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

譯者: Armani Chen 審譯者: 麗玲 辛

當我還在學機器人學的時候，

When I was a student studying robotics,

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

我們決定為教授做一份生日禮物，

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

我們想讓我們做的機器人， 為他切一片生日蛋糕。

We pulled an all-nighter writing the software,

並花了一整晚開發軟體，

and the next day, disaster.

但是隔天...是一場災難。

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

我們為這個機器人寫好程式， 要切柔軟的圓形海綿蛋糕

but we didn't coordinate well.

但我們沒有協調好，

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

我們反而得到了一個 方又硬的冰淇淋蛋糕，

The robot flailed wildly and nearly destroyed the cake.

機器人瘋狂地揮動， 幾乎摧毀了蛋糕。

(Laughter)

（笑聲）

Our professor was delighted, anyway.

但總之，我們的教授很開心。

He calmly pushed the stop button

他冷靜地按下停止按鈕，

and declared the erratic behavior of the robot

並表明說機器人不穩定行為

a control singularity.

是它控制的獨特性，

A robotics technical term.

這是機器人技術術語，

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

那讓我很失望， 但我學到非常重要的教訓。

The physical world,

在實體世界，

with its physics laws and imprecisions,

有其物理定律和不精確性， 是一個比數位世界要求更高的空間。

is a far more demanding space than the digital world.

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

現在我帶領著麻省理工學院 計算機科學和 AI 實驗室，

the largest research unit at MIT.

這是麻省理工學院最大的研究單位。

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

這是我們的大樓， 我在此與其他傑出的研究人員工作，

to invent the future of computing and intelligent machines.

一同研發計算機學和智能機器的未來。

Today in computing,

在現在的計算機學裡，

artificial intelligence and robotics are largely separate fields.

人工智能和機器人學是兩個獨立領域，

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

AI 有令人驚艷的決策和學習能力，

but it remains confined inside computers.

但它仍被限制在電腦內。

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

機器人則具有實體， 可以執行預先編程的任務，

but they're not intelligent.

卻不具智能。

Well, this separation is starting to change.

不過這種分水嶺將開始改變。

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

AI 將擺脫只能在 2D 螢幕互動這項限制，

並進入充滿活力的 實體 3D 世界中，

and enter a vibrant, physical 3D world.

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

在我的實驗室， 我們將 AI 的數位智慧，

with the mechanical prowess of robots.

與機器人的機械能力融合。

將 AI 從數位世界 帶到我們的實體世界，

Moving AI from the digital world into the physical world

is making machines intelligent

使機器更加聰明，

and leading to the next great breakthrough,

並將帶來下一個巨大突破，

what I call physical intelligence.

我稱它為物理智能，

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

物理智能是指當 AI 有理解文字、

圖像和其他網路上資訊的能力，

images and other online information

is used to make real-world machines smarter.

被用來讓真實世界的機器更聰明。

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

這代表 AI 可以使用數據的知識， 幫助預先編程好的機器人，

by using knowledge from data.

更好的完成任務，

With physical intelligence,

借助物理智能，

AI doesn't just reside in our computers,

AI 不僅存在於我們的電腦中，

but walks, rolls, flies

它可以走、滾動、飛行，

and interacts with us in surprising ways.

並用其他令人驚艷的方式與我們互動，

Imagine being surrounded by helpful robots at the supermarket.

想像你在超市時， 周圍都是協助你的機器人，

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

左邊的機器人可以幫你提重的箱子，

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

為實現這一點， 我們需要做一些事。

We need to rethink how machines think.

我們需要重新思考機器的思考方式。

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

我們需要重新構思， 它們是如何設計及學習的。

So for physical intelligence,

因此對於物理智能，

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

AI 必須在適合機器人 體型大小的電腦上運行。

For example, our soft robot fish.

例如，我們的機器魚。

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

但現今的 AI 使用 不合適的伺服器農場。

Today's AI also makes mistakes.

現在的 AI 也會犯錯。

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

這台機器汽車上的 AI 系統， 沒辦法檢測到行人。

For physical intelligence,

為了實現物理智能， 我們需要不會犯錯的小型腦袋。

we need small brains that do not make mistakes.

We're tackling these challenges using inspiration

我們尋找靈感來應對這些挑戰，

from a worm called C. elegans

這靈感來自於一種線蟲， 叫秀麗隱桿線蟲 (C. elegans)，

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

與人腦中數十億個神經元 形成鮮明對比，

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

秀麗隱桿線蟲只靠 302 個 神經元，就過著幸福的生活，

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

而生物學家了解 這些神經元運作的數學原理。

So here's the idea.

這就是我們想法。

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

我們能否用這些神經元的 數學原理來構建 AI ？

We have developed, together with my collaborators and students,

我和合作夥伴及學生一起開發了

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

一種 AI 的新方法 我們稱為「液態網絡」。

And liquid networks results in much more compact

而液態網絡相較於傳統 AI ，

有更簡潔且的更容易解釋的解決方案。

and explainable solutions than today's traditional AI solutions.

Let me show you.

讓我展示給你看。

This is our self-driving car.

這是我們的自動駕駛汽車。

It's trained using a traditional AI solution,

它使用傳統的 AI 方案進行訓練，

the kind you find in many applications today.

這種解決方案在許多 應用程式中都可以找到。

This is the dashboard of the car.

這是汽車的儀表板。

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

在右下角，你會看到地圖，

In the upper left corner, the camera input stream.

在左上角則是攝影機收到的訊號。

And the big box in the middle with the blinking lights

而且中間有閃爍燈的大箱子，

is the decision-making engine.

是決策引擎。

It consists of tens of thousands of artificial neurons,

它由數十萬個人工神經元組成，

and it decides how the car should steer.

而且它決定了汽車應該如何駕駛。

It is impossible to correlate the activity of these neurons

很難將這些神經元的活動 與車輛的行為聯繫起來。

with the behavior of the car.

Moreover, if you look at the lower left side,

此外，如果你看左下角，

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

你會看到圖片中的是決策引擎，

to tell the car what to do.

它會跟汽車說該做什麼。

And you see how noisy it is.

你可以看到它有多嘈雜。

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

而這輛車行駛時， 是看著路邊的灌木和樹木。

on the side of the road.

That's not how we drive.

但這不是我們人類開車的方法。

People look at the road.

我們會看著道路。

Now contrast this with our liquid network solution,

現在，我們拿它和 液態網絡的解決方案對比，

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

它僅由 19 個神經元組成， 而不是數十萬個。

And look at its attention map.

看左下的注意力地圖 (attention map)。

It's so clean and focused on the road horizon

它非常乾淨，並專注於道路的 地平線和兩旁。

and the side of the road.

Because these models are so much smaller,

因為它模型小得多，

we actually understand how they make decisions.

我們實際上了解 它們如何做出每個決定。

So how did we get this performance?

那我們是如何得到這個成果的？

Well, in a traditional AI system,

在傳統的 AI 系統中，

the computational neuron is the artificial neuron,

計算神經元是人工神經元，

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

人工神經元本質上 是一個開/關的計算單元。

It takes in some numbers, adds them up,

它會採用一些數字，將它們相加，

applies some basic math

應用一些基本數學，

and passes along the result.

並把結果傳遞下去。

And this is complex

而這是很複雜的，

because it happens across thousands of computational units.

因為它發生在數千個計算單位中。

In liquid networks,

在液態網絡中，

we have fewer neurons,

它們的神經元較少，

but each one does more complex math.

但每個神經元都能進行更複雜的數學。

Here's what happens inside our liquid neuron.

這是在液態神經元內發生的事。

We use differential equations to model the neural computation

我們使用微分方程模擬神經計算，

and the artificial synapse.

以及人工突觸。

And these differential equations

這些微分方程

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

是生物學家對蠕蟲的 神經結構映射出的結果。

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

我們也以不同的方式連結神經元 以增加資訊流量。

Well, these changes yield phenomenal results.

這些變化帶來了驚人的結果。

Traditional AI systems are frozen after training.

傳統的 AI 系統 在訓練後會被凍結。

That means they cannot continue to improve

這意味著在外面物理世界裡， 當我們部署它們後，

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

這些系統沒辦法再繼續進步，

We just wait for the next release.

我們只能等下一個版本，

Because of what's happening inside the liquid neuron,

由於液態神經元內部發生的事情，

liquid networks continue to adapt after training

液態網絡在訓練後，會根據 所看到的輸入內容來適應改動。

based on the inputs that they see.

Let me show you.

我來展示給你看。

We trained traditional AI and liquid networks

我們用這類夏天影片 來訓練傳統 AI 和液態網絡，

using summertime videos like these ones,

and the task was to find things in the woods.

任務是在樹林中找到東西。

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

所有模型在夏天時 都學會了如何完成任務。

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

我們接著嘗試在秋天 在無人機上使用這些模型。

The traditional AI solution gets confused by the background.

傳統的 AI 會因背景產生混亂，

Look at the attention map, cannot do the task.

你看注意力地圖 (attention map) ， 它無法完成任務。

Liquid networks do not get confused by the background

但液態網絡不會因背景感到混亂，

and very successfully execute the task.

並非常成功地執行任務。

So this is it.

所以就是這樣。

這是前進的一步：

This is the step forward:

AI that adapts after training.

訓練後能適應的 AI 。

Liquid networks are important

液態網絡非常重要，

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

因為它們為我們提供了 一種讓機器思考的新方式，

that is rooted into physics models,

讓機器思考植根於實體模型，

a new technology for AI.

這是一種新的 AI 技術。

We can run them on smartphones, on robots,

我們可以在智能手機、機器人、

on enterprise computers,

企業電腦上運行它們，

and even on new types of machines

甚至在我們現在開始想像、 設計的新類型的機器上運行。

that we can now begin to imagine and design.

The second aspect of physical intelligence.

物理智能的第二方面，

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

現在你可能已經使用 文本生成圖像系統，生成了圖像，

We can also do text-to-robot,

我們也可以用文本生成機器人，

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

但不是使用現今的 AI 方案，

因為它們依賴統計學而不理解物理。

and do not understand physics.

In my lab,

在我的實驗室中，

we developed an approach that guides the design process

我們開發了一種方法，

通過檢查和模擬機器的物理限制， 來引導設計過程。

by checking and simulating the physical constraints for the machine.

We start with a language prompt,

我們從語言提示開始，

"Make me a robot that can walk forward,"

「幫我做一個能向前走的機器人。」

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

我們的系統會產生設計， 包括形狀、材料、執行器、

sensors, the program to control it

感測器、控制它的程式， 以及製作的檔案。

and the fabrication files to make it.

And then the designs get refined in simulation

接下來，設計會在模擬中進行精細化，

until they meet the specifications.

直到符合規格為止。

So in a few hours we can go from idea

因此在幾個小時內，

to controllable physical machine.

我們可以從想法轉為 可以控制的物理機器。

We can also do image-to-robot.

我們也可以用圖像創造機器人。

This photo can be transformed into a cuddly robotic bunny.

這張照片可以變成可愛的兔子機器人。

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

為此，我們的演算法計算 裁切、折疊、列印後的照片，

that gets sliced and folded, printed.

並以 3D 表示。

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

我們接著折疊打印的圖層， 串起一些馬達和感測器。

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

我們寫一些程式， 就得到你現在看到的兔子。

We can use this approach to make anything almost,

我們可以用這種方法 幾乎製作任何東西，

from an image, from a photo.

從一張照片，一張圖片。

So the ability to transform text into images

因此，將文字轉換為圖像，

and to transform images into robots is important,

並將圖像轉換為 機器人的能力非常重要，

because we are drastically reducing the amount of time

因為我們正在大幅減少

做原型和測試新產品所需的時間，

and the resources needed to prototype and test new products,

and this is allowing for a much faster innovation cycle.

而這使得創新周期大大加快。

And now we are ready to even make the leap

現在我們已準備好了， 甚至躍到下一步，

to get these machines to learn.

讓這些機器學習。

The third aspect of physical intelligence.

物理智能的第三個方面。

These machines can learn from humans how to do tasks.

這些機器可以向 人類學習如何完成任務。

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

你可以將它視為人對機器人。

In my lab, we created a kitchen environment

在我的實驗室中， 我們建了一個廚房的環境，

where we instrument people with sensors,

我們為人們佩戴感測器，

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

並收集了大量有關人們 如何執行廚房任務的數據。

We need physical data

我們需要實體數據，

because videos do not capture the dynamics of the task.

因為影片無法捕獲任務的動態。

So we collect muscle, pose, even gaze information

因此我們收集肌肉、姿態， 甚至是視線信息，

about how people do tasks.

來了解人們如何執行任務。

And then we train AI using this data

然後我們使用這些數據訓練 AI ，

to teach robots how to do the same tasks.

來教導機器人如何執行相同的任務。

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

最終的結果是 機器能優雅和敏捷地移動，

as well as adapt and learn.

以及適應和學習。

Physical intelligence.

物理智能，

We can use this approach to teach robots

我們可以使用這種方法 來教導機器人

how to do a wide range of tasks:

完成各式各樣的任務:

food preparation, cleaning and so much more.

包括準備食材、清潔等，甚至更多。

The ability to turn images and text into functional machines,

將圖像和文本 轉換為功能機器的能力，

coupled with using liquid networks

加上使用液態網絡，

to create powerful brains for these machines

來創造強大的機器大腦， 這些大腦可以向人類學習，

that can learn from humans, is incredibly exciting.

是令人非常興奮的。

Because this means we can make almost anything we imagine.

因為這意味著我們幾乎可以製作 我們想像的任何東西。

Today's AI has a ceiling.

今天的 AI 有一個上限。

It requires server farms.

它需要伺服器農場。

It's not sustainable.

它不能永續發展。

It makes inexplicable mistakes.

它會犯令人費解的錯誤。

Let's not settle for the current offering.

我們不能滿足於當前它所提供的價值。

When AI moves into the physical world,

當 AI 進入實體世界時，

the opportunities for benefits and for breakthroughs is extraordinary.

帶來好處以及重大突破的 機會是非凡的。

You can get personal assistants that optimize your routines

你可以擁有個人助理，

來優化你的日常工作， 並預測你的需求，

and anticipate your needs,

bespoke machines that help you at work

客製化的機器來協助你的工作，

and robots that delight you in your spare time.

以及空閒時取悅你的機器人，

The promise of physical intelligence is to transcend our human limitations

物理智能帶給我們的承諾是 超越我們人類的限制，

with capabilities that extend our reach,

擴展我們的影響力，加強我們的力量，

amplify our strengths

and refine our precision

以及提高我們的精確度，

and grant us ways to interact with the world

並賦予我們方法 與我們夢想的世界互動，

we've only dreamed of.

We are the only species so advanced, so aware,

我們是唯一先進且有意識的物種，

so capable of building these extraordinary tools.

如此有能力建造 這些非凡的工具。

Yet, developing physical intelligence

然而發展物理智能，

is teaching us that we have so much more to learn

正在教導我們，還有更多 科技和人類的知識要學習。

about technology and about ourselves.

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

我們需要人類儘早指導 AI 。

After all, we remain responsible for this planet

畢竟我們仍然必須對 這個星球以及生活於此的物種負責。

and everything living on it.

I remain convinced that we have the power

我仍然相信我們有能力，

to use physical intelligence to ensure a better future for humanity

使用物理智能可以確保人類和地球 擁有更美好的未來。

and for the planet.

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

我想邀請你幫助我們完成這個任務。

Some of you will help develop physical intelligence.

在座有些人能協助發展物理智能。

Some of you will use it.

有些人會使用它。

And some of you will invent the future.

而有些人將創造未來。

Thank you.

謝謝。

(Applause)

（掌聲）