IKEA Furniture and the Limits of AI1家具組裝顯人工智能局限

楊雄飛

Humans have had a good run2. But with the most recent breakthrough in robotics， it is clear that their time as masters of planet Earth has come to an end. 人類雖已拔得頭籌，占得先機(jī)，但近年來機(jī)器人領(lǐng)域科學(xué)技術(shù)的重大突破，清楚地表明人類作為地球統(tǒng)治者的時代已宣告結(jié)束。

Computers have already proved better than people at playing chess and diagnosing diseases. But now a group of artificial-intelligence researchers in Singapore have managed to teach industrial robots to assemble an IKEA chair—for the first time uniting the worlds of Allen keys and Alan Turing3. Now that machines have mastered one of the most baffling4 ways of spending a Saturday afternoon， can it be long before AIs rise up and enslave5 human beings in the silicon mines6？

The research also holds a serious message7. It highlights a deep truth about the limitations of automation. Machines excel at the sorts of abstract， cognitive tasks that， to people， signify intelligence—complex board games8， say， or differential calculus9. But they struggle with physical jobs， such as navigating10 a cluttered11 room， which are so simple that they hardly seem to count as12 intelligence at all. The IKEA-bots are a case in point13. It took a pair of them， pre-programmed by humans， more than 20 minutes to assemble a chair that a person could knock together14 in a fraction of15 the time.

AI researchers call that observation Moravec’s paradox16， and have known about it for decades. It does not seem to be the sort of problem that could be cured with a bit more research. Instead， it seems to be a fundamental truth： physical dexterity17 is computationally harder than playing Go18. That humans do not grasp this is a side-effect19 of evolution. Natural selection has had billions of years to attack the problem of manipulating the physical world， to the point where it feels effortless. Chess， by contrast， is less than 2，000 years old. People find it hard because their brains are not wired for20 it.

That is something to bear in mind when thinking about the much-hyped21 effects of AI and automation， especially as AI moves out of the abstract world of data and information and into the real world of things you can drop on your foot22. Machines may soon be able to drive delivery vans. But， at least for now， they could well fail to carry a parcel to a flat at the top of a flight of slippery stairs， especially if the garden was patrolled by a dangerous dog.

Today’s AI systems are limited in other ways， too. They are pattern-recognition engines23， trained on thousands of examples in the hope that the rules they infer24 will continue to apply in the wider world. But they apply those rules blindly， without a human-like understanding of what they are doing or an ability to improvise25 a solution on the spot26. Makers of self-driving cars， for instance， worry constantly about how their machines will perform in “edge cases27”—complicated and unusual situations that cannot be foreseen during training.

計(jì)算機(jī)已證明自己在棋類游戲和疾病診斷方面要優(yōu)于人類。但一個人工智能研究團(tuán)隊(duì)在新加坡進(jìn)行了一場實(shí)驗(yàn)，讓工業(yè)機(jī)器人組裝宜家的椅子，這是第一次將現(xiàn)實(shí)世界與人工智能連接起來。既然機(jī)器已掌握一種最令人困惑的方式來消磨一個周六下午的時間，那么是不是在不久的將來，人工智能便可在硅谷奴役人類？

這項(xiàng)研究還釋放出另一個重要的信號。它揭示了有關(guān)自動化局限性的一個深刻真相。機(jī)器擅長處理一些抽象、認(rèn)知類任務(wù)，例如復(fù)雜的棋類游戲或微積分，對人類來說完成這類任務(wù)意味著智商夠高。但是機(jī)器處理現(xiàn)實(shí)世界的問題就很難，比如穿過一個雜亂房間這種根本算不上智能的簡單任務(wù)。宜家家具組裝機(jī)器人就是一個很好的例子。預(yù)設(shè)定程序的兩個機(jī)器人組裝一把椅子需要20多分鐘，而一般人很短時間就能裝好。

人工智能研究者把這一現(xiàn)象稱為莫拉維克悖論，該悖論已提出了幾十年。這類問題看起來并不能通過更進(jìn)一步的研究去解決。相反，這似乎更像是一個基本法則，物理上的靈巧性對計(jì)算機(jī)而言遠(yuǎn)遠(yuǎn)比玩圍棋難得多。人類沒能掌握這一技能是進(jìn)化的意外結(jié)果。人類歷經(jīng)數(shù)十億年物競天擇的演化，不斷應(yīng)付“如何擺平現(xiàn)實(shí)世界”這個問題，操控現(xiàn)實(shí)世界已易如反掌。相反，國際象棋的歷史不到2000年。人類之所以覺得國際象棋難，是因?yàn)槿四X還沒有將其內(nèi)化為本能。

在思考被大肆炒作的人工智能和自動化的影響時，尤其是人工智能從抽象的數(shù)據(jù)和信息世界進(jìn)入真實(shí)的世界時，應(yīng)將上述結(jié)論銘記于心。機(jī)器不久便能駕駛送快遞的小貨車了。但是，至少現(xiàn)在看來，機(jī)器恐怕沒法順著一段很滑的樓梯爬到公寓頂樓送包裹，尤其是這棟公寓的花園里還有惡狗看門的話。

現(xiàn)今的人工智能系統(tǒng)在其他方面也有局限性。它們是會識別模式的機(jī)器，經(jīng)過成千上萬案例的訓(xùn)練推論出各種規(guī)則，期待這些規(guī)則能繼續(xù)運(yùn)用到更廣闊的領(lǐng)域里。但是機(jī)器對這些規(guī)則的運(yùn)用有盲目性，它們不像人類明白自己在做什么，也缺乏現(xiàn)場解決問題的能力。例如，無人駕駛汽車的研制者就擔(dān)心機(jī)器如何處理一些“極端狀況”，即一些不能在平時訓(xùn)練中預(yù)見的復(fù)雜的非常規(guī)情況。