东西在现实中可能完全不起作用。一旦感知加入循环，这个仿真到现实差距会大得多。

So I mean, in the end, what happens is that the final behavior is better if you do everything right. But that's a big if. The typical thing we refer to as the so-called sim-to-real gap. If you're training things in simulation and deploy them in real life, it's not the same. And things that work really well in simulation might not work at all in real life. That sim-to-real gap is much larger once you have perception in the loop.

我对这个问题的尖锐看法是——我很乐意被证明是错的——但我认为今天没有一个类人机器人真正创造了价值。意思是，可能有一个机器人在工厂或仓库里做了相当接近它应该做的事情，但并不是精确的任务。所以最终，它没有创造价值，因为它没有做它应该做的实际事情。

My hot take on that and I'll be happy to be proven wrong, but I think there is not a single humanoid robot today that actually generates value. Meaning there might be a robot that does something fairly close to what it's supposed to do in a factory or in a warehouse, but it's not the exact task. So in the end, it's not generating value because it's not doing the actual thing it's supposed to do.

一旦你开始模拟深度图像或 RGB 图像，情况就更糟了。这让研究人员或工程师的工作变得更难。你需要跨越机器人的物理和感知输入两方面的仿真到现实差距。你必须仔细模拟传感器。但如果做对了，最终的行为实际上会好得多，因为你可以看到机器人不仅仅是对脚下发生的情况做出反应，而是提前进行相应的规划。那么，对于四足机器人来说，结合视觉的 locomotion 问题是否已经解决了？还是仍然存在一些悬而未决的问题，比如泛化差距？你怎么看？

Once you're simulating either depth images or RGB images is even worse. So that makes the job of the researcher or the engineer harder. You need to cross that sim-to-real gap for both the physics of the robot and the perceptive inputs. You have to simulate the sensors carefully. But if you do it right, then the final behavior is actually much better because you can see that the robot is not simply reacting to whatever is happening under its feet but actually planning accordingly in advance. Is this problem of locomotion solved then with vision, at least for quadrupeds as an example? Or are there still outstanding issues like a generality gap? How do you think about it?

这很有趣。这实际上取决于你对 locomotion 的界定。一旦机器人能够穿越复杂地形，下一步就更多地进入导航领域，这意味着它应该去哪里？它应该爬上去还是避开？目前，我描述的一切主要基于几何信息，所以没有语义。但现在想象一下，机器人从 A 点移动到 B 点，它会走直线，然后径直穿过中间的任何东西，而不是思考该走哪条路。

It's interesting. So it really depends where you draw the line on locomotion. Once the robot can cross complicated terrain, the next step goes more into navigation, which means where should it go? Should it climb that thing or should it avoid it? For now, everything I've described so far was mostly using geometry. So there are no semantics. But now if you imagine the robot moving from point A to point B and it's going to take the straight line and kind of plow through whatever is between here and there, as opposed to think about which way to go.

差不多。如果有一堵大墙，它可能会避开，但大多数情况下它会试图爬上面前的东西。

More or less. If there is a huge wall it might avoid it but then mostly it will try to climb on whatever is in front of it.

好的。现在想象一下，机器人在办公室里跟在我后面走，你肯定不希望它爬上每一张桌子。你想避开一些东西，但也要走在其他东西上，对吧？所以如果有楼梯，你想走楼梯，但不想撞到植物或其他东西。这意味着突然之间，你必须给策略添加语义。

Okay. Now if you imagine the robot walking behind me in the office, you don't really want it to climb on every single desk. You want to avoid some things but also walk on others, right? So if there's stairs, you want to take the stairs, but you don't want to hit plants or whatever else. That means suddenly you have to add semantics to the policy.

给策略添加语义意味着什么？

And what does it mean to add semantics to the policy?

同样，如果你从仿真到现实，这会让仿真到现实差距变得更大，因为突然之间你必须模拟所有这些办公室、所有这些不同的物体。另外，你可能需要给它图像，而不仅仅是深度图像。你需要给它 RGB 图像，这意味着你需要以照片级真实感的方式模拟它。另一种选择，短期内可能更正确，是拆分问题。所以你训练机器人非常擅长在任何东西上行走，但不给它语义信息，然后在上面训练另一个东西，即规划器或更高级别的策略，来引导它。

Once again, if you go from simulation to reality, it makes the sim-to-real gap bigger because suddenly you have to simulate all these offices, all these different objects. Plus, you probably need to give it images, not just depth images. You need to give it RGB images, which means you need to simulate it in a photorealistic way. The other option, which is probably the more correct one in the short term, is to split the problem. So you train the robot to be very good at walking on anything but you don't give it semantic information, and you train another thing on top which will be the planner or a higher-level policy that will steer it around.

从历史上看，在我与机器人专家的对话中，这是一个很大的争论：我们是应该使用能够解决所有问题的端到端深度学习模型，还是采用更模块化的方法。听起来你在说，更模块化的方法仍然是一种克服端到端训练挑战的务实方式。

Historically, in conversations I've had with roboticists, this has been a big debate: whether we should be using end-to-end deep learning models that can figure all this stuff out, or using a more modular approach. It sounds like what you're saying is that a more modular approach can still be a pragmatic way to overcome the challenges of end-to-end training.

是的。这很有趣，因为我整个博士研究都是关于越来越端到端地处理 locomotion 问题，但我现在却在主张我们不应该什么都端到端。我认为最终我们会达到那个目标，但在短期和中期，正如你所说，更务实的方法是拆分问题，对问题的不同部分使用不同的技术。

Yes. It's interesting because my whole PhD was about going more and more end-to-end specifically for locomotion, and still I'm here arguing that we should not do everything end-to-end. I think at some point we'll get there, but in the short and medium term, as you said, the more pragmatic approach is to split the problem and use different techniques for different parts of the problem.

我们把问题拆分成一个 locomotion 模型和一个规划器模型。规划器的目标是什么？当你试图……你知道，用自然语言描述的目标是“我希望这个东西智能地选择最佳路径”。但什么是“最佳路径”？你怎么定义它？你如何围绕它创建一个目标？是最节能的路径，还是让你更快到达或距离最短的路径？你如何平衡所有这些？

We're splitting the problem into a locomotion model and a planner model. What's the objective for the planner? When you're trying to, you know, the English language objective is 'I want this thing to intelligently choose the best path.' But what is 'best path'? How do you define that? How do you create an objective around that? Is it the path that's most power efficient, or the path that gets you there faster or least distance? How do you balance all that?

有几种不同的方法。如果你仍然选择强化学习路线，你可以用强化学习来训练这些规划器，通常你需要定义一个奖励函数。它会包括诸如“不要撞到任何东西”（避开物体）和“不要移动太快”，因为那是让机器人看起来很危险的一个因素。这些强化学习策略会试图优化一切，所以它们会非常快速地到达目标。但这并不是你想要的在人类周围操作的机器人。通常我们实际上是想尽可能让它们慢下来。嗯，大致就是这样。然后这真的取决于机器人面前有什么。比如，有些东西可以走上去，有些则不行。还有另一种方法。一旦你把问题分成两半，你可以用纯强化学习在仿真中训练 locomotion，但你可以用其他数据训练规划器。例如，人类在办公室走动的视频，你可以从中提取轨迹。这样你就不再需要强化学习了。你训练一个行为克隆、模仿学习策略，它会像人类走路一样引导机器人。

There are different ways to do this. If you still choose the RL route, when you're doing that you can train these planners with reinforcement learning, and typically you have to define a reward function. It would include things like 'don't hit anything' (avoid objects) and 'don't move too fast' because that's one thing that makes robots seem very dangerous. These reinforcement learning policies will try to optimize everything, so they will go very quickly to the goal. This is not really what you want with a robot that operates around humans. Typically we're actually trying to slow them down as much as possible. And yeah, that's mostly it. Then it depends really on what is in front of the robot. Like there might be things that it's okay to walk on, others that it's not. There is also another approach. Once you split the problem in half, you could train the locomotion with pure reinforcement learning in simulation, but you could train the planner with other data. For example, videos of humans walking around the office, and you can extract the trajectories from that. Then you don't really need reinforcement learning anymore. You train a behavior cloning, imitation learning policy that will just steer the robot just like a human would walk.

到目前为止，我们主要讨论的是四足机器人。所有这些有多少可以从四足机器人迁移到人形机器人？

Thus far, we've been talking primarily about quadrupeds. How much of all of this translates from quadrupeds to humanoid robots?

所有技术都可以迁移。这就是强化学习的魔力：这些策略并不关心它们控制的是四足机器人还是人形机器人。这也是从我的博士研究到 Flexion 的重大转变，我们在那里主要研究人形机器人。我们已经看到完全相同的技术可以迁移。人形机器人有一个有趣的现象。

All of it transfers. That is the magic of reinforcement learning: these policies don't really care if they're controlling a quadruped or a humanoid. And this was the big switch from my PhD to Flexion where we're working mostly on humanoids. And we've seen that the exact same techniques transfer. There is one interesting thing that happens with humanoids.

这对规划器来说说得通。但对于 locomotion 模型来说就不那么直观了。也许我想到的 locomotion 模型是分成多个组件的。但让我更具体一点。我把控制步进电机等的输出也包括在 locomotion 中。那是 locomotion 模型训练的一部分吗？然后我想，如果你要改变机器人的外形，你至少需要调整它或做其他事情。

That makes sense to me for a planner. But it's less intuitive for a locomotion model. And maybe I'm thinking of the locomotion model that is split into multiple components. But let me be more specific. I'm including in locomotion the outputs that control stepper motors and all that kind of stuff. Is that part of what is trained in the locomotion model? And then I would think that you would need to at least tune it or do something else if you're going to change the form factor of your robot.

不，这是一个很好的观点。当我说它迁移时，是通用技术迁移，模型本身不迁移。所以你肯定需要为新的机器人重新训练一个新的策略、一个新的控制器。

No, this is a very good point. When I say it transfers, the general techniques transfer, the models themselves don't. So you for sure need to retrain a new policy, a new controller for the new robot.

如果你的所有仿真流水线都足够通用，那么只需更改输入文件（描述机器人的 URDF），重新训练，就可以部署了。调优部分很有意思，因为人形机器人比四足机器人更难调。我不认为这与两条腿或四条腿有关。个人觉得，主要是因为我们对人形机器人应该如何行走有非常具体的期望。四足机器人走路姿势和狗稍有不同没问题，但人形机器人如果手臂摆动不对、膝盖弯太多或走路有点歪，人类会有很强烈的反应。我最近看到一条推文也提到这个：一个人形机器人像四足动物一样移动，背朝下，手臂这样摆，移动很快。推文的主要观点是，人形机器人之所以像人形机器人那样移动，是因为那是我们的期望，但即使有这种外形，也可能存在更高效的移动方式，只是对我们来说看起来不对劲。

If all your simulation pipelines are general enough, it can be as easy as changing the input file, the URDF that describes the robot, retraining it, then you're ready to deploy. The tuning part is interesting because it is a bit harder for humanoids compared to quadrupeds. I don't think it's related to having two or four legs. Personally, I think it's mostly because we have very specific expectations of how a humanoid robot should walk. A quadruped walking slightly differently from a dog is fine. But a humanoid that doesn't move its arms correctly, bends its knees too much, or walks sideways triggers a strong reaction in humans. I saw a tweet recently touching on this: a humanoid robot locomoting like a quadruped, on its back with arms like this, moving quickly. The main point was that humanoid robots move like humanoid robots because that's our expectation, but there might be more efficient ways to move with that form factor, even if they seem wrong to us.

没错。如果我们希望这些机器人在人类周围工作，就必须建立信任。所以如果能让人类更舒适，我们得走那条不那么最优的路线。我们研究机器人已经很长时间了，但过去一年，通过视频演示，我们看到进展非常快。我之前问过这个问题，但我想让你分析一下如何看待这些视频。年初我们看到人形机器人走路，现在它们能跑、能洗碗等等。制作这样的演示需要什么？这些演示在说明机器人能力方面有什么局限性？

Yeah, that's true. If we want these robots operating around humans, we have to create some trust. So we have to take the less optimal route if it makes humans feel more comfortable. We've been working on robots for a long time, but over the past year, we're seeing advancements via video demos very quickly. I asked this before, but I want you to parse how to think about these videos. We saw humanoid robots walking at the beginning of the year, now they're running, doing dishes, etc. What goes into creating a demo like that, and what are the limitations of what it says about the robot's capabilities?

首先，整个生态系统发展如此之快令人兴奋。每天都有新视频展示机器人做新事情，这很好。我们在硬件和 AI 方面都看到了进步，机器人能做的事情越来越多。不过，当我看到演示时，我的方法是思考实现这一目标的最简单方式，而通常这就是实际做法。作为一个生态系统，我们在展示机器人应该做什么的愿景，但幕后有时并非如此。例如，如果机器人站在桌子旁进行操作，比如叠床单，通常有两种情况：要么有人在另一个房间远程操控（约占三分之一），要么机器人是自主的。但要实现自主，需要 100 个人远程操控 100 个机器人，收集数千小时的数据（可能在类似环境中），训练一个策略来模仿这些数据，然后才能自主部署。这和你看到的视频不一样，视频让人觉得机器人能适应任何环境，直接进入你家。我们还没到那一步，但很快就会实现。

First, it's exciting that the whole ecosystem is moving so quickly. Every day there's a new video of a robot doing something, which is great. We see progress in hardware and AI, in what robots can do. However, when I see a demo, my approach is to think about the absolute easiest way to achieve that, and that's typically how it's done. As an ecosystem, we're showing a vision of what robots should do, but behind the scenes it's sometimes different. For example, if a robot stands and does manipulation on a table, folding sheets, one of two things is true: either someone is teleoperating the robot from another room (about one-third of cases), or the robot is autonomous. But to get there, 100 people had to teleoperate 100 robots to collect thousands of hours of data in similar environments, train a policy to imitate that data, then deploy autonomously. That's not what you see in the video, which makes it seem like the robot can adapt to anything and come to your home. We're not there yet, but we'll get there soon.

提醒我一下，那家开始接受人形机器人预售的公司叫什么？就是那种“准备就绪”可以进家用的。你指的是 1x 吗？

Remind me the name of the company that started taking pre-orders for a humanoid robot ready for the home, quote unquote. Referring to 1x?

1x。我看了那个，也聊了很多关于机器人现状的话题。我在想，我们是不是真的比我以为的要进步得多，还是另有隐情。早期买家会成为测试者，机器人到家后可能好用也可能不好用。你对人形机器人进入家庭的准备程度有什么看法？不一定是那家公司。

1x. I looked at that and had many conversations about where robots are. I question whether we're much further along than I think, or if something else is happening. Early buyers will be beta testers, and it may or may not work once it gets to the house. Do you have any takes on the readiness of humanoid robots for the home, not necessarily that specific company?

1x 不是唯一一家。还有几家也宣布了类似计划。从某种程度上说，这很好。明年就把机器人卖到人们家里，这个目标非常雄心勃勃。我认为他们面临巨大挑战。我们看看明年能走多远。但这些公司通常也很坦诚，说这只是测试版的测试版，是 Alpha 项目，只面向早期用户。真正能买到机器人并送进家门，还需要几年时间。这也是我们专注于工业应用的部分原因。工业应用也有其他挑战，比如性能很重要，你需要很快。但你有更多控制权。例如，在一个新仓库部署 10 台机器人更容易：我们派一名工程师去一两天，检查一切正常，微调一下，然后让它们工作。这在每个人的家里是做不到的。

1x is not the only one. A few others announced similar things. In a way, it's good. It's very ambitious to sell robots into people's homes next year. I think they have a big challenge ahead. Let's see how far they get next year. But these companies are usually honest that it's a beta of a beta, an alpha program, just for early adopters. It will take a few more years before you can really buy these robots and send them into homes. That's partly why we focus on industrial use cases. There are other challenges there, like performance is critical; you need to be fast. But you have more control. For example, deploying 10 robots in a new warehouse is easier: we send an engineer for a day or two to check everything, fine-tune, then let them work. That's not possible in everyone's home.

而且工业场景中的任务更重复、更一致，变化更少，不像“跑到冰箱拿瓶可乐”那样多变。

And the tasks in industrial settings are more repetitive, consistent, with less variation than, say, run to the fridge and grab me a Coke.

对，没错。我们可以决定先处理哪些任务，把哪些留到以后。我们从更简单的事情开始：把物体从 A 点移到 B 点，搬运箱子，打开箱子，取出物品，把物品放进箱子，再把箱子装上卡车。这看起来明年或后年就能实现。但即使你看过机器人做这些的视频，也不意味着它现在已经准备好，不再处于开发阶段。这样说公平吗？

Yeah, that's right. We get to decide which tasks to tackle and which to leave for later. We start with simpler things: moving objects from point A to B, moving boxes, opening boxes, taking items out, putting objects into boxes, putting the box in a truck. This seems within reach next year or the year after. But even if you've seen a video of a robot doing that, it doesn't mean it's ready now without being in a development phase. Is that fair?

嗯，公平。我的大胆看法是（我很乐意被证明是错的），目前还没有任何一台人形机器人真正创造了价值。

Yeah, that's fair. My hot take on that, and I'll be happy to be proven wrong, is that there is not a single humanoid robot today that actually generates value.

意思是，可能有一个机器人在工厂或仓库里做了一些相当接近它应该做的事情，但并不是精确的任务。所以最终，它没有产生价值，因为它没有做它应该做的实际事情。

Meaning there might be a robot that does something fairly close to what it's supposed to do in a factory or in a warehouse, but it's not the exact task. So in the end, it's not generating value because it's not doing the actual thing it's supposed to do.

意思是它在做某种变体，或者就像有一个操作员在收拾，你知道，在机器人搞得一团糟之后清理……

Meaning it's doing some variant of the thing or it's like there's a handler that's fixing up, you know, cleaning up after the robot as it makes a mess across the...

完全正确。是的。而且通常会有比之前更多的人手来操作。你可以说价值是负的，但我们会解决这个问题。我们谈到了如何创建这些演示，以及要么是实时远程操作，要么是很多人进行远程操作来收集训练数据。那么，在通过远程操作收集训练数据之后，训练的方法是什么？这些数据是作为强化学习的一部分使用，还是更像是一种监督学习方法？

Exactly. Yeah. And typically would have more handlers than you had people before. You could argue the value is negative, but once again, we'll fix that. Yeah, we talked about like how you create these demos and the idea that there's either real-time teleoperation or many people doing teleoperation to collect training data. Talk a little bit about then, after that training data is collected via teleoperation, what the approach is for training. Is that data then used as part of RL or is that more a supervised learning type of approach?

通常是监督学习。所以你记录数据：来自摄像头的图像，以及远程操作员发送给机器人的指令，通常是如何在空间中移动手以及如何移动手指。这些被记录下来，然后训练一个大型 Transformer 来从相同的图像中产生相同的动作。有趣的是，整个领域从仅仅从头训练这些 Transformer 转向使用在互联网规模数据上预训练的视觉和语言编码器。

Typically it is supervised. So you record the data: images from the camera and the commands that the teleoperator sent to the robot, which typically is like how you should move your hands in space and how you should move your fingers. So that is recorded, and then a big transformer is trained to produce the same actions from the same images. Now what's interesting is the whole field shifted a little bit from just training these transformers from scratch to using vision and language encoders pre-trained on internet-scale data.

比如 VLM，现成的 VLM。你拿一个 VLM，去掉输出层，然后在上面训练网络的新部分，然后称之为 VLA。所以这是一个视觉-语言-动作模型，其中视觉-语言部分是预训练的，而动作部分是从头训练的。

Like VLMs, off-the-shelf VLMs. You take a VLM, you remove the output, and you train a new part of the network on top, and then you call that a VLA. So it's a vision-language-action model where the vision-language part was pre-trained before and the action part is trained from scratch.

明白了。所以与预测下一个语言 token 不同，你现在预测一个动作 token，然后将其转换为独立的电机运动之类的。

Got it. Got it. So as opposed to predicting next language token, you're now predicting an action token which is then translated into a separate motor motion or something like that.

是的，完全正确。

Yeah, exactly.

把这种方法与之前人们所做的进行比较和对比。我们这样做是因为它很酷，还是因为它让我们摆脱了通用 Transformer？从一个预训练模型开始能为我们节省多少？

Kind of compare and contrast that approach with what folks were doing before. Are we doing that because it's cool or because it saves us from the generic transformers? How much does having a pre-trained model to start with save us?

这是一个非常好的问题。普遍的想法是它有助于泛化，但由于语言和视觉编码器是在互联网规模数据上训练的，它们应该能够泛化。一个典型的案例是，如果你不这样做，你会在白天训练一个机器人，然后如果晚上灯光变暗，它就无法再执行任务。尽管仍然有光，一切应该正常工作，人类甚至看不出区别。但因为图像嵌入发生了轻微变化，机器人就无法执行了。我相信预训练编码器能改善这一点。老实说，泛化能力仍有待证明。我认为 Russ Tedrake 在斯坦福大学做了一场精彩的演讲，他们讨论了在丰田研究所的努力，比较了在非常具体的任务上使用少量数据训练策略与使用大量数据训练更通用的策略。他们看到了一些泛化的迹象，但我不想直接引用他的话。似乎还不完全清楚泛化有多少来自那里。

It's a very good question. The general thought is that it helps with generalization, but since the language and vision encoders were trained on internet-scale data, they're supposed to generalize. A typical case was that if you don't do that, you would train a robot during the day, then if the lights go down at night, it won't be able to perform anymore. Even though there's still light, everything should just work, a human wouldn't even see the difference. But because the image embedding changes a bit, the robot doesn't perform anymore. I believe that this gets better with pre-trained encoders. To be completely honest, the generalization capabilities still need to be proven. I think Russ Tedrake has had an amazing talk at Stanford where they were talking about their efforts at Toyota Research Institute, comparing training policies on a very specific task with little data versus training more general policies with a lot of data. They were seeing some signs of generalization, but I don't want to quote him directly. It seemed like it's not fully understood yet how much of the generalization is coming from that.

现在把这两个——少量数据、大量数据——映射到 Transformer VLM 的讨论中。VLM 是少量数据，而 Transformer 是大量数据，因为我们假设 VLM 是预训练的，还是反过来？

Now map those two — little data, lots of data — to the transformer VLM discussion. The VLM would be the little data and the transformer was the lots of data, because we're assuming the VLM was pre-trained, or is it reversed?

不，实际上是相反的。如果你把预训练视为免费获得的数据，那么你就有大量的数据用于预训练，然后只需要添加较少的数据来微调动作头。

No, it's actually reversed. If you include the pre-training as data that you get for free, then you have a massive amount of data for pre-training and just then you can add less data for fine-tuning the action head.

我想哪个是哪个并不重要，因为我听到的结果有些不确定。我们需要更详细地讨论，我在这里引用别人的话，所以很难假设。我认为拥有预训练的视觉编码器和语言编码器非常有意义，因为你不希望每次想让机器人做某事时都重新学习语言。语言就是语言，而且顺便说一句，我们现在有很棒的 VLM，所以不妨使用它们。更多的问题在于这个动作头：你应该在大量随机数据上训练它，还是只在最终希望机器人做的事情上训练它？这仍然是一个开放的问题。

I guess it doesn't matter which one is which because the results were somewhat inconclusive, is what I'm hearing. We need to go into more detail, and I'm quoting other people here, so it's a bit hard to hypothesize. I think it makes a lot of sense to have pre-trained visual encoders and language encoders because you don't want to relearn language every single time you want the robot to do something. Language is language, and by the way, we have amazing VLMs now, so might as well use them. There's more of a question of this action head: should you train it on a lot of random data or just on the things that you want the robot to do in the end? And this is still an open question.

这让我想到一个问题：我刚刚进行了一次对话，我们谈到 VLM 通常会忽略大量视觉信息，而更依赖语言信息。在机器人场景中，这似乎对你试图完成的任务更加有害。你遇到过这个挑战吗？

One question that raises for me is: I just had a conversation where we were talking about how with VLMs generally, they kind of ignore a lot of the visual information and really rely more heavily on the language information. And it seems like in a robotic scenario, that would be even more harmful to what you're trying to accomplish. Do you run into that as a challenge?

我听说过同样的事情。但在 VLA 案例中我没有看到。我猜这是因为机器人无法忽略视觉输入；它是主要的信息来源。然而，往往发生的是它们忽略了语言输入。如果你训练机器人总是做同样的事情——我不知道，如果你有一个盒子，里面有一个物体，它总是要把它拿出来——它会完全忽略语言。它只会做同样的事情。它会试图从图像中猜测它应该做什么。

I've heard the same thing. I haven't seen it in the VLA case. I would guess it's because the robot cannot ignore the visual input; it's the main source of information. What tends to happen, however, is that they ignore the language inputs. If you train the robot to always do the same thing — I don't know, if you have a box and you have an object inside, it always has to take it out — it will completely ignore the language. It will just do the same thing. It will try to guess from the image what it's supposed to do.

你提到了 sim-to-real 差距。我们在缩小这一差距方面取得了良好进展。谈谈你的经验：今天要在仿真中创建模型并在现实中运行需要什么？你是必须明确地做些什么来解决现实世界的问题，还是只是模型更好了、流程更好了，你不再真正考虑这个问题，它就能正常工作？

You mentioned the sim-to-real gap. We've been making good progress on closing that gap. Talk a little bit about your experience: what is required today to create a model in sim and have it run in real? Do you have to do things explicitly to address the real world, or is it just that the models are better, the process is better, and you don't really think about that anymore and it just kind of works?

你需要非常明确地做很多事情。挑战在于，要跨越 sim-to-real 差距，你需要对两个世界——仿真及其工作原理，以及现实世界——都有非常深入的理解。这意味着如果你想让一个机器人在仿真中像应该的那样行走，你需要深入钻研。

You need to do a lot of things very explicitly. The challenge is that to cross the sim-to-real gap, you need to have a very deep understanding of both worlds — of the simulation and how it works, and of the real world. Which means that if you want to have a robot that walks around as it should in sim, you need to go very deep.

你需要确切了解从策略输出的指令一直到电机扭矩之间发生了什么。通常，仅软件层面就有大约 10 层不同的转换，从高级指令到电机中的实际电流。忽略这一点非常诱人，但通过理解每一层并了解所有不同的转换，你才能正确模拟它，这才能真正释放更好的性能。

You need to know exactly what's happening between a command that the policy outputs and then all the way down to torque in the motors. And there are typically 10 different layers of transformations even just on software of how we go from a high level command to actual current in the motors. And it's very tempting to ignore that. But by understanding every single layer and knowing all the different transformations, then you can properly simulate it. And this really unlocks better performance.

所以这意味着你做的模拟并不是那种，你打开一个模拟环境，拿一个通用的人形机器人，训练一个模型然后部署到别的东西上。而是你有一个高保真模拟环境中的人形机器人数字孪生，并且你在非常精细的细节层面上进行训练，这听起来计算成本非常高。

So that suggests like the level of simulation that you're doing isn't like, you know, you pull up your sim environment and get generic, you know, humanoid robot and you're going to train some model and deploy it to something else. It's like you have a digital twin of your humanoid robot in a high-fidelity simulation environment and you're training to a very fine level of detail which sounds very computationally expensive.

不，我们更接近你描述的第一种情况。我们有一个非常通用的模拟环境，但有一些非常具体的东西很重要。一个明显的例子是电机的扭矩和速度限制是什么。你不能期望它做真实机器人上不可能的事情。所以你需要添加这些限制，还有一些类似的东西，比如指令之间会有怎样的延迟。所以你需要识别出其中一些参数，我们通常通过一个称为“真实到模拟”的过程来做。我们取来真实机器人，把它挂在空中，让它稍微晃动一下，收集所有不同电机的数据，然后我们就知道哪些是需要识别的重要效应。我们识别它们并添加到模拟器中。但模拟速度是最重要的。所以你无法负担模拟所有那些不同的效应、电流、磁场等。你需要把它们全部抽象掉。

No. So we're much closer to what you described first. So we have a very generic simulation environment but there are some very specific things that are important. One clear example is what are the torque and velocity limits of a motor. You cannot expect it to do something that is not possible on the real robot. So you need to add those limits and there are a few more things like that like what kind of delay can you expect between a command. So you need to identify a few of those parameters and we are actually doing it usually in what we call a real to sim process. So we take the real robot, we hang it in the air, we let it shake a little bit, collect data of all the different motors and then we know which are those important effects that we need to identify. We identify them and add them to the simulator. But simulation speed is the most important thing. So you cannot afford to simulate all those different effects, currents, magnetic fields, etc. You need to abstract all of it away.

这听起来非常困难且昂贵。我个人看法是，即使你没有模拟它们，你仍然需要理解它们。

That sounded very hard and expensive. And my personal take is that you still need to understand them even though you're not simulating them.

明白。

Got it.

那么这个过程的结果是不是一个可以部署到任何人形机器人的通用模型，而是特定于你为其进行“真实到模拟”的那个人形机器人的模型？你知道，那些关键参数，但因为你能够将其抽象成少数几个关键参数，所以做新机器人相对容易。

And so is it the result of that process is not a general model that you could deploy to any humanoid robot, but one that is specific to the humanoid robot for which you took the real to sim. You know, those key parameters, but that, because you're able to abstract it out to these, you know, some handful of key parameters, like it's relatively easy to do new robots.

是的。这也是今年关键学习之一中令人惊讶的部分：切换机器人相当容易。只要硬件性能合理。所以现在作为一家公司，我们与几个不同的机器人供应商以及几个密切合作的伙伴合作。我们已经在大概 5 到 10 个不同的机器人上部署了控制器。我们现在看到，让一个新机器人行走基本上只需要几天的工作，而且应该更少。如果我们优化一些流程，应该少于一天。现在，让它们执行新任务则更具挑战性。这需要更多的工程工作，这也是我们目前关注的。所以我们的一个关键指标是，将新机器人带到新任务需要多少人力。新机器人现在很容易，新任务是我们正在努力的方向。

Yeah. And this was also the surprising part of one of the key learnings of this year is switching robots is fairly easy. As long as the hardware performs reasonably well. So now as a company we work with a few different suppliers of robots and a few different partners as well with whom we're working closely. We've deployed controllers on let's say between five and 10 different robots. And we see now that making a new robot walk is basically a few days of work and it should be less. It should be less than one day of work if we optimize some of our processes. Now bringing them to a new task. This is a bit more challenging. This requires more engineering today and this is what we're focusing on. So one of our key metrics is how much human effort is involved in bringing a new robot to a new task. New robot is very easy today. New task is something we're working on.

在这个背景下，我们稍微讨论过这一点，但一个任务具体到什么程度？意思是，任务是“抓取和放置”，还是“这个仓库里的机器人从这条线上抓取并放入这些箱子”？

And in this context, like how we've kind of talked a little bit about this, but how specific is a task? Meaning like is a task pick and place or is a task robot in this warehouse picking off of this line and placing into these bins?

这是个好问题。更像是“抓取和放置”。但有一个有趣的概念：我们试图利用大型 VLM 中包含的信息来编排和分解复杂任务为清晰的子任务，尽管这不是我们目前关注的重点。烹饪是一个很好的例子，一个很好的比喻。如果你想训练机器人烹饪地球上的每一顿饭，并且你说每顿饭都是一个独立的任务，那你永远也完不成，对吧？任务集是巨大的。但你可以做的是，把食谱给一个 VLM。你也给它机器人看到的图像。如果食谱说“切一根黄瓜”，VLM 会说，拿起刀，拿起黄瓜，然后做这样的动作来切它。然后你可以将其分解成更简单的原语，比如切东西、拿锅、把它放在某处、倒一杯水、倒水等等。突然之间，这些原语的集合就不再是无限的了。挑战在于，我们现在需要一个更高层次的智能来编排所有这些原语。但有趣的是，这部分基本上已经通过 VLM 解决了。它还没有 100% 到位，但它的进展比实际执行所有这些动作的物理交互要快得多。

That's a great question. More like pick and place. But there's an interesting concept there which is we are trying to leverage the information contained in large VLMs to orchestrate and break down complex tasks into clear subtasks even though that's not what we're focusing on today. Cooking is a great example, a great metaphor. If you wanted to train the robot to cook every single meal on the planet and you would say each meal is its own task, you would never finish that, right? The set of tasks is huge. But what you could do is you give the recipe to a VLM. You also give it images of what robot sees. If the recipe says cut a cucumber, the VLM would say, you know, grab the knife, grab the cucumber and do this sort of motion to cut it. And then you can break it down into much simpler primitives like cutting things, holding a pan, putting it down somewhere, filling a glass of water, pouring it, things like that. And suddenly the set of these primitives is not infinite anymore. The challenge is that now we need a higher level intelligence that will orchestrate all these primitives. But what's interesting is that that part is basically solved with a VLM. It's not 100% there, but it's moving much faster than the actual physical interaction of doing all these motions.

那么你能详细说明一下吗？编排问题解决了，如果是的话，它是如何解决的？这种编排与我们讨论的这些大模型的推理能力之间是什么关系？是同一回事还是相关的？

So can you elaborate on that? The orchestration is solved and if so how is it solved and what's the relationship between that orchestration and what we talk about as the reasoning capabilities of these large models? Is it the same thing or related?

类似。也许描述这一点的一种方式是，我们的网站上有两个视频。一个视频是机器人在森林里行走并捡垃圾。这主要是为了展示可能性，也稍微利用了我们瑞士的角度来展示自然。另一个视频是机器人在我们办公室里做同样的事情。在第二个视频中，它是 100% 自主的。你给它一个文本提示。我想我们说的是类似“捡起你面前的玩具，然后把它们放到末尾的篮子里”这样的话。我们为此使用了一个现成的 VLM。它的工作方式是，我们把图像给 VLM，并允许它调用机器人的特定技能。所以 VLM 会说：“哦，我看到地上有一个玩具。我们走到玩具那里。”而这个“走到玩具那里”是一个技能，实际上由机器人触发并执行。一旦我们到了那里，它会触发捡起玩具，然后走到篮子那里放下它。因此，通过拥有几个这样的原语，比如走到某物——我的意思是，行走本身就像我们讨论的那样相当复杂，你可以走楼梯，可以在各种复杂地形上行走——但通过拥有行走、从地上捡起东西、然后放到别处这些原语，我们可以以许多不同的方式重新组合它们，无需任何重新训练，只需提示一个现成的 VLM。

It's similar. Maybe one way to describe this is on our website we have two videos. We have a video of a robot walking in forest and picking up trash. This is mostly there to showcase what's possible and also play a little bit on our Swiss angle to using our nature. We have another video where the robot is doing the same thing in our office. And in that second video, it's 100% autonomous. You give it a text prompt. I think we're saying something like pick up the toys in front of you and drop them in the basket at the end. And we're using an off-the-shelf VLM for that. The way this works is we're giving the images to the VLM and we're allowing it to call specific skills of the robot. So the VLM would say, "Oh, I see a toy there on the ground. Let's walk to the toy." And this "let's walk to the toy" is a skill that is actually triggered and executed by the robot. Once we're there, it will trigger pick up the toy, then go to the basket and drop it off. And so by having a few of those primitives which are walk to things, I mean the walking is locomotion as we discussed is itself fairly complicated. So you can walk on stairs, you can walk on a bunch of different complex terrain, but by having like walking, picking things up from the ground, and then dropping them somewhere else, we can recombine this in many many different ways without any retraining just by prompting an off-the-shelf VLM.

是这样吗？那么你怎么看：如果我们有一堆工具，也就是一堆独立的模型或模块，它们实际上是分开训练的吗，还是说它们更通用一些？

Is that the case? Then how do you think about if we have a bunch of tools, separate models or modules, are they actually separate and trained separately, or are they more universal somehow?

这是另一个很好的问题。具体到我们目前的情况，这些模型是分开的，但我们正在积极努力将它们合并成一个单一的模型，一个更通用的模型。这样做的希望是能看到一些泛化，也就是不同模型之间的插值。我们的做法是：仍然分别训练那些不同模态的基元，然后用它们作为数据生成器，在仿真中收集大量数据，再训练一个更大的 VLA 覆盖整个数据集，使其能执行所有任务。我们在公司内部已经看到了初步成果。事情正朝着那个方向发展，但我们还需要证明这确实能带来我们所说的泛化。

That's another great question. In our case specifically today, they are separate, but we are actively working on merging them together into one single model, a more general model. The hope with that is that you see some generalization, some interpolation between those different models. The way we would do that is by still training those different modalities primitives separately, then using them as data generators to collect a massive amount of data in simulation to train one larger VLA across the whole dataset so it can perform everything. We're seeing early results of that in our company. Things are going in that direction, but we still need to prove that this actually leads to the generalization we're talking about.

听你这么说，我经常想到师生方法或蒸馏来得到更小的模型，但你并不一定在这么做。从模型推理的角度谈谈硬件能力，以及我们在模型规模方面处于什么位置。

Hearing you describe that, I often think of student-teacher approaches or distillation to get smaller models, but you're not necessarily doing that. Talk a bit about hardware capabilities from a model inference perspective and where we are in terms of model size.

在我们的计划中，如果走完整个过程，我们会训练大约 50 个这样的基元。我们会蒸馏所有内容。我们仍在开发一个层级流水线，其中有三个模型相互交互。首先是一个相对较大的 VLM，它不允许使用推理。它的输出是：VLM 从一个非常抽象的任务分解为清晰的子任务。例如，如果你有一个机器人，我让它去冰箱里拿东西，它会说：转身，穿过门，打开冰箱，抓住东西，关上冰箱，等等。清晰的指令。然后这些清晰的指令会传给之前描述的 VLA。如果它收到像“打开冰箱”这样的指令，它会规划手臂的运动学动作，抓住冰箱把手并在未来几秒内打开它。最后，我们有一个所谓的全身跟踪器，它会接收这个关于手和全身如何移动的计划，然后控制该特定机器人的电机来执行该动作。

In our plan, if we go through that whole process, we train, let's say, 50 of those primitives. We distill everything. We are still developing a hierarchical pipeline where we have three models interacting with each other. It would start with a relatively large VLM that would not be allowed to use reasoning. The output of that one would be that VLM would go from a very abstract task to clear subtasks. For example, if you have a robot here and I tell it to go pick something up in the fridge, it would say: turn around, go through the door, open the fridge, grab the thing, close the fridge, etc. Clear instructions. Then those clear instructions would go to the VLA we described before. If it receives an instruction like 'open the fridge', it will plan a kinematic motion for the arm to grab the fridge handle and open it just a few seconds into the future. Finally, we have what we call a whole body tracker, which will receive this plan of how the hand should move and how the whole body should move, and then control the motors of that specific robot to execute that motion.

好的。

Okay.

这些模型的大小和频率差异很大，这就是为什么我们认为用三个模型是合理的。最后一个，全身跟踪器，是一个非常简单的、非常小的模型。通常是一个很小的 Transformer。老实说，它甚至不需要是 Transformer，但如今什么都是 Transformer。所以它可以很容易地以 50 Hz 运行，每秒 50 次，甚至在机器人板载计算机的 CPU 上运行，而不是 GPU，因为发送到 GPU 再取回需要更多时间。我先跳过 VLA。来说 VLM。这个模型通常很难在机器人上板载运行。目前，它是在外部运行的，要么在我们办公室的服务器机架上，要么在云端。这在你想要在仓库中部署 100 个机器人时会带来一些挑战：要么你有非常好的互联网连接，要么你必须在仓库里安装服务器机架。所以我们希望机器人计算能力不断进步，最终能把这些东西放在机器人本身上，通常是在 Jetson 上。然后是 VLA，这是目前计算最受限的地方，因为我们不能把它放在外部。它仍然需要运行得相当快，比如每秒 10 次，延迟最小，所以它必须板载。它们通常也使用扩散模型，这意味着你不只是推理一次；至少网络的一部分要推理多个步骤。所以这是计算最关键的地方。机器人的板载计算是最关键的。

Now, the size and frequencies of these models are very different, and that's why we think it makes sense to have three of them. The final one, the whole body tracker, is a very simple, very small model. Typically, it's a very small transformer. To be completely honest, it doesn't even need to be a transformer, but today everything should be a transformer. So that can run very easily at 50 Hz, 50 times per second, even on the CPU of the onboard computer of the robot, not even the GPU, because it takes more time to send it to GPU and get it back. I'll skip the VLA for now. Going to the VLM. That one is typically fairly hard to run on board the robot. For now, it's running offboard, either in our office in a server rack or even in the cloud, which creates some challenges once you want to deploy 100 robots in a warehouse. Either you have an amazingly good internet connection, or you have to install server racks in that warehouse. So we are hoping that robot compute keeps progressing so we can finally fit those things on the robots themselves, typically on a Jetson. Then the VLA, this is where compute is the most limiting today, because we cannot really put it offboard. It still needs to run fairly fast, let's say 10 times per second, with minimal delay, so it needs to be on board. They also typically use diffusion, which means you don't infer it just once; at least part of the network is inferred multiple steps. So this is where compute is the most critical. The onboard compute of the robot is the most critical.

是的。我之前没想过，在 1X 和其他家用机器人的情况下，它们的大脑基本上在云端。我原以为我们能让这些模型足够小，可以在本地运行，这看起来很多，但从延迟的角度来看，很难想象这能特别可行和一致。

Yeah. I hadn't thought about, in the case of the 1X and these other home robots, that they're essentially like their brains are in the cloud. I assumed we were able to get these models small enough to run locally, which seemed like a lot, but from a latency perspective, it's hard to imagine that being particularly tenable and consistent as well.

所以这些模型中的一些，澄清一下，一些模型可以放在机器人上。但我们今天看到的是，如果你想做一些更抽象的推理，比如给它一个非常抽象的任务，它需要协调几分钟的事情，那么你真的会从更大的模型中受益。

So some of these models, just to clarify, some of these models can fit on the robot. But what we're seeing today is if you want to do some of this more abstract reasoning, so you give it a very abstract task and it has to orchestrate something for multiple minutes, you would really benefit from larger models.

是的。而且我想，在家庭环境中，对于消费者任务，你会需要比工业环境中更抽象的模型。是这样吗？

Yeah. And I would imagine that you would want even more abstract models in the home for consumer tasks than you would require in an industrial setting. Is that true?

是的，我认为是这样。因为在工业环境中，如果任务是重复性的，你或多或少可以预先计算那些非常抽象的指令，或者从非常抽象的指令转到清晰的指令。在家里，如果有人直接告诉你一些事情，你肯定需要大规模模型。

Yes, I would say that's true. Because in an industrial setting, if the task is repetitive, you can more or less precompute those very abstract instructions or go from very abstract instructions to clear instructions. In a home, if you have a human just telling you something, for sure you need the scale of large models.

你们是使用现成的强化学习环境，还是说你们创建的部分内容包括用于创建这些模型的仿真环境？

Are you using off-the-shelf RL environments, or is part of what you're creating the simulation environment for creating these models?

这是我们创建内容的重要一步。我们不是自己构建模拟器。我们使用现有的模拟器，包括 Nvidia 的。我们也测试和实验了许多其他模拟器。但我们的一个关键专长是如何正确地构建那些仿真环境，并且我们在其之上有自己的自定义强化学习算法，以尽可能从中受益。

It is a big step of what we're creating. So we are not building simulators ourselves. We're using existing simulators, including ones from Nvidia. We also test and experiment with many others. But one of our key know-how is how to properly build those simulation environments, and we have our own custom RL algorithms on top to benefit from that as much as possible.

明白了。所以模拟器和仿真环境是不同的。

Got it. So the simulator and the simulation environment are distinct.

你说的仿真环境是指模拟器中的环境配置吗？模拟器像是平台，而仿真环境是你针对场景创建的东西？还是说只有这两个抽象层次，或者有三个？我猜你会把强化学习算法作为第三个组件，它与两者交互，但仅此而已。

Is the simulation environment when you say that is that the configuration of the environment in the simulator like the simulator is like the platform and the simulation environment is like the thing that you create about your scenario? Or are there just those two levels of abstraction or there are three levels of abstraction? I guess you would add the RL algorithm as a third component that interacts with both, but that's it.

模拟器本身基本上是一个物理引擎和渲染器，然后你需要把机器人放进去。如果你想在楼梯上行走，你必须创建楼梯，但你不能让机器人随机摸索如何在非常复杂的楼梯上行走。所以你必须创建一整套我们所谓的难度课程。你会从非常小的楼梯开始，然后逐步增加难度。所有任务都是如此。当我们训练机器人开门时，我们必须创建一个模拟版本的门，然后帮助机器人。所以我们必须在这些场景之上，精确地设计所有训练过程。

The simulator itself is basically a physics engine and a renderer, and then you have to put a robot in there. If you wanted to walk on stairs, you have to create stairs, but you cannot just ask your robot to randomly figure out how to walk on very complex stairs. So you have to create a whole what we call a curriculum of difficulty. You would start with very small stairs and progressively make them harder. The same is true for all sorts of tasks. When we're training a robot to open a door, we have to create a simulated version of the door, then we have to help the robots. So we have to figure out exactly all these training processes on top of just a scenario itself.

我们讨论过，你把强化学习和模仿学习定位为两种替代方案。但有没有可能将模仿学习与强化学习结合使用，来引导学习，帮助机器人更快地掌握楼梯？这仍然是一个研究问题，还是我们现在已经能在实践中做到的事情？

As we've talked about this, you've positioned RL and imitation as these two alternatives. But is it also possible to use imitation in conjunction with RL to bootstrap learning and help the robot figure out stairs more quickly? Is that still a research problem or something we're able to do in practice now?

这是个好问题。这仍然是一个研究问题，但我们看到了积极的迹象。我可以谈两种不同的结合方式。一种是用少量演示来辅助强化学习过程。这是我们正在积极推进的。如果有人演示任务，你可以从中提取少量信息来帮助强化学习的探索过程，这样机器人就不会只是随机乱动、从头摸索，而是被引导向正确的解决方案。另一种完全不同的方法是我们在其他公司和学术界看到的：用更多的模仿学习进行预训练，然后在此基础上加入一些强化学习来改进行为。一个关键问题是，如果模仿学习预训练是在没有模拟器的情况下完成的，你是否需要突然再次引入模拟来进行微调？我认为这仍然是一个非常开放的问题。

That's a good question. It's still a research problem, but we are seeing good signs of life. I can talk about two different ways to combine them. One way is to use a few demonstrations to help the RL process. This is something we're doing very actively. If you have a human showing the task, you can extract just a little bit of information to help the exploration process of reinforcement learning, so the robot is not just randomly shaking and trying to figure out everything from scratch, but you're guiding it a little bit towards the right solution. A completely different way to approach imitation learning plus RL is what we're seeing more in other companies and academia: doing more imitation learning for pre-training and then adding some flavor of RL on top to try to improve the behavior after the fact. A big question there is if the imitation learning pre-training was done without any simulator, do you suddenly need to add a simulation again to do the fine-tuning or not? I think this is still a very open question.

回想我在机器人领域最早的一些对话，比如 Peter Reveal 这些人，那可能是在 Transformer 之前。一些最早的工作在 Google 用了数百台真实机器人，他们开始尝试强化学习。机器人在做强化学习，成本非常高。他们不需要处理仿真到现实的差距。你更专注于仿真和缩小差距的方法，但仍有支持在现实中进行强化学习的人。你如何看待比较和对比这些方法？

When I think back to some of the earliest conversations I had in robotics, like folks like Peter Reveal, this was pre-Transformer maybe. Some of that earliest work had hundreds of real robots at Google, and they were starting to experiment with RL. The robots were RL-ing and it was very expensive. They didn't have to deal with the sim-to-real gap. You're more focused on simulation and ways to close that gap, but there are still proponents of RL in real life. How do you think about comparing and contrasting those approaches?

确实有人大力支持在现实中做强化学习，或者完全反对仿真。在某些情况下，这是有充分理由的，因为有些东西从根本上很难仿真。我可以谈谈我们正在关注的一个具体任务：机器人操作纸箱。机器人必须行走、拿起箱子、把它带到某处、放在桌子上、打开它、取出里面的东西、然后放到架子上。大多数子任务都非常容易仿真，除了一个特定环节：打开箱子，因为箱子上可能有胶带。你需要用刀割开胶带才能打开。这是可能的，但模拟胶带与纸板的相互作用以及刀如何切割仍然需要大量工作。所以我们试图做的是识别那些仿真仍然有限的特定情况，只对这些特定情况使用真实数据，然后将其与其他所有仿真数据混合。我们认为这能两全其美，尽可能多地使用仿真。随着模拟器的发展，它们将覆盖更多的任务。但在存在差距的情况下，我们会为这些特定情况复用真实数据。

There are definitely people who are big proponents of RL in real life, or against simulation at all. In some cases, there are good reasons for that because some things are fundamentally hard to simulate. I can talk about a specific task we're focusing on: the robot manipulating cardboard boxes. The robot has to walk, pick up a box, bring it somewhere, put it on a table, open it, take what is inside out, and put it on a shelf. Most of the subtasks are very well simulatable except one specific piece: opening the box, because there might be tape on the box. You have to take a knife, cut through the tape to open it. It's possible but still a lot of effort to simulate the interaction of the tape with the cardboard and exactly how the knife cuts through it. So what we are trying to do is identify those specific cases where simulation is still limited and use real data only for those very specific cases, then mix it with simulated data of everything else. We think this gives us the best of both worlds, getting as much simulation as possible. As simulators develop, they'll take more of the whole set of tasks. But while there is a gap, we'll reuse real data for those specific cases.

对于那些说在现实中做强化学习更好的人，对于那个特定场景，它会在哪些方面更好？听起来你只是会消耗很多箱子，但在现实中你仍然有同样的问题，对吧？

For the folks that say RL in real is better, in what ways would it be better for that particular scenario? Sounds like you would just go through a lot of boxes, but you still have the same problem in real, right?

我同意你的看法。我认为在现实中做比在仿真中难得多，尤其是使用强化学习。你可以做的是针对那个特定案例进行少量模仿学习，这比让机器人从头学习所有东西要容易得多。我猜纯粹在现实中做强化学习的唯一理由是，你不需要处理仿真，而仿真可能非常困难，特别是如果你内部没有创建仿真环境和调整模拟器使其良好运行的专长。在现实中，你某种程度上已经有了完美的模拟器。但另一方面，你有非常昂贵的硬件，任何失败和重置的成本都比仿真高得多。

I agree with you. I think it is much harder to do it in real life compared to simulation, especially with reinforcement learning. What you could do is a little bit of imitation learning for that specific case, and that's much easier than letting the robot learn everything from scratch. I would guess the only argument for pure real-life RL would be that you don't need to deal with simulation, which can be very hard especially if you don't have expertise in house in terms of how to create those simulated environments and how to tune the simulators to behave nicely. When everything is in real life, you already have the perfect simulator in a way. But on the other hand, you have very expensive hardware, and any failure and reset is way more expensive than in simulation.

是的，我明白为什么它很有吸引力，也为什么它令人向往。作为人类，我们不会通过仿真世界来学习，我们在世界中探索并以此学习，所以我希望我的机器人也能做到这一点。但机器人的样本效率远不及人类。所以你会弄坏很多机器人或箱子才能达到目标，而且你只解决了一个任务。我认为另一个有趣的点是，人类的奖励信号极其复杂。

Yeah, I think it's clear why it's compelling and also why it is aspirational. The idea that as humans we don't simulate the world to learn things, we explore in the world and we learn that way, so I'd want my robot to be able to do that. But we're nowhere near the sample efficiency in robots as we are in humans. So you would end up breaking a lot of robots or boxes to get there, and then you've only solved one task. I think another interesting point is that the human reward signal is extremely complicated.

如果你想想用手做任务，从所有神经、皮肤以及疲劳的肌肉中获取的信息量极其复杂。而机器人根本没有这些信息。通常，触觉感知非常原始。所以如果机器人慢慢损坏自己，你直到电机坏了才会知道。如果你在现实中进行强化学习，消除会损坏电机行为根本行不通，因为你可能每周只遇到一次事件。所以奖励信号不存在。这正是仿真再次发挥作用的地方，因为在仿真中我们拥有关于一切的完美信息。例如，我们可以设计奖励函数来避免损坏电机或破坏机器人的机械结构。

If you think about doing some tasks with your hands, the amount of information you get from all the nerves, your skin, and your tired muscles is extremely complicated. We don't have that information at all with a robot. Typically, tactile sensing is very primitive. So if the robot is slowly damaging itself, you wouldn't know until a motor breaks. If you're doing reinforcement learning in real life, getting rid of behaviors that would damage the motors simply won't work because you'll get maybe one event per week. So the reward is not there. This is where simulation helps once again, because in simulation we have perfect information about everything. We can design reward functions that will avoid breaking motors or damaging the mechanics of the robot, for example.

根据你之前的观点，加入视觉会降低性能或使模型更难收敛，一种方法是直接添加皮肤或额外传感器。但每增加一个传感器都会增加这些模型的计算负担。

Based on your earlier point about incorporating vision reducing performance or making it more difficult to converge on a model, one approach is to just add skin or additional sensors. But any additional sensor increases the computational burden on these models.

完全正确。而且，从机械上讲，这会让一切变得更脆弱。对于摄像头，我认为我们今天已经做到了。我们可以给机器人加摄像头，它们非常便宜且可靠。但触觉传感今天还做不到。

Absolutely. Plus, mechanically, it makes everything more brittle. With cameras, I would say we are there today. We can add cameras to our robots. They're very cheap and very reliable. Tactile sensing is just not there today.

设计奖励函数是任何类型强化学习的关键。对于给定任务，它们有多标准化？还是差异很大，需要大量手动调参？第二个问题，在编码智能体中，有很多讨论关于引入价值函数，在最终目标之前为积极行为提供信号。价值函数在今天的机器人领域是实用的，还是仍处于研究阶段？

Crafting that reward function is key in any type of RL. How standardized are they for a given task, or do they vary widely and require a lot of hand tuning? As a secondary question, in coding agents, there's talk about incorporating value functions to provide signal for positive behavior before the end objective. Are value functions practical in robotics today, or is it still research?

完全正确。价值函数是强化学习算法本身的一部分。我们主要使用 PPO 的变体进行训练，这是一种演员-评论家算法，意味着同时训练演员和评论家。评论家本质上是一个价值函数。它在部署时不使用，只在训练过程中帮助训练演员。现在我们看到一些研究探讨如何在部署时也使用它。我认为这更多是研究方面，尚未得到验证。关于奖励调参，这是一个大话题。我们公司有 35 人，其中不少人都花时间调奖励。通常，这被视为负面的事情——你不想让人调奖励。但我认为必须区分两种调参。一种是来自任务本身的一般奖励。例如，如果机器人在行走，你说任务是让机器人从 A 点走到 B 点。但实际上更复杂：你希望机器人从 A 到 B，但不要消耗太多能量，不要撞地太猛，不要到处打滑，手臂不要乱晃。当你描述真正想要的东西时，可能已经有 15 行，这对应 15 个不同的奖励函数需要调参。我个人认为这部分调参没问题。另一种调参与探索相关：一旦你描述了完美任务，如何引导策略训练过程走向那个目标？例如，如果你想让机器人开门，你可能需要告诉它把手靠近门把手，然后合拢手指，再拉门。这些方法不能跨任务扩展。这是我们尽量要避免的。这就是为什么我们正在研究其他技术，利用人类的一两次演示来帮助学习过程，而不是手动调奖励函数。

Absolutely. Value functions are part of the RL algorithm itself. We train mostly using some variant of PPO, an actor-critic algorithm, which means training both an actor and a critic. The critic is basically a value function. It's not used at deployment, only to help train the actor during training. Now we're seeing some research into how it could be used even at deployment. I think this is more on the research side; it's not proven yet. About reward tuning, it's a big topic. We have 35 people in the company, and quite a few spend hours tuning rewards. Typically, this is referred to in a negative way—you don't want people tuning rewards. But I think you have to distinguish between two types of tuning. There are general rewards that come from the task itself. For example, if a robot is walking, you say the task is to go from point A to point B. But in reality, it's more complicated. You want the robot to go from A to B without using too much energy, without hitting the ground too hard, without slipping everywhere, without the arms doing crazy motions. By the time you describe what you actually want, you have maybe 15 lines, which translates to 15 different reward functions to tune. My personal take is that part of tuning is fine. There's another kind of tuning related to exploration: once you've described the perfect task, how do you guide the policy training process towards that? For example, if you want a robot to open a door, you might need to tell it to put its hand close to the handle, then close its fingers, then pull. These things don't scale across tasks. This is something we're trying to avoid as much as possible. That's why we're working on other techniques where we can use one or two demonstrations from a human to help the learning process instead of manually tuned reward functions.

现在是年底，正是做预测的时候。你对未来一年或几年在机器人领域有什么预测？你怎么看待未来？

We're at the end of the year, a natural time for predictions. Do you have any predictions for the upcoming year or several years in terms of robotics? How do you think about the future?

我之前说过，我认为当今世界上没有一个能提供价值的人形机器人。我的预测是，这种情况将在明年年底或 2027 年初改变。很难准确预测会发生什么，但我认为不会出现一个 ChatGPT 时刻，突然有十亿个机器人到处都是，因为你需要制造硬件，这不像访问 ChatGPT 那样可以扩展。但我预测，到明年年底，我们将开始看到机器人做实际工作。一开始只是零星的几个。然后在 2027 年和 2028 年，我们将扩展每个任务的机器人数量以及它们能完成的不同任务集。在未来几年，我们将从几百个机器人发展到几千个，然后很快到几百万甚至几千万个。而且，这很可能首先发生在工业环境中，然后是消费领域。

I said before that I think there isn't a single humanoid robot providing value in the world today. My prediction is that will change around the end of next year, maybe beginning of 2027. It's hard to predict exactly what will happen, but I don't think we'll get a ChatGPT moment where suddenly there are a billion robots everywhere, because you need to build the hardware; it doesn't scale like getting access to ChatGPT. But I predict that around the end of next year, we'll start seeing robots doing actual work. It will be just a few here and there. Then in 2027 and 2028, we'll scale both the numbers of robots per task and the set of different tasks they can do. In the coming years, we will go from hundreds of robots to thousands, then very quickly to millions and tens of millions. And presumably, this will happen first in industrial settings, then consumer.

而且你预计这首先发生在工业环境，然后是消费领域。

And presumably you see that happening first in industrial settings and then consumer.

这是我的预测。是的，我们首先会在工业领域看到，然后是消费领域，比如家庭。我希望在那之后我们能进入更疯狂的应用，比如我们应该在人类登陆火星之前，先把人形机器人送到那里建造殖民地。

This is my prediction. Yes, that we'll see that first in industrial, then consumer. At home. And I hope that after that we can go to crazier applications, like we should really send humanoid robots to Mars to build colonies before humans land there.

当你想到目前见过或使用过的机器人时，它们都一样吗？比如，所有的机器狗都一样，所有的人形机器人大致相同？还是从硬件角度看有很大差异？有没有哪些让你特别兴奋？

When you think about the currently available robots you've seen or worked with, are they all the same? Like, are all the dogs the same, all the humanoids roughly the same? Or do you see big differences from a hardware perspective? Are there any particularly exciting ones for you?

这是个好问题。

It's a great question.

在设计人形机器人时，大概有三四种不同的策略，具体来说就是使用什么样的执行器、什么样的齿轮箱。由于有很多公司都在探索这个领域，一切都在并行推进。无论你采用哪种策略，美国有几家公司，欧洲可能有一家，而中国大概有 50 家在做完全相同的事情。所以竞争非常激烈。目前硬件还没到位的一个部分是末端执行器，也就是手。是否需要非常灵巧的手仍然存在争议。我认为很多公司开发高灵巧度手部（比如超过 20 个自由度）的一个主要原因是他们使用模仿学习。他们模仿人类，这意味着你需要能够模仿人类所做的一切。一旦你走上这条路，你就可以用更简单的夹爪学习其他类型的行为。但这个问题仍然没有定论。

There are probably three or four different strategies you can take in terms of how you're designing your humanoid robot, specifically what kind of actuators you use, what kind of gearboxes. And since there are many companies exploring that space, everything is happening in parallel. No matter which of those strategies you take, there are a few companies in the US, maybe one in Europe and probably 50 in China building that exact thing. So the competition is really fierce. One part where hardware is not there yet today is on the end effectors, on the hands. It's still debatable if you actually need very dexterous hands. I think one of the big reasons why many companies develop hands with high dexterity, so like more than 20 degrees of freedom in a hand, is because they're using imitation learning. They're imitating humans, which means that you need to be able to imitate everything a human does. Once you go that route, you can learn other kinds of behaviors with much simpler grippers. But the debate is still open on that.

我们一直在讨论机器狗和人形机器人。但有一个更广泛的问题：人形是机器人的最佳形态吗？我们是否应该制造有两只手臂、两条腿的机器人？你认为这是正确的方向吗？还是说我们只是因为这是人类的形态而固守于此，但实际上有更好的形态？

Again, we've been talking about dogs and humanoids. But there's this broader question: is humanoid the best form factor for a robot? Should we be making robots with two arms, two legs? Do you think that's the way to go? Or are we kind of anchored on this because it's our form, but there are better forms that you've seen or think about?

这是另一个很好的问题。到目前为止，我参与过超过 25 个机器人项目。你能想象到的任何数量的腿和手臂，从零到大概四、五、六条腿，世界上有各种机器人的空间。老实说，我和我们公司使用“人形”这个词是因为没有更好的词。我们指的不是人类形态，而是人类能力。基本上，我们希望机器人能够去人类能去的地方，并以类似人类的方式操作环境。所以你可能需要至少两只手臂和某种末端执行器来与环境互动。至于用腿还是轮子，两者都可以，各有各的应用场景。尤其是在工业领域，轮式平台可以走得很远。我们也在使用这样的机器人。话虽如此，轮式平台很容易被卡住，这令人惊讶。一个很常见的例子是，如果地面不完全平坦，有电缆，或者当然还有楼梯，轮式平台就会被卡住。但另一个非常重要的方面是占地面积。对于这些轮式平台，你有两个选择：要么做得很大，这样它们天生稳定，因为占地面积大，但无法通过狭窄空间。尤其是在一些较老的工业环境中，很快你就会遇到狭窄空间。

That's another great question. So I've worked on more than 25 robots by now, I think. Any number of legs and arms you can imagine, from zero to probably four, five, six legs, and there is room for all sorts of robots in the world. Honestly, I mostly use, and as a company we use the word 'humanoid' for lack of a better word. What we mean by that is not the human form factor but human capabilities. Very basically, we want robots that can go where humans go and can manipulate the environment in a similar way to how humans do that. So probably you would need at least two arms with some sort of end effector to interact with the environment. And then whether you have legs or wheels, both are fine. Both have their own applications. So you can go a long way, especially in industry, with a wheeled platform. And we're working with robots like that as well. Having said that, it's surprising how quickly wheeled platforms get stuck. A very common thing that is easy to imagine is if the floor is not perfectly flat, if you have cables, or of course stairs, your wheeled platform is stuck. But another very important part is also the footprint. So with those wheeled platforms, you have two choices: either you make them very large and then they're stable by default because they have a very large footprint, but they don't fit through tight spaces anymore. And very quickly, especially in slightly older industrial settings, you have tight spaces.

另一种选择可能是类似陀螺仪平衡车的东西。

And the alternative is maybe some gyroscopic Segway-like thing.

那可能是一种。通常我们看到的是，你只有一个很小的平台，这意味着你必须非常小心地移动上面的躯干，因为如果倾斜太多，它就会翻倒。

That could be one. I typically what we see is that you just have a small platform, which means that you have to be very careful how you move the torso on top because if you lean too far it just falls over.

我还没见过陀螺仪平台。

I haven't seen the gyroscopic platform yet.

好吧。我仍然认为三足机器人可能是我见过的最酷的机器人，但它可能不是地球上工业任务中最适用的。

Okay. I still think a three-legged robot is probably the coolest robot I've seen so far, but it's maybe not the most applicable for industrial tasks on Earth.

现在有很多机器人套件，机器人对于对这个领域感兴趣、想动手玩但没有条件接触人形机器人的人来说越来越容易获得。有什么酷的东西是现在就能订购的，也许在假期或之后不久就能到手，而不是像预购 2027 年的产品？如果你在给一个跃跃欲试的人提建议，你会告诉他们从哪里开始？

There are quite a few robotics kits now, and robots are getting more accessible for folks interested in the space who want to play but don't have access to a humanoid robot. What are some cool things that someone can order now, maybe by the holidays or soon thereafter, not like pre-order for 2027, and start playing around? If you were advising someone excited about getting their hands dirty, what would you tell them to start doing?

Hugging Face 和他们的 LeRobot 项目有一个很棒的社区，他们提供非常便宜的机械臂，可以帮助你学习整个遥操作、数据收集、训练和部署流程。这是一个非常好的学习途径。对于更偏向运动强化学习的方面，难度会大一些，因为你可能需要一个带腿的机器人，这也意味着机器人应该能够摔倒并站起来而不完全损坏。我认为最好的选择是中国的四足机器人，它们已经变得相当便宜。虽然仍然要几千美元，但对于大学或学校来说是可以承受的，如果你真的想在家里深入研究的话也可以。

There's an amazing community around Hugging Face and their LeRobot project, where they have very cheap robot arms and they can help you learn about the whole teleoperation, data collection, training, and deployment pipeline with those arms. So that's a really good way to learn about that. For the more locomotion reinforcement learning aspect, it's a little bit harder because you probably want a robot with legs, which also means that the robot should be able to fall and stand up without completely breaking. I think the best bet there is the Chinese quadrupeds that are getting fairly cheap. It's still multiple thousands of dollars, but it's affordable for a university or a school, or if you really want to go much deeper into that at home as well.

那么，你拿到四足机器人并拆箱后，可以用它做什么？或者从哪里开始尝试做一些实验？

And you know, you get your quadruped and you unbox it. What can you do with it, or where do you start trying to do some experiments with it?

当你拆箱时，它通常已经能做很多事情了。它能走路。有些甚至还有 SLAM 管道，可以进行一些导航、避障等。但挑战在于，你想摆脱所有这些软件，基本上从头开始重建。同样，网上有很多社区，有很多 GitHub 仓库可以帮助你入门。宇树科技的 Go 可能是最标准的平台。所以我会从那里开始，已经有人开源了从训练到部署代码、在这些机器人上部署策略的一切。

So when you unbox it, it can typically already do quite a lot. So it will be able to walk. In some cases they even have things like SLAM pipeline, so it can do some navigation, avoid obstacles, things like that. But then the challenge is that you want to get rid of all that software and basically recreate it from scratch. And there again, there are many communities online. There are many GitHub repos that help you get started. But the Unitree Go is probably the most standard platform. So I would start there, and there are people who open source already everything from training to deploying code, deploying these policies on those robots.

好的，太棒了。Nikita，非常感谢你参加节目，分享你正在做的事情。非常酷。

Okay, cool. Awesome. Well, Nikita, thanks so much for jumping on and sharing a bit about what you're up to. Very cool stuff.

非常感谢。我真的很享受这次对话。

Thank you so much. Really enjoyed this.

谢谢。

Thank you.