MyOOPS開放式課程
請加入會員以使用更多個人化功能
來自全球頂尖大學的開放式課程,現在由世界各國的數千名義工志工為您翻譯成中文。請免費享用!
課程來源:TED
     

 

Vijay Kumar 談會飛行及互相合作的機器人

Vijay Kumar: Robots that fly ... and cooperate

 

Photo of three lions hunting on the Serengeti.

講者:Vijay Kumar

2012年2月演講,2012年3月在TED2012上線

 

翻譯:TED

編輯:朱學恆、洪曉慧

簡繁轉換:洪曉慧

後製:洪曉慧

字幕影片後制:謝旻均

 

影片請按此下載

MAC及手持裝置版本請按此下載

閱讀中文字幕純文字版本

 

關於這場演講

Vijay Kumar和他的團隊在賓州大學實驗室裡製造了使用四個螺旋槳飛行、小巧而靈活的機器人,這些機器人可以聚集在一起,感應彼此的存在,並組成特殊的任務團隊-進行建築工程、災區偵查及其它的任務。

 

關於Vijay Kumar

Vijay Kumar在賓夕法尼亞大學研究機器人團隊的控制及協調。

 

為什麼要聽他演講

賓夕法尼亞大學「機器人科學、自動、感測與感知」(GRASP)實驗室研發了可組成特殊隊形、以四個螺旋槳飛行的機器人,它們能自行排列成完美的隊伍,即使其中一個脫隊,其他機器人也能自行填補空缺。你或許看過這部廣為流傳的影片:在掛著網子的GRASP實驗室中以四個螺旋槳轉動的機器人。(它們能表演特技、穿過呼拉圈!)

 

Vijay Kumar於1998至2004年領導這個實驗室;他現在是費城賓夕法尼亞大學工程與應用科學學院副院長,持續進行研發機器人的工作,將計算機科學與機械工程融合,創造出下一代機器人奇蹟。

 

Vijay Kumar的英語網上資料

Lab: GRASP Lab

 

[TED科技‧娛樂‧設計]

已有中譯字幕的TED影片目錄(繁體)(簡體)。請注意繁簡目錄是不一樣的。

 

Vijay Kumar 談會飛行及互相合作的機器人

 

早安,今天我想談談會自動飛行的海灘球。不,是靈巧的飛行機器人,就像這個。我想告訴大家製作這種東西的挑戰性,以及運用這種技術一些很棒的可能性。所以這些機器人算是一種無人飛行器,然而,如你所見,它們的尺寸比較大,有幾千磅重,一點都不靈巧,它們甚至無法自動操作。事實上,這些飛行器大多由飛行小組操作,包括好幾位駕駛員同時操控感應器及任務協調器。

 

我們想開發像這樣的機器人-左邊兩張照片是可以買到的現成機器人;這是一架有四個螺旋槳的直昇機,長寬大約是一公尺,重達好幾磅,於是我們改良它的感應器與處理器,讓這些機器人能不靠GPS在室內飛行。

 

我手中拿的機器人就是這種飛行器,這是由兩位學生-Alex及Daniel製作,重量大約是十分之一磅,消耗的能量大約是15瓦。如你所見,它的直徑大約是8英吋,讓我替大家簡單介紹一下這些機器人的原理。

 

它有四個螺旋槳,當四個螺旋槳速度相同時,機器人會懸浮在空中,如果這些螺旋槳的速度增加,機器人會飛起來,往上加速。當然,如果機器人向水平面傾斜,它會往這個方向前進。所以,如果想讓它傾斜,有兩種方法可以辦到。在這張圖片中,你可以看見4號螺旋槳的轉速變快,2號螺旋槳的轉速變慢,當這種情況發生時,機器人就會翻轉。另一種情況是,當3號螺旋槳的速度上升,1號螺旋槳的速度下降時,機器人就會往前傾斜。

 

最後一種可能是,當相對的一組螺旋槳轉得比另外一組快時,機器人就會往垂直方向偏移。其中有一個內置處理器,監控它該進行什麼動作,並將這些動作組合-以每秒600次的速度,決定該對這些螺旋槳下達什麼指令,這就是它的基本操作概念。

 

這項設計的優點之一是,當你將它的尺寸縮小時,機器人自然會變得很靈巧,所以這裡的R代表機器人的特徵長度,事實上相當於它的半徑。當你將R縮小時,許多物理係數會跟著變動,其中最重要的是慣性-或抵抗運動狀態改變的性質。所以結果是,控制了角運動的慣性大小約是R的5次方,所以當R變小時,慣性會急遽下降。結果是,角加速度-這裡以希臘字母α表示,變成了1 / R,和R成反比;尺寸越小,它轉的越快。

 

這部影片可以清楚說明這一點。在右下角,你可以看見一個機器人,在不到1 / 2秒的時間內進行360度翻轉,多次翻轉只需稍微長一點的時間。所以內置處理器接收加速器及陀螺儀傳回的資訊,然後進行計算,如我之前所說的,以每秒600次的速度發出指令,讓機器人保持平衡。在左下角,你可以看見Daniel正將機器人拋向空中,這顯示了它的操控能力有多強大,無論你怎麼丟,機器人都會恢復平衡,回到他手中。

 

為什麼要製造像這樣的機器人?好,這種機器人可以做很多應用。你可以將它派遣到像這樣的建築物裡擔任先遣部隊,尋找入侵者,或尋找生化物質外洩或瓦斯外洩等;你也可以將它們運用在例如建築上面,這些機器人正運送橫樑、柱子,組合成立方體形狀的建築物,我稍後會再詳細說明。這些機器人可用來運送貨櫃,但這些小機器人的問題在於它們的負重能力有限,所以你會希望多一點機器人來搬運重物。這是我們近期實驗的照片-事實上已經不算近期了-這是地震後的仙台市,這種機器人可被送入傾倒的建築物裡,評估天災造成的損害,或被送入反應爐裡勘查輻射等級。

 

想讓這些機器人自動化必須先解決一個基本問題,就是必須讓它能判斷怎麼從A點到B點。這有一些難度,因為這個機器人的動力學相當複雜,事實上它們活在12維空間裡,所以我們運用了一些技巧,我們將這個彎曲的12維空間轉換成一個平面的四維空間,在這個四維空間中包含了 X, Y, Z 和偏移的角度。

 

所以這個機器人所做的是,找出我們所謂的最小震盪軌跡。複習一下物理參數,我們有位置、衍生速度,然後是加速度,再來是角加速度,然後是震盪,所以機器人將震盪最小化,實際上的結果是,產生平順而優雅的動作,還能避開障礙物。這個平面空間中的最小震盪軌跡,必須被轉換回複雜的12維空間所使用的形式,機器人才能進行控制及執行任務。

 

讓我給大家看一些例子,說明最小震盪軌跡是什麼模樣。在第一段影片中,你可以看見機器人經過中繼點,然後由A點到達B點(最小震盪軌跡),所以機器人確實可執行任何曲線軌跡。這些是環狀軌跡,機器人牽引著大約2 G的重力,頂端有個動態影像攝影機,它會以每秒100次的速度告訴機器人它身處何處,它也會告訴機器人這些障礙物的位置。移動的障礙物也行。你會看見Daniel將這個鐵環丟向空中,機器人會計算鐵環的位置,試著找出穿過鐵環的最佳方式。身為一位學術人員,我們總是被訓練得能克服萬難來籌措研究經費,所以我們製造的機器人也能克服萬難。

 

(掌聲)

 

這個機器人還能做另一件事,就是記住軌跡模式,這是藉由經驗學習或事先輸入。所以你可以看見,機器人會組合一項動作,讓它產生動量,然後改變方向,再回復原狀。它必須這麼做,因為這個窗口只比機器人的寬度稍微大一點,就像是跳水選手站在跳板上,跳起來產生動量,然後快速旋轉,做兩圈半的空翻,最後優雅地恢復原始狀態,基本上這就是機器人所做的事。它懂得如何結合這些零碎的軌跡,來達成這些相當困難的任務。

 

我想換個話題。這些小機器人的缺點之一就是尺寸。如同先前提過的,我們想使用大量機器人來解決尺寸的限制,其中一個困難是如何協調這麼多的機器人?所以我們向自然界借鏡。我想讓大家看一段影片,內容是沙漠盤腹蟻在Stephen Pratt教授實驗室裡搬運東西。這其實是一小塊無花果,事實上你可以將任何東西沾上無花果汁,螞蟻們就會將它搬回巢穴。這些螞蟻並沒有中樞協調者,牠們能感覺到身旁的鄰居,不用進行明確的溝通,但因為牠們能感覺到鄰居、感覺到物體,所以這個團體擁有隱性協調能力。

 

所以這就是我們希望機器人擁有的協調能力。當一個機器人被其它機器人包圍時-看看機器人 I 和機器人 J,我們希望機器人做的是,當它們以特定隊形飛行時,可同時偵測彼此間的距離,你希望能確保這個距離是在可接受的範圍內。同樣地,機器人偵測這個誤差值,然後以每秒100次的速度計算控制指令,接著以每秒600次的速度將其轉換成螺旋槳控制指令,這也必須以沒有中央控制的方式進行。同樣地,如果有許多機器人,中央協調訊息的速度根本不足以使所有機器人完成任務,再加上機器人必須依靠偵測鄰近機器人來獲得訊息,才能進行動作,最後,我們堅持機器人必須無法預知鄰近機器人會是誰,這就是我們所謂的匿名方式。

 

接下來我要給大家看的影片是20個小機器人以特定隊形飛行。它們正在偵測鄰近機器人的位置,它們保持這個隊形,這些隊形可以改變,可以是平面隊形,也可以是三維空間隊形,如你們所見,它們從三維空間隊形轉變成平面隊形,穿越障礙物時,它們可在飛行中調整隊形。同樣地,在這個8字形的飛行隊伍中,這些機器人的距離非常接近,相距只有幾英吋而已。儘管這些螺旋槳葉片之間有空氣動力的交互影響,它們仍然能維持穩定飛行。

 

(掌聲)

 

一旦知道如何以特定隊形飛行,就能準確地合力拿起物體。這是要讓大家知道,將機器人組成小組後,我們可將它們的力量放大兩倍、三倍、四倍。如你們所見,這麼做的其中一個缺點是,當你將規模放大後-如果你讓很多機器人搬運同一個物體,一定會有效地增加慣性,你將因此付出代價-它們會失去靈活性。但相對地,你會獲得負重能力。

 

我想給大家看的另一項應用是-同樣地,這也是我們實驗室所進行的,這是由Quentin Lindsey完成的,他是一位研究生。他的演算法告訴這些機器人如何自動地將建築材料建造成立體建築,他的演算法告訴機器人該拿起哪個部份,以及什麼時候該把它放在哪裡。你可以在這部影片中看到-這是以10倍、14倍速播放-你可以看見這些機器人建造了三種不同建築,同樣地,一切都自動化進行。Quentin只需要給機器人一張建築物的設計藍圖。

 

你們所看見的這些實驗、這些展示,都使用了動作擷取系統。所以,如果離開實驗室,走進真實世界,會變成怎樣?如果沒有GPS會怎樣?這個機器人裝置了一具攝影機、一具雷射測距儀,雷射掃描器,它使用這些感應器來製作一張環境配置圖,這張地圖包含一些環境特徵-例如門、窗戶、人、家具-它能辨識出本身相對於環境特徵的位置。所以這裡並沒有整體座標系統,座標系統由機器人本身所定義,藉由它所在位置及它所看到的東西,它會偵察這些環境特徵。

 

我想給大家看一段影片,關於Frank Shen及Nathan Michael教授所開發的演算法。這個機器人第一次進入一棟建築,然後在飛行中製作這張地圖,於是機器人偵察出其中的環境特徵、製成地圖,它知道自己相對於環境特徵的位置,然後以每秒100次的速度估算出自己的位置,讓我們可以利用剛剛說過的控制演算法。事實上,這個機器人正由Frank進行遠端遙控,但這個機器人也能自行判斷它該往哪裡走。假設我把它送進一棟建築物,我完全不知道建築物內部情形,我可以命令機器人進入,製作一張地圖,然後回來告訴我建築物的情形。所以機器人並不只是解決如何從地圖上的A點到B點這個問題,它也知道每一次的最佳B點是哪個位置,所以它知道該往哪裡尋找資訊缺乏之處,這就是它製作完整地圖的方法。

 

最後,我想再給大家看一項應用。這個技術有許多應用方式,我是個教授,我們對教育充滿熱情,這種機器人可以改變我們實施12年國教的方式。我們在南加州,很靠近洛杉磯,所以我想用與娛樂有關的例子來作最後結尾,我想用一段音樂影片來做結尾。我要為大家介紹影片製作者-Alex和Daniel。

 

(掌聲)

 

在我播放影片之前,我想告訴大家,他們接到Chris的電話,三天後就將這段影片製作完成,影片中演奏的機器人是完全自動化的,你會看見9個機器人演奏6種不同的樂器,當然,這是為了TED 2012特別製作的,我們一起欣賞吧!

 

(音樂聲)

 

(掌聲)

 

以下為系統擷取之英文原文

About this Talk

In his lab at Penn, Vijay Kumar and his team build flying quadrotors, small, agile robots that swarm, sense each other, and form ad hoc teams -- for construction, surveying disasters and far more.

About the Speaker

At the University of Pennsylvania, Vijay Kumar studies the control and coordination of multi-robot formations.Full bio »

Transcript

Good morning. I'm here today to talk about autonomous, flying beach balls. No, agile aerial robots like this one. I'd like to tell you a little bit about the challenges in building these and some of the terrific opportunities for applying this technology. So these robotsare related to unmanned aerial vehicles. However, the vehicles you see here are big. They weigh thousands of pounds, are not by any means agile. They're not even autonomous. In fact, many of these vehicles are operated by flight crews that can include multiple pilots,operators of sensors and mission coordinators.

What we're interested in is developing robots like this -- and here are two other pictures --of robots that you can buy off the shelf. So these are helicopters with four rotors and they're roughly a meter or so in scale and weigh several pounds. And so we retrofit these with sensors and processors, and these robots can fly indoors without GPS.

The robot I'm holding in my hand is this one, and it's been created by two students, Alex and Daniel. So this weighs a little more than a tenth of a pound. It consumes about 15 watts of power. And as you can see, it's about eight inches in diameter. So let me give you just a very quick tutorial on how these robots work.

So it has four rotors. If you spin these rotors at the same speed, the robot hovers. If you increase the speed of each of these rotors, then the robot flies up, it accelerates up. Of course, if the robot were tilted, inclined to the horizontal, then it would accelerate in this direction. So to get it to tilt, there's one of two ways of doing it. So in this picture you see that rotor four is spinning faster and rotor two is spinning slower. And when that happensthere's moment that causes this robot to roll. And the other way around, if you increase the speed of rotor three and decrease the speed of rotor one, then the robot pitches forward.

And then finally, if you spin opposite pairs of rotors faster than the other pair, then the robot yaws about the vertical axis. So an on-board processor essentially looks at what motions need to be executed and combines these motions and figures out what commands to send to the motors 600 times a second. That's basically how this thing operates.

So one of the advantages of this design is, when you scale things down, the robot naturally becomes agile. So here R is the characteristic length of the robot. It's actually half the diameter. And there are lots of physical parameters that change as you reduce R.The one that's the most important is the inertia or the resistance to motion. So it turns out,the inertia, which governs angular motion, scales as a fifth power of R. So the smaller you make R, the more dramatically the inertia reduces. So as a result, the angular acceleration, denoted by Greek letter alpha here, goes as one over R. It's inversely proportional to R. The smaller you make it the more quickly you can turn.

So this should be clear in these videos. At the bottom right you see a robot performing a 360 degree flip in less than half a second. Multiple flips, a little more time. So here the processes on board are getting feedback from accelerometers and gyros on board and calculating, like I said before, commands at 600 times a second to stabilize this robot. So on the left, you see Daniel throwing this robot up into the air. And it shows you how robust the control is. No matter how you throw it, the robot recovers and comes back to him.

So why build robots like this? Well robots like this have many applications. You can send them inside buildings like this as first responders to look for intruders, maybe look for biochemical leaks, gaseous leaks. You can also use them for applications like construction. So here are robots carrying beams, columns and assembling cube-like structures. I'll tell you a little bit more about this. The robots can be used for transporting cargo. So one of the problems with these small robots is their payload carrying capacity.So you might want to have multiple robots carry payloads. This is a picture of a recent experiment we did -- actually not so recent anymore -- in Sendai shortly after the earthquake. So robots like this could be sent into collapsed buildings to assess the damage after natural disasters, or sent into reactor buildings to map radiation levels.

So one fundamental problem that the robots have to solve if they're to be autonomous is essentially figuring out how to get from point A to point B. So this gets a little challengingbecause the dynamics of this robot are quite complicated. In fact, they live in a 12-dimensional space. So we use a little trick. We take this curved 12-dimensional spaceand transform it into a flat four-dimensional space. And that four-dimensional spaceconsists of X, Y, Z and then the yaw angle.

And so what the robot does is it plans what we call a minimum snap trajectory. So to remind you of physics, you have position, derivative, velocity, then acceleration, and then comes jerk and then comes snap. So this robot minimizes snap. So what that effectively does is produces a smooth and graceful motion. And it does that avoiding obstacles. So these minimum snap trajectories in this flat space are then transformed back into this complicated 12-dimensional space, which the robot must do for control and then execution.

So let me show you some examples of what these minimum snap trajectories look like.And in the first video, you'll see the robot going from point A to point B through an intermediate point. So the robot is obviously capable of executing any curve trajectory. So these are circular trajectories where the robot pulls about two G's. Here you have overhead motion capture cameras on the top that tell the robot where it is 100 times a second. It also tells the robot where these obstacles are. And the obstacles can be moving. And here you'll see Daniel throw this hoop into the air, while the robot is calculating the position of the hoop and trying to figure out how to best go through the hoop. So as an academic, we're always trained to be able to jump through hoops to raise funding for our labs, and we get our robots to do that.

(Applause)

So another thing the robot can do is it remembers pieces of trajectory that it learns or is pre-programmed. So here you see the robot combining a motion that builds up momentum and then changes its orientation and then recovers. So it has to do this because this gap in the window is only slightly larger than the width of the robot. So just like a diver stands on a springboard and then jumps off it to gain momentum, and then does this pirouette, this two and a half somersault through and then gracefully recovers,this robot is basically doing that. So it knows how to combine little bits and pieces of trajectories to do these fairly difficult tasks.

So I want change gears. So one of the disadvantages of these small robots is its size.And I told you earlier that we may want to employ lots and lots of robots to overcome the limitations of size. So one difficulty is how do you coordinate lots of these robots? And so here we looked to nature. So I want to show you a clip of Aphaenogaster desert ants in Professor Stephen Pratt's lab carrying an object. So this is actually a piece of fig. Actually you take any object coated with fig juice and the ants will carry them back to the nest. So these ants don't have any central coordinator. They sense their neighbors. There's no explicit communication. But because they sense the neighbors and because they sense the object, they have implicit coordination across the group.

So this is the kind of coordination we want our robots to have. So when we have a robotwhich is surrounded by neighbors -- and let's look at robot I and robot J -- what we want the robots to do is to monitor the separation between them as they fly in formation. And then you want to make sure that this separation is within acceptable levels. So again the robots monitor this error and calculate the control commands 100 times a second, which then translates to the motor commands 600 times a second. So this also has to be donein a decentralized way. Again, if you have lots and lots of robots, it's impossible to coordinate all this information centrally fast enough in order for the robots to accomplish the task. Plus the robots have to base their actions only on local information, what they sense from their neighbors. And then finally, we insist that the robots be agnostic to who their neighbors are. So this is what we call anonymity.

So what I want to show you next is a video of 20 of these little robots flying in formation.They're monitoring their neighbors' position. They're maintaining formation. The formations can change. They can be planar formations, they can be three-dimensional formations. As you can see here, they collapse from a three-dimensional formation into planar formation. And to fly through obstacles they can adapt the formations on the fly. So again, these robots come really close together. As you can see in this figure-eight flight,they come within inches of each other. And despite the aerodynamic interactions of these propeller blades, they're able to maintain stable flight.

(Applause)

So once you know how to fly in formation, you can actually pick up objects cooperatively.So this just shows that we can double, triple, quadruple the robot strength by just getting them to team with neighbors, as you can see here. One of the disadvantages of doing thatis, as you scale things up -- so if you have lots of robots carrying the same thing, you're essentially effectively increasing the inertia, and therefore you pay a price; they're not as agile. But you do gain in terms of payload carrying capacity.

Another application I want to show you -- again, this is in our lab. This is work done by Quentin Lindsey who's a graduate student. So his algorithm essentially tells these robotshow to autonomously build cubic structures from truss-like elements. So his algorithm tells the robot what part to pick up, when and where to place it. So in this video you see --and it's sped up 10, 14 times -- you see three different structures being built by these robots. And again, everything is autonomous, and all Quentin has to do is to get them a blueprint of the design that he wants to build.

So all these experiments you've seen thus far, all these demonstrations, have been done with the help of motion capture systems. So what happens when you leave your lab and you go outside into the real world? And what if there's no GPS? So this robot is actually equipped with a camera and a laser rangefinder, laser scanner. And it uses these sensors to build a map of the environment. What that map consists of are features -- like doorways, windows, people, furniture -- and it then figures out where its position is with respect to the features. So there is no global coordinate system. The coordinate system is defined based on the robot, where it is and what it's looking at. And it navigates with respect to those features.

So I want to show you a clip of algorithms developed by Frank Shen and Professor Nathan Michael that shows this robot entering a building for the very first time and creating this map on the fly. So the robot then figures out what the features are. It builds the map. It figures out where it is with respect to the features and then estimates its position 100 times a second allowing us to use the control algorithms that I described to you earlier.So this robot is actually being commanded remotely by Frank. But the robot can also figure out where to go on its own. So suppose I were to send this into a building and I had no idea what this building looked like, I can ask this robot to go in, create a map and then come back and tell me what the building looks like. So here, the robot is not only solving the problem, how to go from point A to point B in this map, but it's figuring out what the best point B is at every time. So essentially it knows where to go to look for places that have the least information. And that's how it populates this map.

So I want to leave you with one last application. And there are many applications of this technology. I'm a professor, and we're passionate about education. Robots like this can really change the way we do K through 12 education. But we're in Southern California,close to Los Angeles, so I have to conclude with something focused on entertainment. I want to conclude with a music video. I want to introduce the creators, Alex and Daniel, who created this video.

(Applause)

So before I play this video, I want to tell you that they created it in the last three days after getting a call from Chris. And the robots that play the video are completely autonomous.You will see nine robots play six different instruments. And of course, it's made exclusively for TED 2012. Let's watch.

(Music)


留下您對本課程的評論
標題:
您目前為非會員,留言名稱將顯示「匿名非會員」
只能進行20字留言

留言內容:

驗證碼請輸入6 + 4 =

標籤

現有標籤:1
新增標籤:


有關本課程的討論

目前暫無評論,快來留言吧!

Creative Commons授權條款 本站一切著作係採用 Creative Commons 授權條款授權。
協助推廣單位: