System.Xml.XmlElement

The enormous advance in AI capabilities over the last several years has in large part been driven by the availability of vast amounts of training data.

过去几年人工智能能力的巨大进步在很大程度上是由大量训练数据的可用性推动的。

Gwern’s influential 2020 essay “ The Scaling Hypothesis ,” written after the release of GPT-3, argues that recent advances in AI have been the result of scale: bigger models, bigger datasets, and more compute spent training them.

Gwern 在 GPT-3 发布后撰写的颇具影响力的 2020 年论文《扩展假设》认为，人工智能的最新进展是规模化的结果：更大的模型、更大的数据集以及用于训练它们的更多计算。

Gwern writes:

格温写道：

The scaling hypothesis regards the blessings of scale as the secret of AGI: intelligence is ‘just’ simple neural units & learning algorithms applied to diverse experiences at a (currently) unreachable scale.

规模假设将规模的好处视为 AGI 的秘密：智能“只是”简单的神经单元和学习算法，应用于（目前）无法达到的规模的不同体验。

As increasing computational resources permit running such algorithms at the necessary scale, the neural networks will get ever more intelligent.

随着计算资源的增加允许以必要的规模运行此类算法，神经网络将变得更加智能。

Likewise, pseudonymous AI researcher Nostalgebraist wrote in 2024 that “the entire current landscape dates back to the realization that ML models improve steadily and predictably as you train them on more data.

同样，匿名人工智能研究员 Nostalgebraist 在 2024 年写道，“当前的整个形势可以追溯到这样一种认识：当你用更多数据训练机器学习模型时，它们会稳步且可预测地改进。

And that everything since then consists of training bigger models on larger datasets, enabled by years of rapid improvement in hardware.” Epoch AI found that since the early 2010s, the size of training datasets for AI language models has doubled about every six months .

从那时起，一切都包括在更大的数据集上训练更大的模型，这得益于多年来硬件的快速改进。” Epoch AI 发现，自 2010 年代初以来，AI 语言模型的训练数据集大小大约每六个月翻一番。

The biggest AI models today are trained on tens of trillions of tokens.

当今最大的人工智能模型是在数万亿代币上进行训练的。

As more and more effort gets devoted to applying modern AI to robotics, an important question is “where will the training data come from?” Language models could be trained on huge existing corpuses of text.

随着越来越多的努力致力于将现代人工智能应用于机器人技术，一个重要的问题是“训练数据从哪里来？”语言模型可以在现有的大量文本语料库上进行训练。

GPT-3, for instance, was trained on about 300 billion tokens’ worth of data scraped from the internet .

例如，GPT-3 接受了从互联网上抓取的价值约 3000 亿代币的数据的训练。

In 2024 Epoch AI estimated the total existing stock of quality, accessible text data is around 300 trillion tokens .

Epoch AI 估计，到 2024 年，优质、可访问的文本数据的现有存量总量约为 300 万亿个代币。

But there’s no equivalent stock of good, public robot training data.

但没有相应的良好的公共机器人训练数据。

Figuring out where the data for robotics will come from and how much of it we’ll need has become a key problem for the field.

弄清楚机器人技术的数据从哪里来以及我们需要多少数据已经成为该领域的一个关键问题。

Training an AI model, whether it’s an LLM or a model used for controlling robots, fundamentally requires lots and lots of examples of something being done correctly (or enough knowledge of what a correct action looks like that you can score an attempt).

训练人工智能模型，无论是法学硕士还是用于控制机器人的模型，从根本上来说都需要大量正确完成某些事情的示例（或者足够了解正确的动作是什么样子，以便您可以为尝试评分）。

You feed a model some set of inputs, you compare the model’s output to the actual, correct output, and you use the difference between the two to modify the parameters of the model to reduce this difference via some flavor of gradient descent.

您向模型提供一些输入集，将模型的输出与实际的正确输出进行比较，然后使用两者之间的差异来修改模型的参数，以通过某种梯度下降来减少这种差异。

(We went into this process in some depth earlier this month .)

（本月早些时候我们深入研究了这个过程。）

With text, a training example can be just an existing chunk of text; you take the chunk, feed the model part of it, and have the model predict what comes next in the sequence.

对于文本，训练示例可以只是现有的文本块；您获取该块，将其中的一部分提供给模型，然后让模型预测序列中接下来会发生什么。

One chunk of text can thus generate many training examples by feeding the model different portions of the text.

因此，通过向模型提供文本的不同部分，一大块文本可以生成许多训练示例。

To train a robot, an example is typically some sort of recording of a task being accomplished or attempted.

为了训练机器人，一个典型的例子是对正在完成或尝试的任务的某种记录。

There are a variety of different methods used to create these recordings and turn them into training data.

有多种不同的方法可用于创建这些录音并将其转换为训练数据。

Historically, one of the most common methods for creating these recordings is called “teleoperation.” With teleoperation, a person manipulates some controls that are directly connected to a robot, and uses them to move a robot through some particular task (folding a shirt, picking up and moving objects, and so on).

从历史上看，创建这些录音的最常见方法之一称为“远程操作”。通过远程操作，人们可以操纵一些直接连接到机器人的控件，并使用它们来移动机器人完成某些特定任务（折叠衬衫、拾取和移动物体等）。

The image below, for instance, shows a person teleoperating a pair of robot arms, using them to pour water from a pitcher into a glass.

例如，下图显示了一个人远程操作一对机器人手臂，用它们将水从水罐倒入玻璃杯中。

As the person performs the task, data of the robot’s state is recorded, along with images from cameras that are recording the task (often mounted to the robot itself), which are combined to create a training example.

当人执行任务时，机器人状态的数据以及记录任务的摄像机（通常安装在机器人本身上）的图像都会被记录下来，这些图像组合起来创建一个训练示例。

An example might consist of feeding a model some data on the robot’s state and what the cameras can see at some particular point in the sequence.

一个例子可能包括向模型提供一些有关机器人状态的数据以及相机在序列中某个特定点可以看到的数据。

The model will try to predict what the next robot actions are (in the form of something like joint angles), and then compare that to what the actual actions were.

该模型将尝试预测机器人的下一个动作（以关节角度等形式），然后将其与实际动作进行比较。

This sort of training strategy, using examples generated by an expert (a human) and then training a model to mimic them, is known as behavioral cloning , and it’s part of the broader category of “imitation learning.”

这种使用专家（人类）生成的示例，然后训练模型来模仿它们的训练策略被称为行为克隆，它是更广泛的“模仿学习”类别的一部分。

The major benefit of this method is that the data it generates very closely matches the robot’s actual movement capabilities, because it’s produced from the actual robot you want to train, and it remains widely used today.

这种方法的主要好处是，它生成的数据与机器人的实际运动能力非常匹配，因为它是由您想要训练的实际机器人产生的，并且至今仍被广泛使用。

You sometimes see videos or pictures from huge data collection factories , filled with people doing nothing but teleoperating robots to complete various tasks, and you sometimes see teleoperation accidents when the person forgets that they’re controlling the robot and suddenly moves in a way that the robot can’t handle.

有时你会看到来自巨大数据收集工厂的视频或图片，里面充满了人们除了远程操作机器人来完成各种任务之外什么也不做的情况，有时你会看到远程操作事故，因为人们忘记了他们正在控制机器人，突然以机器人无法处理的方式移动。

But teleoperation has drawbacks.

但远程操作也有缺点。

The biggest one is probably expense: to get the data, you need to have an actual robot, along with a person to sit there and manipulate it, which makes it hard to scale this method to produce very large amounts of data.

最大的一个可能是费用：要获取数据，您需要有一个真正的机器人，以及一个坐在那里操纵它的人，这使得很难扩展这种方法以产生大量数据。

Another drawback is that it’s somewhat clunky for a person to control a robot this way; they often can’t get any sort of tactile feedback from the object being manipulated to guide their motions, so this won’t necessarily produce really good data on dexterous manipulation.

另一个缺点是，人用这种方式控制机器人有点笨拙。他们通常无法从被操纵的物体上获得任何形式的触觉反馈来指导他们的动作，因此这不一定会产生关于灵巧操纵的真正好的数据。

The clunkiness can also make it hard to collect good teleoperation data, and as I understand it, only a fraction of total teleoperation data collected ends up being useful for training.

这种笨拙还使得收集良好的远程操作数据变得困难，据我了解，收集到的全部远程操作数据中只有一小部分最终对训练有用。

It’s also possible to generate training data using an existing robot by having it move around on its own, recording its actions, and using some sort of system to categorize these behavioral sequences as successful or not.

还可以使用现有的机器人生成训练数据，方法是让机器人自行移动，记录其动作，并使用某种系统将这些行为序列分类为成功或失败。

This sort of strategy, where the robot moves autonomously and you score its actions in some way, is known as “reinforcement learning.” Google was doing something like this around 2016 with an “arm farm”: a series of robot arms moving autonomously , with success scored by whether an arm picked up or grasped an object or not.

这种机器人自主移动并以某种方式对其动作进行评分的策略被称为“强化学习”。谷歌在 2016 年左右做了类似的事情，推出了一个“手臂农场”：一系列自主移动的机器人手臂，成功与否取决于手臂是否拾起或抓住物体。

This sort of arm farm doesn’t seem to be widely used today, but using real-world deployed robots will probably eventually be the largest source of robot training data.

这种手臂农场如今似乎并未得到广泛使用，但使用现实世界中部署的机器人最终可能会成为机器人训练数据的最大来源。

You often hear robotics folks talk about the need for a “data flywheel,” where you have robots deployed in the real world, you use the data from them to train even better models, resulting in more robot deployments, resulting in better models, and so on.

你经常听到机器人技术人员谈论对“数据飞轮”的需求，即在现实世界中部署机器人，使用它们的数据来训练更好的模型，从而部署更多的机器人，从而产生更好的模型，等等。

Robotic AI company Physical Intelligence, for instance, is experimenting with this sort of “deploy a robot to do real work, and use it to gather data to improve the model” proto-flywheel at a chocolate factory near its offices.

例如，机器人人工智能公司Physical Intelligence正在其办公室附近的一家巧克力工厂试验这种“部署机器人来做实际工作，并用它收集数据来改进模型”的原型飞轮。

But while this might work eventually, right now there’s a very small number of AI-controlled robots doing anything resembling real work.

不过，虽然这最终可能会奏效，但目前只有极少数人工智能控制的机器人在做类似实际工作的事情。

Real-world data also tends to be extremely repetitive — robots doing the same task over and over again in very similar conditions — which makes it much less useful for training.

现实世界的数据也往往极其重复——机器人在非常相似的条件下一遍又一遍地执行相同的任务——这使得它对于训练的用处要小得多。

So for now most data needs to come from other sources.

因此，目前大多数数据需要来自其他来源。

Another very common strategy for generating robotic training data is simulation: instead of running a physical robot in the real world, run a digital one in a simulated world instead.

生成机器人训练数据的另一种非常常见的策略是模拟：不是在现实世界中运行物理机器人，而是在模拟世界中运行数字机器人。

This is also commonly done with reinforcement learning, in which you have the robot try to navigate some simulated space, after which you score it on how well it does.

这通常也是通过强化学习来完成的，在强化学习中，你让机器人尝试在一些模拟空间中导航，然后你根据它的表现进行评分。

Another flavor of this is the teacher/student paradigm where you have a “teacher” model that has full access to all the data within the simulated environment, and a “student” model that can only see what the real robot would see (images from its cameras and its own internal state), and you try to get the student to mimic the teacher.

另一种风格是教师/学生范式，其中你有一个“教师”模型，可以完全访问模拟环境中的所有数据，以及一个只能看到真实机器人会看到的内容（来自其相机的图像和其自身内部状态）的“学生”模型，并且你尝试让学生模仿教师。

The benefits of simulation are that it’s much cheaper and faster to simulate a robot than it is to deploy one in the real world, and that you can get very large amounts of data by simulating the behavior of thousands of robots.

模拟的好处在于，模拟机器人比在现实世界中部署机器人便宜得多、速度快得多，而且您可以通过模拟数千个机器人的行为来获取大量数据。

Figure AI taught its humanoid robot to walk via reinforcement learning on thousands of simulated robots .

Figure AI 通过对数千个模拟机器人进行强化学习，教会其人形机器人行走。

OpenAI taught a robotic hand to solve a Rubik’s Cube via the equivalent of 13,000 years of simulated experience .

OpenAI 通过相当于 13,000 年的模拟经验，教会了机械手解决魔方问题。

By 2019, Waymo had driven 10 million miles in the real world, but 10 billion miles in simulation .

到 2019 年，Waymo 在现实世界中行驶了 1000 万英里，在模拟中行驶了 100 亿英里。

But simulation also has drawbacks.

但模拟也有缺点。

The biggest one is that the quality of the data it produces is a function of the accuracy of the simulation.

最重要的一点是，它生成的数据的质量是模拟精度的函数。

For some things, such as locomotion (having a robot move through some physical space), simulation is fairly accurate, because the behavior is mostly well-characterized enough that we can represent it digitally.

对于某些事情，例如运动（让机器人在某个物理空间中移动），模拟相当准确，因为行为大多具有足够的特征，我们可以用数字方式表示它。

So you can train a quadruped robot to navigate some terrain, or a bipedal robot to dance, or a drone to fly along some path in simulation, and the simulated environment will be a reasonably good approximation of the real world.

因此，您可以训练四足机器人在某些地形中导航，或者双足机器人跳舞，或者无人机沿着模拟路径飞行，并且模拟环境将与现实世界相当接近。

But for things such as object manipulation (particularly for soft or deformable objects), it’s much harder to build a simulation that accurately captures real-world behavior, and using it as a source of training data is much more difficult.

但对于诸如对象操纵（特别是对于柔软或可变形对象）之类的事情，构建准确捕获现实世界行为的模拟要困难得多，并且将其用作训练数据源也更加困难。

This is generally referred to as the “ sim-to-real gap .”

这通常被称为“模拟与真实的差距”。

Because of simulation inaccuracies, one popular simulation technique is called “domain randomization.” Since you don’t know what various parameters like friction will be in the real world, you decide to vary the parameters randomly when running the simulation, in the hopes of creating a training envelope that encompasses the parameters that will be encountered in the real world.

由于模拟不准确，一种流行的模拟技术称为“域随机化”。由于您不知道现实世界中的各种参数（例如摩擦力）是什么，因此您决定在运行模拟时随机改变参数，希望创建一个包含现实世界中会遇到的参数的训练范围。

When OpenAI trained its robotic hand to solve a Rubik’s Cube, for instance, it continuously expanded the range of parameters being tested, forcing the model to learn to solve the cube in more and more difficult environments, in the hopes of giving it capabilities that would generalize to the real world.

例如，当 OpenAI 训练其机械手解魔方时，它不断扩大所测试的参数范围，迫使模型学习在越来越困难的环境中解魔方，希望赋予其能够推广到现实世界的能力。

The robotics team there used this strategy specifically because accurate physical modeling was too difficult :

那里的机器人团队专门使用了这种策略，因为准确的物理建模太困难了：

The biggest challenge we faced was to create environments in simulation diverse enough to capture the physics of the real world.

我们面临的最大挑战是在模拟中创建足够多样化的环境来捕捉现实世界的物理现象。

Factors like friction, elasticity and dynamics are incredibly difficult to measure and model for objects as complex as Rubik’s Cubes or robotic hands and we found that domain randomization alone is not enough.

对于像魔方或机器人手这样复杂的物体，摩擦、弹性和动力学等因素非常难以测量和建模，我们发现仅域随机化是不够的。

To overcome this, we developed a new method called Automatic Domain Randomization (ADR), which endlessly generates progressively more difficult environments in simulation.

为了克服这个问题，我们开发了一种称为自动域随机化 (ADR) 的新方法，它可以在模拟中不断生成越来越困难的环境。

This frees us from having an accurate model of the real world, and enables the transfer of neural networks learned in simulation to be applied to the real world.

这使我们摆脱了现实世界的精确模型，并使在模拟中学到的神经网络能够迁移到现实世界。

But despite this technique, and really quite a lot of simulation (13,000 simulated years), performance was only OK — for maximally difficult starting arrangements, success was only 20% — which is maybe why this technique doesn’t seem to be widely used for manipulation training currently.

但是，尽管采用了这种技术，并且进行了大量的模拟（13,000 次模拟年），但性能还算可以——对于难度最大的起始安排，成功率只有 20%——这也许就是为什么这种技术目前似乎没有广泛用于操纵训练的原因。

Environmental parameter randomization does seem to be fairly standard for locomotion training though, used for things like randomly varying the coefficient of friction of the surface being traversed.

不过，环境参数随机化似乎确实是运动训练的相当标准，用于随机改变所经过表面的摩擦系数等。

Another way to use simulation is to create synthetic training data: taking real training data as an input and using it to create lots of similar training examples.

使用模拟的另一种方法是创建合成训练数据：将真实的训练数据作为输入，并使用它来创建大量类似的训练示例。

Nvidia, for instance, has a page showing how you can use its Isaac GR00T robot simulation platform to generate synthetic robot training data using teleoperation data as an input.

例如，Nvidia 有一个页面展示了如何使用其 Isaac GR00T 机器人模拟平台，以远程操作数据作为输入来生成合成机器人训练数据。

Synthetic data generation also seems to be widely used for training autonomous vehicles .

合成数据生成似乎也广泛用于训练自动驾驶车辆。

We can also create robot training data by recording humans doing tasks and creating some sort of mapping from the recorded task data to a robot.

我们还可以通过记录人类执行任务并创建从记录的任务数据到机器人的某种映射来创建机器人训练数据。

This might be as simple as fitting a person with a head-mounted camera (creating what’s known as “egocentric video”), and using it to record the person doing various tasks with their hands.

这可能就像给一个人安装头戴式摄像头（创建所谓的“以自我为中心的视频”）一样简单，并用它来记录该人用手执行各种任务。

The video is then processed so the person’s motions are interpreted as robot motions, and used to create training examples.

然后对视频进行处理，将人的动作解释为机器人动作，并用于创建训练示例。

(This is one reason for the popularity of the humanoid form factor in robotics: humans doing tasks will be a reasonably good source of training data.) Figure AI, for instance, recently announced Index, its program for gathering training data by sending people head-mounted cameras to record themselves doing various tasks (alternatively, Figure will also send a crew equipped with these cameras to your home or business to do chores for you).

（这是人形外形在机器人中流行的原因之一：执行任务的人类将成为相当好的训练数据来源。）例如，Figure AI 最近宣布了 Index，它的程序是通过向人们发送头戴式摄像头来记录自己执行各种任务的情况来收集训练数据（或者，Figure 还将派遣配备这些摄像头的工作人员到您的家中或企业为您做家务）。

Companies like Shift are also doing this, sending people with hat-mounted cameras to folks’ houses to record themselves doing chores.

Shift 等公司也在这样做，派人带着戴着帽子的摄像机到人们家中记录自己做家务的情况。

You can also add more sensors to collect more elaborate human task data.

您还可以添加更多传感器来收集更详细的人工任务数据。

The now-defunct startup Eidon AI stuck sensors on folks’ arms to collect data on joint positions for training robot arms.

现已解散的初创公司 Eidon AI 将传感器贴在人们的手臂上，收集关节位置的数据，用于训练机器人手臂。

Neocambrian AI recently announced “Mitts,” which it describes as “state of the art tactile gloves that capture high-precision touch data for robots.”

Neocambrian AI 最近发布了“Mitts”，它将其描述为“最先进的触觉手套，可为机器人捕获高精度触摸数据”。

Another flavor of this strategy is to use what’s called a Universal Manipulation Interface , or UMI.

该策略的另一种风格是使用所谓的通用操作界面（UMI）。

With UMI, a person holds a pair of specially designed manipulators, that typically have a simple pincer grip at the end (though some have multiple fingers ).

使用 UMI，一个人握住一对专门设计的机械手，通常在末端有一个简单的钳形握把（尽管有些机械手有多个手指）。

The person does various tasks with the manipulators, which have cameras mounted to them that record exactly what the manipulators are doing.

人们使用操纵器执行各种任务，操纵器上安装有摄像头，可以准确记录操纵器正在做什么。

This video is then used to create training examples for robots that have similar manipulators.

然后，该视频用于为具有类似操纵器的机器人创建训练示例。

Robot model startups Generalist AI and Sunday Robotics , for instance, seem to be primarily using this strategy for gathering training data.

例如，机器人模型初创公司 Generalist AI 和 Sunday Robotics 似乎主要使用这种策略来收集训练数据。

The benefit of using human-collected task data is that it’s much cheaper to collect than teleoperation data: all you need is a camera, or some camera-equipped manipulators, rather than an entire robot.

使用人类收集的任务数据的好处是，它的收集成本比遥操作数据要便宜得多：您所需要的只是一个摄像头或一些配备摄像头的操纵器，而不是整个机器人。

And it’s probably much easier to get a wide variety of different task recordings by just having a person walk around and do different things, rather than constantly repositioning a robot to do them.

只需让一个人四处走动并做不同的事情，而不是不断地重新定位机器人来完成这些任务，就可能更容易获得各种不同的任务记录。

And because the data is gathered in the real world, by people manipulating real objects, it’s more accurate than simulation.

由于数据是在现实世界中通过人们操纵真实物体收集的，因此它比模拟更准确。

The drawback is that because this data comes from a person rather than a robot, it won’t map onto a robot’s capabilities perfectly well.

缺点是，由于这些数据来自人而不是机器人，因此它不能很好地映射到机器人的能力。

Even UMI, which minimizes the human-robot differences by having a person do tasks using a sort of robotic manipulator, won’t necessarily map perfectly, though the discrepancy is small and many of the most impressive robotics demos we’ve seen (such as Generalist AI’s ) have been trained using UMI data.

即使 UMI 通过让人使用某种机器人操纵器执行任务来最小化人类与机器人的差异，也不一定能完美映射，尽管差异很小，而且我们见过的许多最令人印象深刻的机器人演示（例如 Generalist AI ）都是使用 UMI 数据进行训练的。

It’s possible to use multiple training strategies in a single robot model.

可以在单个机器人模型中使用多种训练策略。

Agility Robotics’ Digit , for instance, walks using a small, 1-million-parameter neural network trained in simulation using reinforcement learning, while object manipulation seems to have been trained using behavioral cloning.

例如，Agility Robotics 的 Digit 使用一个小型的 100 万参数神经网络行走，该神经网络通过强化学习进行模拟训练，而对象操作似乎是使用行为克隆进行训练的。

Dyna Robotics similarly seems to have used a combination of reinforcement learning in simulation and human video for training data.

类似地，Dyna Robotics 似乎也结合了模拟中的强化学习和人类视频的训练数据。

It’s also often useful to train a robot model using data other than task completion data.

使用任务完成数据以外的数据来训练机器人模型通常也很有用。

Vision-Language-Action models, as we noted earlier , are built from Vision-Language models trained on text and images.

正如我们之前提到的，视觉-语言-动作模型是根据在文本和图像上训练的视觉-语言模型构建的。

And because of the ubiquity of general video data on things like YouTube or TikTok, there’s a great deal of interest in finding ways to use generic web videos as a source of training data.

由于 YouTube 或 TikTok 等通用视频数据无处不在，人们对寻找使用通用网络视频作为训练数据源的方法非常感兴趣。

Nvidia’s Cosmos , for instance, is a foundation model trained on video data that can be used as a backbone for a robot policy model.

例如，Nvidia 的 Cosmos 是一个基于视频数据训练的基础模型，可以用作机器人策略模型的骨干。

How much of this action-relevant training data have we gathered so far, and how much do we need?

到目前为止，我们已经收集了多少与行动相关的培训数据，我们需要多少？

As with AI models more generally, robotic training datasets seem to be getting bigger.

与更普遍的人工智能模型一样，机器人训练数据集似乎越来越大。

An early dataset that was considered large at the time was the Open X-Embodiment (OXE) dataset, created by Google DeepMind in 2023 by stitching together dozens of existing datasets.

当时被认为很大的早期数据集是 Open X-Embodiment (OXE) 数据集，由 Google DeepMind 在 2023 年通过将数十个现有数据集拼接在一起创建。

OXE consisted of over 1 million robot action trajectories spanning more than 160,000 different tasks, totaling around 4,000 hours’ worth of task data.

OXE 由超过 100 万条机器人动作轨迹组成，涵盖超过 160,000 个不同的任务，总计约 4,000 小时的任务数据。

Physical Intelligence’s Pi 0 model , released in October of 2024, was trained on 10,000 hours of robot task data.

Physical Intelligence 的 Pi 0 模型于 2024 年 10 月发布，接受了 10,000 小时的机器人任务数据训练。

Generalist AI’s GEN 1 model, announced earlier this year, was trained on around 500,000 hours of UMI data.

今年早些时候宣布的 Generalist AI GEN 1 模型接受了约 500,000 小时的 UMI 数据训练。

Figure AI’s Index announcement stated that it was gathering 30 minutes of video data every second, or 43,000 hours a day.

Figure AI 的 Index 公告称，它每秒收集 30 分钟的视频数据，即每天 43,000 小时。

Dyna Robotics and Nature Will both claim that their latest models are trained on over 1 million hours of human egocentric video data.

Dyna Robotics 和 Nature Will 都声称他们的最新模型是根据超过 100 万小时的人类自我中心视频数据进行训练的。

I saw one estimate that the robot data collection industry is collectively producing around 500,000 hours of training data a month.

我看到一项估计，机器人数据收集行业每月总共产生约 500,000 小时的训练数据。

It’s not amazingly easy to compare these datasets to the text datasets used to train LLMs.

将这些数据集与用于训练法学硕士的文本数据集进行比较并不容易。

For one, different robotic training sets have different levels of quality — teleoperation or UMI data is probably higher quality than pure human egocentric video data for training purposes, and a highly variable dataset with lots of different sorts of tasks is much more useful than an extremely repetitive one.

其一，不同的机器人训练集具有不同的质量水平——远程操作或 UMI 数据可能比用于训练目的的纯人类自我中心视频数据的质量更高，并且包含许多不同类型任务的高度可变的数据集比极其重复的数据集有用得多。

And video data for robotics training is informationally different from text data: for one, video data tends to be highly correlated — on short timescales, one frame is pretty similar to another frame.

用于机器人训练的视频数据在信息上与文本数据不同：首先，视频数据往往高度相关——在很短的时间尺度上，一帧与另一帧非常相似。

Thus even though video recordings contain a lot of data, they’re probably less informationally dense for training purposes than text data.

因此，尽管视频记录包含大量数据，但出于训练目的，它们的信息密度可能不如文本数据。

As robotics researcher Chris Paxton notes , “ machine learning works best the closer you are to an independent and identically distributed (iid) dataset; neither web data nor robot data is iid, but robot data is far less iid than web data.

正如机器人研究员 Chris Paxton 指出的那样，“机器学习越接近独立同分布 (iid) 数据集，效果越好；网络数据和机器人数据都不是独立同分布的，但机器人数据的独立同分布程度远低于网络数据。

”

”

But we can still try.

但我们仍然可以尝试。

One very rough way to compare dataset sizes, used by Physical Intelligence researcher Kevin Black in this presentation and robotics professor Ken Goldberg in this presentation , is to estimate how long a dataset of text tokens would take for a person to read, and compare that to the hours of robot training data we have.

物理智能研究员 Kevin Black 在本次演示中和机器人学教授 Ken Goldberg 在本次演示中使用的一种比较数据集大小的非常粗略的方法是估计一个人阅读文本标记数据集需要多长时间，并将其与我们拥有的机器人训练数据的时间进行比较。

Under this metric, GPT-2 was trained on the equivalent of around 475,000 hours of data, GPT-3 was trained on 26 million hours, and GPT-4 was trained on 685 million hours.

在此指标下，GPT-2 接受了相当于约 475,000 小时数据的训练，GPT-3 接受了 2600 万小时的训练，GPT-4 接受了 6.85 亿小时的训练。

Using this (somewhat dubious) metric, even several-year-old AI models were trained on something like 1,000 times the data that current robot models are trained on.

使用这个（有点可疑的）指标，即使是已有几年历史的人工智能模型所训练的数据量也是当前机器人模型所训练数据量的 1000 倍。

Goldberg calls this the “ 100,000-year data gap .”

戈德堡称之为“10万年的数据差距”。

On his Substack “It Can Think!” Paxton makes an estimate from the other direction , converting hours of robot data into tokens, and comes up with a fairly similar estimate.

在他的子堆栈上“它可以思考！”帕克斯顿从另一个方向进行了估计，将机器人的数小时数据转换为标记，并得出了相当相似的估计。

He guesses that for video running at 10 frames per second, each frame will contain the equivalent of one useful training token.

他猜测，对于每秒 10 帧运行的视频，每一帧将包含相当于一个有用的训练令牌。

Every hour of video thus contains about 10 × 60 × 60 = 36,000 training tokens.

因此，每个小时的视频大约包含 10 × 60 × 60 = 36,000 个训练标记。

He then divides this by 10, on the assumption that you can’t expect every second of video data to actually be useful, getting you about 3,600 training tokens per hour of video.

然后，他将其除以 10，假设您不能指望每一秒的视频数据实际上都是有用的，这样您就可以得到每小时视频约 3,600 个训练令牌。

This would make Dyna Robotics’s 1 million hours of training equal to around 3.6 billion tokens, on the order of 5,000 times smaller than the datasets used to train recent AI models.

这将使 Dyna Robotics 的 100 万小时训练相当于约 36 亿个代币，比用于训练最新 AI 模型的数据集小约 5,000 倍。

If you assume that we need about 1,000 times as much real-world robot training data as we actually have — on the order of 1 billion hours of data — how much would that cost to generate?

如果您假设我们需要大约 1000 倍于实际机器人训练数据的数据（大约 10 亿小时的数据），那么生成数据的成本是多少？

Eidon AI reports robot data prices of $60 to $100 per hour.

Eidon AI 报告机器人数据价格为每小时 60 至 100 美元。

Pantheon similarly reports UMI data costs from $60 to $150 per hour, but that its engineers developed a pipeline capable of generating data for $10 per hour.

Pantheon 类似地报告 UMI 数据成本为每小时 60 美元到 150 美元，但其工程师开发了一个能够以每小时 10 美元生成数据的管道。

Shift claims you can earn up to $20 per hour recording egocentric video data and selling it to them.

Shift 声称，录制以自我为中心的视频数据并将其出售给他们，每小时可赚取高达 20 美元的收入。

So a naive estimate suggests that getting enough data to get robot models to a similar level of capability as recent AI models might cost on the order of $10 billion to $100 billion.

因此，天真的估计表明，获取足够的数据以使机器人模型达到与最新人工智能模型相似的能力水平可能需要花费 100 亿至 1000 亿美元。

How much data we actually need, however, seems to be something of an open question.

然而，我们实际需要多少数据似乎是一个悬而未决的问题。

If we don’t care about generalization, and only want to train robot models to perform various specific tasks, the data requirements might be substantially less.

如果我们不关心泛化，只想训练机器人模型来执行各种特定任务，那么数据需求可能会少得多。

Alternatively, it might turn out that frontier general AI models like Fable or Astra are extremely capable robot controllers after being fed a comparatively small amount of robotic training data; some folks are already predicting that special-purpose robot AI models might get replaced with more generic AI models.

或者，事实证明，像《Fable》或《Astra》这样的前沿通用人工智能模型在输入相对少量的机器人训练数据后，可能会成为功能非常强大的机器人控制器；一些人已经预测专用机器人人工智能模型可能会被更通用的人工智能模型所取代。

Likewise, we might find robotic AI model architectures that are much more sample efficient — learning much more from each training example — than current models.

同样，我们可能会发现机器人人工智能模型架构比当前模型的样本效率更高——从每个训练示例中学到更多东西。

On the other hand, it could be the case that moving around and manipulating objects in the real world is a fundamentally more difficult task than manipulating text on a computer, and that models require a lot more training to become competent at this real-world behavior.

另一方面，在现实世界中移动和操作对象可能比在计算机上操作文本根本上是一项更困难的任务，并且模型需要更多的训练才能胜任这种现实世界的行为。

Small errors in generated text, for instance — misspellings, or slightly wrong word choice — don’t necessarily compromise the text as a whole, but when you’re manipulating objects a single tiny positioning error can be the difference between success and complete failure.

例如，生成文本中的小错误（拼写错误或稍微错误的单词选择）不一定会影响整个文本，但当您操作对象时，一个微小的定位错误可能会导致成功和彻底失败。

Likewise, Ken Goldberg argues that robotic movement is a much higher-dimensional problem than generating text (though I’m not sure I buy the version of the argument he articulates in that presentation, but I have seen this sentiment echoed by other robotics researchers).

同样，肯·戈德堡（Ken Goldberg）认为，机器人运动是一个比生成文本更高维的问题（尽管我不确定我是否认同他在演讲中阐述的观点，但我看到其他机器人研究人员也认同这种观点）。

LLMs showed the benefits of scale when training AI models — bigger models, more data, trained using much more computation.

法学硕士在训练人工智能模型时展示了规模化的好处——更大的模型、更多的数据、使用更多的计算进行训练。

We’re now in the process of seeing how well this scaling works for robotics, and robotics researchers are working hard to gather enormous amounts of robot task training data in the hopes of seeing the same sort of trends with robotic capabilities that we have with AI more generally.

我们现在正在了解这种扩展对机器人技术的效果如何，机器人研究人员正在努力收集大量的机器人任务训练数据，希望看到机器人能力与我们更普遍的人工智能所具有的相同趋势。