How AI Will Step Off the Screen and into the Real World | Daniela Rus | TED

翻译人员: Yip Yan Yeung

When I was a student studying robotics,

当我还是个研究机器人的学生时，

a group of us decided to make a present for our professor's birthday.

我们一群人决定为我们的教授 做一份生日礼物。

We wanted to program our robot to cut a slice of cake for him.

我们想给机器人写个程序， 让它为他切一片蛋糕。

We pulled an all-nighter writing the software,

我们通宵写了这个软件，

and the next day, disaster.

第二天就翻车了。

We programmed this robot to cut a soft, round sponge cake,

我们为这个机器人写了程序， 让它切一个柔软的圆形海绵蛋糕，

but we didn't coordinate well.

但我们没协调好。

And instead, we received a square hard ice cream cake.

我们拿到了一个 方形的坚硬冰淇淋蛋糕。

The robot flailed wildly and nearly destroyed the cake.

机器人张牙舞爪， 差点毁了这个蛋糕。

(Laughter)

（笑声）

Our professor was delighted, anyway.

不管怎样，我们的教授很高兴。

He calmly pushed the stop button

他冷静地按下了停止按钮，

and declared the erratic behavior of the robot

宣布机器人的异常行为

a control singularity.

是一个“控制异变”。

A robotics technical term.

机器人技术术语。

I was disappointed, but I learned a very important lesson.

我很失望， 但我吸取了非常重要的教训。

The physical world,

物理世界

with its physics laws and imprecisions,

有其物理定律和不准确度，

is a far more demanding space than the digital world.

是一个比数字世界要求更严格的空间。

Today, I lead MIT's Computer Science and AI lab,

如今，我领导着麻省理工学院的 计算机科学和人工智能实验室，

the largest research unit at MIT.

这是 MIT 最大的研究单位。

This is our buildingm where I work with brilliant and brave researchers

这是我们的大楼，我在这里 与优秀勇敢的研究人员合作，

to invent the future of computing and intelligent machines.

共同发明计算机 和智能机器的未来。

Today in computing,

如今，在计算领域，

artificial intelligence and robotics are largely separate fields.

人工智能和机器人技术 是泾渭分明的两个领域。

AI has amazed you with its decision-making and learning,

AI 用决策和学习惊艳了你，

but it remains confined inside computers.

但它仍然被困在计算机里。

Robots have a physical presence and can execute pre-programmed tasks,

机器人有物理的实体， 可以执行预先编程的任务，

but they're not intelligent.

但它们并不智能。

Well, this separation is starting to change.

这种隔阂要开始改变了。

AI is about to break free from the 2D computer screen interactions

AI 即将摆脱二维计算机屏幕的交互，

and enter a vibrant, physical 3D world.

进入一个充满活力的物理三维世界。

In my lab, we're fusing the digital intelligence of AI

在我的实验室里， 我们正在将 AI 的数字智能

with the mechanical prowess of robots.

与机器人的机械能力融为一体。

Moving AI from the digital world into the physical world

将 AI 从数字世界转移到物理世界

is making machines intelligent

正在让机器变得智能，

and leading to the next great breakthrough,

带来下一个重大突破，

what I call physical intelligence.

我称之为“物理智能”。

Physical intelligence is when AI's power to understand text,

“物理智能”是指利用 AI 理解文本、

images and other online information

图像和其他在线信息的能力，

is used to make real-world machines smarter.

使现实世界中的机器变得更智能。

This means AI can help pre-programmed robots do their tasks better

这意味着 AI 可以从数据中获取知识，

帮助预编程的机器人更好地完成任务。

by using knowledge from data.

With physical intelligence,

借助物理智能，

AI doesn't just reside in our computers,

AI 不仅存在于我们的计算机中，

but walks, rolls, flies

还能行走、滚动、飞翔、 以惊人的方式与我们互动。

and interacts with us in surprising ways.

Imagine being surrounded by helpful robots at the supermarket.

想象一下，在超市里 周围都是能干的机器人。

The one on the left can help you carry a heavy box.

左边的那个可以帮你 搬一个沉重的箱子。

To make it happen, we need to do a few things.

为了实现这一目标， 我们得做几件事。

We need to rethink how machines think.

我们得重新思考 机器人是如何思考的。

We need to reorganize how they are designed and how they learn.

我们需要重构 它们的设计方式和学习方式。

So for physical intelligence,

对于物理智能来说，

AI has to run on computers that fit on the body of the robot.

AI 必须运行于可以适配 机器人机身的计算机上。

For example, our soft robot fish.

例如，我们的软体机器鱼。

Today's AI uses server farms that do not fit.

现在 AI 使用的服务器群 就放不进去。

Today's AI also makes mistakes.

如今的 AI 也会犯错误。

This AI system on a robot car does not detect pedestrians.

机器人汽车上的这个 AI 系统 无法检测到行人。

For physical intelligence,

对于物理智能，

we need small brains that do not make mistakes.

我们需要不会犯错误的小大脑。

We're tackling these challenges using inspiration

我们受到了一种名为 “秀丽隐杆线虫”的启发，

from a worm called C. elegans

应对这样的挑战。

In sharp contrast to the billions of neurons in the human brain,

与人脑中数十亿个神经元 形成鲜明对比的是，

C. elegans has a happy life on only 302 neurons,

秀丽隐杆线虫仅靠 302 个神经元过着快乐的生活，

and biologists understand the math of what each of these neurons do.

生物学家了解 每个神经元功能的奥秘。

So here's the idea.

于是有了这样的想法。

Can we build AI using inspiration from the math of these neurons?

我们可以通过这些神经元的奥秘 启发我们创造 AI 吗？

We have developed, together with my collaborators and students,

我与我的合作者和学生们 一起开发了

a new approach to AI we call “liquid networks.”

一种新的 AI 方法， 我们称之为“液态网络”。

And liquid networks results in much more compact

液态网络可以拿出 比当今传统的 AI 解决方案

and explainable solutions than today's traditional AI solutions.

更紧凑、更可解释的方案。

Let me show you.

我来给你看看。

This is our self-driving car.

这是我们的自动驾驶汽车。

It's trained using a traditional AI solution,

它是通过传统的 AI 解决方案训练的，

the kind you find in many applications today.

如同现在很多应用采取的方式。

This is the dashboard of the car.

这是汽车的仪表板。

In the lower right corner, you'll see the map.

右下角可以看到地图。

In the upper left corner, the camera input stream.

左上角是摄像机输入信号。

And the big box in the middle with the blinking lights

中间那个闪烁着亮光的大盒子

is the decision-making engine.

是决策引擎。

It consists of tens of thousands of artificial neurons,

它由成千上万个人造神经元组成，

and it decides how the car should steer.

决定了汽车应该如何转向。

It is impossible to correlate the activity of these neurons

不可能将这些神经元的活动

with the behavior of the car.

与汽车的行为相关联。

Moreover, if you look at the lower left side,

此外，请看左下角，

you see where in the image this decision-making engine looks

你会看到这个决策引擎 会根据图像上的什么位置，

to tell the car what to do.

告诉汽车该怎么做。

And you see how noisy it is.

可以看到噪声有多大。

And this car drives by looking at the bushes and the trees

这辆车是通过观察 路边的灌木和树木来行驶的。

on the side of the road.

That's not how we drive.

我们不是这样开车的。

People look at the road.

人们会看着路。

Now contrast this with our liquid network solution,

我们来将其与我们的 液态网络解决方案对比，

which consists of only 19 neurons rather than tens of thousands.

我们的方案只有 19 个神经元， 而不是成千上万个。

And look at its attention map.

来看一下它的注意力图。

It's so clean and focused on the road horizon

非常干净， 关注水平路面和路边。

and the side of the road.

Because these models are so much smaller,

由于这些模型要小得多，

we actually understand how they make decisions.

我们其实是了解它们的决策方式的。

So how did we get this performance?

我们是如何得到这样的表现的呢？

Well, in a traditional AI system,

在传统的 AI 系统中，

the computational neuron is the artificial neuron,

计算神经元是人工神经元，

and the artificial neuron is essentially an on/off computational unit.

而人工神经元本质上 是一个开/关的计算单元。

It takes in some numbers, adds them up,

输入一些数字，加起来，

applies some basic math

应用一些基本的数学运算，

and passes along the result.

传递结果。

And this is complex

这很复杂，

because it happens across thousands of computational units.

因为它发生在 数千个计算单元上。

In liquid networks,

在液态网络中，

we have fewer neurons,

我们的神经元较少，

but each one does more complex math.

但每个神经元都进行更复杂的数学运算。

Here's what happens inside our liquid neuron.

以下是我们的液体神经元 内部发生的事情。

We use differential equations to model the neural computation

我们使用微分方程对神经计算 和人工突触进行建模。

and the artificial synapse.

And these differential equations

而这些微分方程

are what biologists have mapped for the neural structure of the worms.

就是生物学家 为蠕虫的神经结构建立的方程。

We also wire the neurons differently to increase the information flow.

我们还为增大信息流 重新联结了神经元。

Well, these changes yield phenomenal results.

这些改变产生了惊人的结果。

Traditional AI systems are frozen after training.

传统的 AI 系统 在训练后会被冻结。

That means they cannot continue to improve

也就是说当我们把它们 部署在物理世界的自然环境中时，

when we deploy them in a physical world in the wild.

它们是无法被持续改进的。

We just wait for the next release.

我们只能等待下一个版本。

Because of what's happening inside the liquid neuron,

由于液态神经元的内部运作，

liquid networks continue to adapt after training

液态网络会在训练之后

based on the inputs that they see.

根据看到的输入持续调整。

Let me show you.

我来给你看。

We trained traditional AI and liquid networks

我们训练了传统的 AI 和液态网络，

using summertime videos like these ones,

基于这种夏季的视频，

and the task was to find things in the woods.

任务是在森林里找东西。

All the models learned how to do the task in the summer.

所有的模型都在夏天 学会了如何完成任务。

Then we tried to use the models on drones in the fall.

然后到了秋天， 我们尝试在无人机上使用这些模型。

The traditional AI solution gets confused by the background.

传统的 AI 解决方案 被背景搞糊涂了。

Look at the attention map, cannot do the task.

看看注意力图， 无法完成任务。

Liquid networks do not get confused by the background

液态网络不会被背景搞糊涂，

and very successfully execute the task.

可以非常成功地执行任务。

So this is it.

就是这样。

This is the step forward:

这就是向前迈出的一步：

AI that adapts after training.

训练后还会适应的 AI。

Liquid networks are important

液态网络之所以重要，

because they give us a new way of getting machines to think

是因为它们为我们提供了 一种让机器思考的新方式，

that is rooted into physics models,

这种方式植根于物理模型，

a new technology for AI.

一种新的 AI 技术。

We can run them on smartphones, on robots,

我们可以将它们 运行在智能手机、机器人、

on enterprise computers,

企业计算机上，

and even on new types of machines

甚至是在那些我们现在可以 开始想象和设计的新型机器上。

that we can now begin to imagine and design.

The second aspect of physical intelligence.

物理智能的第二个方面。

So by now you've probably generated images using text-to-image systems.

到目前为止，你可能已经 用文本生成图像系统生成了图片。

We can also do text-to-robot,

我们也能做到“文本生成机器人”，

but not using today's AI solutions because they work on statistics

但没有利用现在的 AI 解决方案， 因为它们基于统计数字，

and do not understand physics.

不懂物理学。

In my lab,

在我的实验室，

we developed an approach that guides the design process

我们开发了一种方法 引导设计过程，

by checking and simulating the physical constraints for the machine.

方法是检查和模拟机器的物理限制。

We start with a language prompt,

我们从语言提示开始，

"Make me a robot that can walk forward,"

“让我成为一个能向前行走的机器人”，

and our system generates the designs including shape, materials, actuators,

然后我们的系统生成的设计 包括形状、材料、驱动器、

sensors, the program to control it

传感器、控制它的程序

and the fabrication files to make it.

和用于制造它的制造文件。

And then the designs get refined in simulation

然后，在仿真中对设计进行完善，

until they meet the specifications.

直到它们符合规格。

So in a few hours we can go from idea

在几个小时内，我们可以将想法

to controllable physical machine.

变成可控的物理机器。

We can also do image-to-robot.

我们也可以做到“图像生成机器人”。

This photo can be transformed into a cuddly robotic bunny.

这张照片可以 变成可爱的机器人兔子。

To do so, our algorithm computes a 3D representation of the photo

为此，我们的算法计算出 照片的三维表现形式，

that gets sliced and folded, printed.

然后进行切片、折叠、打印。

Then we fold the printed layers, we string some motors and sensors.

然后我们折叠打印层， 安装一些马达和传感器。

We write some code, and we get the bunny you see in this video.

我们写一些代码，然后我们得到了 你在这段视频中看到的那只兔子。

We can use this approach to make anything almost,

我们可以使用这种方法 制作几乎任何东西，

from an image, from a photo.

从图像、从照片中制作任何东西。

So the ability to transform text into images

因此，将文本转换为图像

and to transform images into robots is important,

以及将图像转换为 机器人的能力非常重要，

because we are drastically reducing the amount of time

因为我们正在大大减少原型设计

and the resources needed to prototype and test new products,

和测试新产品所需的时间和资源，

and this is allowing for a much faster innovation cycle.

这使得创新周期大大提速。

And now we are ready to even make the leap

我们现在甚至准备迈出一大步，

to get these machines to learn.

让这些机器学习。

The third aspect of physical intelligence.

物理智能的第三个方面。

These machines can learn from humans how to do tasks.

这些机器可以 向人类学习如何完成任务。

You can think of it as human-to-robot.

你可以把它看作是“人到机器人”。

In my lab, we created a kitchen environment

在我的实验室里， 我们打造了一个厨房环境，

where we instrument people with sensors,

给人装上传感器，

and we collect a lot of data about how people do kitchen tasks.

收集了大量人们完成厨房任务的数据。

We need physical data

我们需要物理数据，

because videos do not capture the dynamics of the task.

因为视频无法捕捉任务的动态。

So we collect muscle, pose, even gaze information

于是我们收集了人们完成任务时的

about how people do tasks.

肌肉、姿势，甚至凝视信息。

And then we train AI using this data

然后我们使用这些数据训练 AI，

to teach robots how to do the same tasks.

教机器人如何完成同样的任务。

And the end result is machines that move with grace and agility,

最终结果是机器既能优雅灵活地移动，

as well as adapt and learn.

又能适应和学习。

Physical intelligence.

物理智能。

We can use this approach to teach robots

我们可以使用这种方法教机器人

how to do a wide range of tasks:

如何完成各种各样的任务：

food preparation, cleaning and so much more.

准备食物、打扫卫生等等。

The ability to turn images and text into functional machines,

将图像和文本 转换为实用机器的能力，

coupled with using liquid networks

再加上液态网络，

to create powerful brains for these machines

为这些可以向人类学习的机器

that can learn from humans, is incredibly exciting.

创造强大的大脑， 真是激动人心。

Because this means we can make almost anything we imagine.

因为这意味着我们几乎 可以做到任何我们想象的东西。

Today's AI has a ceiling.

如今的 AI 有上限。

It requires server farms.

它需要服务器群。

It's not sustainable.

这是不可持续的。

It makes inexplicable mistakes.

它会犯莫名其妙的错误。

Let's not settle for the current offering.

我们不能安于现状。

When AI moves into the physical world,

当 AI 进入物理世界时，

the opportunities for benefits and for breakthroughs is extraordinary.

受惠和突破的机会是与众不同的。

You can get personal assistants that optimize your routines

你可以获得帮你优化日常工作、

and anticipate your needs,

预测需求的私人助理，

bespoke machines that help you at work

为辅助你工作量身打造的机器，

and robots that delight you in your spare time.

还有在业余时间取悦你的机器人。

The promise of physical intelligence is to transcend our human limitations

物理智能的前景是超越人类的局限，

with capabilities that extend our reach,

这些能力可以扩大我们的触达范围，

amplify our strengths

增强我们的优势，

and refine our precision

提高我们的精度，

and grant us ways to interact with the world

为我们提供与梦寐以求的 世界互动的方式。

we've only dreamed of.

We are the only species so advanced, so aware,

我们是唯一一个 如此先进、如此敏锐、

so capable of building these extraordinary tools.

有能力打造 这些非凡工具的物种。

Yet, developing physical intelligence

然而，发展物理智能

is teaching us that we have so much more to learn

告诉我们，我们还有 很多东西需要学习，

about technology and about ourselves.

学习技术，学习我们自己。

We need human guiding hands over AI sooner rather than later.

我们需要尽快而不是拖延 让人类引导 AI。

After all, we remain responsible for this planet

毕竟，我们还是要对这个星球

and everything living on it.

和身处其中的一切负责。

I remain convinced that we have the power

我仍然坚信，我们有能力

to use physical intelligence to ensure a better future for humanity

使用物理智能来保障人类

和地球更美好的未来。

and for the planet.

And I'd like to invite you to help us in this quest.

我想邀请你帮助我们完成这个任务。

Some of you will help develop physical intelligence.

有些人将协助开发物理智慧。

Some of you will use it.

有些人会使用它。

And some of you will invent the future.

有些人会创造未来。

Thank you.

谢谢。

(Applause)

（掌声）