Annals of Technology

Are Humanoid Robots Ready to Be Deployed?

人形机器人准备好投入应用了吗?
Neo and a dozen other robots with human forms are scheduled to hit the market. Experts are nervous.
Neo 和其他十几款人形机器人即将上市。专家们感到不安。
Robot holding a coffee cup upside down.
Illustration by La Boca

On a recent sunny day in Silicon Valley, I visited the industrial headquarters of 1X Technologies. Security was tight, so I had to put a sticker over my cellphone’s camera and talk my way out of signing an N.D.A. before I was brought into an enormous space to meet Neo, the company’s home robot.

前不久,在硅谷一个晴朗的日子,我造访了 1X Technologies 的工业总部。安保很严,我不得不用贴纸遮住手机摄像头,费了一番口舌才免签保密协议,随后被带进一个巨大的空间,见到了这家公司的家用机器人 Neo。

Neo stands five feet six and has no facial features except for two black cameras in place of eyes. The robot is a humanoid—its design is inspired by the human form—and its proportions are a blend of those of the median American male and those of the median American female. But Neo has no skin. Instead, it wears a beige nylon turtleneck bodysuit, gloves, and padded shoes over a see-through carapace. Under that is a skeleton made up of more than a hundred whizzing motors and cordlike artificial tendons that control Neo’s limbs. Neo’s cozy, minimalist aesthetic allows it to blend into the background. If it served me an espresso at a café, I’m not certain I would look up from my phone.

Neo 身高五英尺六英寸,脸上没有任何五官,只有两只黑色摄像头充当眼睛。这是一款人形机器人——其设计灵感来自人体——身体比例融合了美国男性和女性的中位数体型特征。但 Neo 没有皮肤,取而代之的是一件米色尼龙高领连体衣、手套和带衬垫的鞋子,套在半透明的外壳上。外壳之下,是由一百多个嗡嗡作响的电机和绳索状人造肌腱组成的骨架,控制着 Neo 的四肢。Neo 这种温馨、极简的美学风格,让它能融入背景之中。如果它在咖啡馆给我端上一杯浓缩咖啡,我不确定自己会不会从手机上抬起头来。

The robot weighs just sixty-six pounds, and I was able to pick it up in a bridal carry. It communicates through a speaker in its chest, using several different voices; the default one is in a calm but authoritative masculine register, an A.I.-modulated mixture of several voice actors. Neo can talk, listen, and respond to commands. In its skull is a squarish motherboard, roughly the size of a slice of bread and studded with white capacitors that look like teeth. This is Neo’s brain. Using A.I., it tries to translate your words into physical actions. If you ask Neo to pick up a cup from a table, Neo will attempt to pick up the cup. If you ask Neo to put the milk in the refrigerator, Neo will attempt to put the milk in the refrigerator. But Neo doesn’t always succeed at these tasks, and no one really knows what it can or cannot do.

这台机器人仅重六十六磅,我能用新娘抱的姿势把它抱起来。它通过胸部的扬声器与人交流,使用几种不同的声音;默认声音是平静而带权威感的男性音色,由人工智能混合多位配音演员的声音调制而成。Neo 能听、能说,也能响应指令。它的头骨里有一块方形主板,大约一片面包大小,上面布满白色电容,看起来像牙齿。这就是 Neo 的大脑。借助人工智能,它试图将你的话语转化为身体动作。如果你让 Neo 从桌上拿起一个杯子,Neo 会尝试去拿。如果你让 Neo 把牛奶放进冰箱,Neo 会尝试去做。但 Neo 并非总能成功完成这些任务,而且没人真正知道它到底能做什么、不能做什么。

The robot is the creation of Bernt Børnich, the C.E.O. of 1X Technologies. Børnich, a forty-two-year-old Norwegian, has been obsessed with robots since he was a child. His firm used to be called Halodi, and it sold wheeled security robots from offices south of Oslo. But in 2022 Børnich relocated to Silicon Valley, changed the company’s name, grew out his hair, and started wearing designer streetwear. Neo was announced shortly afterward.

这台机器人是 1X Technologies 首席执行官伯恩特·伯尼奇(Bernt Børnich)的作品。四十二岁的挪威人伯尼奇从小就痴迷于机器人。他的公司前身叫 Halodi,在奥斯陆以南的办公室里销售轮式安保机器人。但 2022 年,伯尼奇迁至硅谷,改了公司名字,留起长发,开始穿设计师品牌的街头服饰。Neo 随后不久便对外公布。

Børnich believes in the supremacy of the human form. “We are incredibly well designed,” he has said. “I would argue this is kind of, like, why we won evolution.” The environment in which humanoids will operate is already built with able-bodied humans in mind. To open a door handle, you need a hand; to climb a flight of stairs, you need feet. Many roboticists are now working on humanoids, but Børnich takes biomimicry further than his peers: Neo uses tendons where other humanoids use motors, and Neo’s motherboard is in its head, while most humanoids have it stored in the chest. Børnich says humanoids have the potential “to do almost anything.”

伯尼奇信奉人体形态的至高无上。“我们的设计精妙绝伦,”他曾说,“我甚至认为,这差不多就是我们赢得进化竞赛的原因。”人形机器人将要运行的环境,本就是为体格健全的人类而建造的。要转动门把手,你需要一只手;要爬一段楼梯,你需要双脚。许多机器人专家现在都在研发人形机器人,但伯尼奇在仿生方面比同行走得更远:其他人形机器人使用电机的地方,Neo 用的是肌腱;Neo 的主板在头部,而大多数人形机器人将其置于胸腔。伯尼奇说,人形机器人有潜力“做几乎任何事情”。

Mention Neo—or Børnich—to other roboticists and you will be met with a slight pause. Børnich’s technical foundation is solid, but many people feel that he’s getting ahead of himself. Compared with the A.I. that generates language and images, the “physical” A.I. that runs a robot remains underdeveloped. “The world still doesn’t have a ChatGPT equivalent for a robot,” Deepu Talla, who oversees robotics at the microchip designer Nvidia, told me.

向其他机器人专家提起 Neo——或伯尼奇——你会遇到片刻的停顿。伯尼奇的技术基础很扎实,但许多人觉得他有些操之过急。与生成语言和图像的人工智能相比,运行机器人的“物理”人工智能仍不成熟。“世界上还没有一个机器人版的 ChatGPT,”芯片设计公司英伟达(Nvidia)负责机器人业务的迪普·塔拉(Deepu Talla)告诉我。

“I think home comes later,” Carolina Parada, the head of robotics at Google DeepMind, said. “I don’t think it’s a lack of interest, or even capability.” Rather, she explained, it was a question of safety. Jeff Cardenas, the C.E.O. of the industrial-humanoid manufacturer Apptronik, told me, “Around small pets, around small children, there’s still work to be done.”

“我认为进入家庭是以后的事,”谷歌 DeepMind 机器人业务负责人卡罗琳娜·帕拉达(Carolina Parada)说。“我不认为这是缺乏兴趣,甚至不是能力问题。”她解释说,这更是一个安全问题。工业人形机器人制造商 Apptronik 的首席执行官杰夫·卡德纳斯(Jeff Cardenas)告诉我:“在小型宠物和小孩周围,还有很多工作要做。”

Børnich agrees that Neo should not be used around small children. Still, he is keen to test consumers’ tolerance for what he calls “robotics slop”: tasks that are clumsily executed. For example, in a demonstration for the Wall Street Journal in October, Neo took longer than a minute to retrieve a bottle of water. Although Neo has a list price of twenty thousand dollars, more than ten thousand customers have already placed a deposit for one.

伯尼奇同意 Neo 不应在小孩周围使用。尽管如此,他仍急于测试消费者对他所称的“机器人瑕疵”的容忍度:即那些执行得笨手笨脚的任务。例如,去年十月在《华尔街日报》的一次演示中,Neo 花了一分多钟才取回一瓶水。尽管 Neo 标价两万美元,但已有一万多名客户支付了定金。

The customers are who you’d expect—wealthy early adopters from the Bay Area, Los Angeles, and New York City. Will DePue, a former OpenAI employee, lives in San Francisco with tech-enthusiast roommates. “I think our house has three orders on humanoid robots,” he told me. “This is kind of like the new iPhone.” Sometime later this year, workers at 1X’s factory, near Oakland, will pack a Neo into a container that looks like an oversized AirPods case, load it into the back of a truck, and deliver it to DePue’s residence. “Neo will probably not be a perfectly polished product, but at least it’s, like, an early version of what’s to come,” DePue said.

这些客户不出所料——来自湾区、洛杉矶和纽约市的富裕早期尝鲜者。前 OpenAI 员工威尔·德普(Will DePue)与几位科技爱好者室友住在旧金山。“我觉得我们这栋房子订了三台人形机器人,”他告诉我。“这有点像新款 iPhone。”今年晚些时候的某一天,1X 位于奥克兰附近工厂的工人会把一台 Neo 装进一个看起来像超大号 AirPods 充电盒的容器里,装上卡车后厢,送到德普的住处。“Neo 大概不会是一款完美无瑕的产品,但至少是未来产品的一个早期版本,”德普说。

At 1X’s headquarters, my guide was Dar Sleeper, the head of product and design and Børnich’s informal second-in-command. Sleeper, who is twenty-seven, has already had a memorable career. After college, he joined the fashion brand Yeezy, then moved on to industrial design at Tesla. In other words, his first boss was Kanye West, and his second was Elon Musk. Børnich reminded him of Musk, Sleeper told me, if you “graft in some human empathy.”

在 1X 总部,我的向导是产品与设计负责人、伯尼奇非正式的二把手达尔·斯利珀(Dar Sleeper)。二十七岁的斯利珀已经拥有了一段令人难忘的职业生涯。大学毕业后,他加入了时尚品牌 Yeezy,随后转到特斯拉从事工业设计。换句话说,他的第一任老板是坎耶·韦斯特(Kanye West),第二任是埃隆·马斯克(Elon Musk)。斯利珀告诉我,伯尼奇让他想起马斯克,如果你“嫁接进一些人类的同理心”的话。

In 2024, Børnich asked Sleeper to design a robot that didn’t scare children. Sleeper’s first attempt failed. “Basically, anyone under the age of a hundred was scared of it,” he said. (It had an eerily smooth black face.) His next iteration was better, but it still frightened children under the age of twelve. He added and removed facial features, experimented with skin textures, and went through hundreds of configurations of eye sockets. He settled, finally, on stripping as many features from Neo as he could, then covering it with soft textiles. “Even children under five seem to like it,” Sleeper said. “Babies are still scared, though.”

2024 年,伯尼奇请斯利珀设计一款不会吓到孩子的机器人。斯利珀的第一次尝试失败了。“基本上,一百岁以下的人都被它吓到了,”他说。(它有一张诡异光滑的黑色面孔。)他的下一个版本好了一些,但仍会吓到十二岁以下的孩子。他添加又移除面部特征,试验皮肤纹理,尝试了数百种眼窝配置。最终,他决定尽可能从 Neo 身上剥离特征,然后用柔软的纺织品覆盖它。“现在连五岁以下的孩子似乎都喜欢它,”斯利珀说。“不过婴儿还是会害怕。”

Sleeper led me through a labyrinthine industrial space at the headquarters. About eight hundred people work there, in roles involving assembly, testing, and research and development. We passed lathes, metal mills, and racks of 3-D printers, all in continuous operation. We passed a safety-testing zone, where technicians in goggles instructed robotic hands to karate-chop ripe cantaloupes. We passed Neo’s staging environments, which included a bedroom, a bathroom, a kitchen, and a dojo with tatami mats on the floor. We stopped at a motion-capture studio, where, off to one side, two dozen decommissioned models knelt together, as if in prayer.

斯利珀带我穿过总部迷宫般的工业空间。大约有八百人在那里工作,从事装配、测试和研发等岗位。我们经过车床、金属铣床和一排排持续运转的 3D 打印机。我们经过一个安全测试区,戴着护目镜的技术人员正指令机械手对手刀劈开成熟的哈密瓜。我们经过 Neo 的模拟环境,包括一间卧室、一间浴室、一间厨房和一间铺着榻榻米的道场。我们在一间动作捕捉工作室停下,一旁有二十几个退役的模型跪在一起,仿佛在祈祷。
Batman talks on phone and looks at BatSignal out window while in bed with computer and chips.
“Yeah, I see it. I’m on my way as we speak.”
“看到了,我这就出发,正说着呢。”
Cartoon by Ellis Rosen
Copy link to cartoon
Shop

This was where Sleeper had taught Neo how to walk. He told me that he had used his own body as the reference for Neo’s gait. (Sleeper stands well over six feet tall, with Vitruvian proportions.) Referring to his engineering co-workers, Sleeper said, “A lot of these guys walk like this,” then pantomimed a hunched scurry, his hands clutching imaginary backpack straps. “But I was an athlete in college,” he continued. “I actually had an anatomically perfect run.”

这里就是斯利珀教 Neo 走路的地方。他告诉我,他用自己的身体作为 Neo 步态的参考。(斯利珀身高远超六英尺,拥有维特鲁威式的比例。)谈到他的工程师同事,斯利珀说:“这帮家伙很多人走路是这个样子,”然后模仿起一种弓着背、紧赶慢赶的步态,双手像抓着想象中的背包带。“但我大学时是运动员,”他继续说,“我实际上拥有解剖学意义上完美的跑步姿态。”

Sleeper puts the “bro” in tech bro. He played lacrosse at the University of Michigan, and, despite the controversies surrounding Kanye West, describes him as “one of the best creative geniuses of all time.” At one point, we were discussing a portable battery pack that Sleeper is designing for Neo. “It’s actually a nontrivial problem,” he told me. I asked him if he had ever used the word “nontrivial” before moving to Silicon Valley. “Probably not,” he said, and laughed. “I didn’t say a lot of shit before I showed up here.”

斯利珀把“兄弟”放进了科技兄弟。他在密歇根大学打过长曲棍球,尽管坎耶·韦斯特争议缠身,他仍将其描述为“有史以来最杰出的创意天才之一”。有一次,我们讨论斯利珀正在为 Neo 设计的一款便携式电池组。“这其实是个非平凡问题,”他告诉我。我问他,在搬到硅谷之前是否用过“非平凡”这个词。“大概没有,”他说着笑了。“我来这儿之前,好多屁话都不说的。”

Sleeper has built a world around Neo. He oversees an in-house woodshop, which he uses to design sets, props, and furniture, all to mimic the environments that Neo will have to navigate. His plans for the 1X campus involve both a television studio and a forest. On the day of my tour, Sleeper wore a striking beige sweater, which he’d designed himself, in the style of Neo’s bodysuit.

斯利珀围绕 Neo 构建了一个世界。他管理着一个内部木工坊,用来设计布景、道具和家具,所有这些都是为了模拟 Neo 将要应对的环境。他对 1X 园区的规划包括一个电视演播室和一片森林。在我参观那天,斯利珀穿着一件引人注目的米色毛衣,是他自己设计的,风格与 Neo 的连体衣一致。

I have been reporting on Silicon Valley for some while, which is another way of saying that I have been served a lifetime supply of bullshit. Before my visit to 1X, I had been ready to call bullshit on Neo, too. The company had posted some splashy videos on social media a few months earlier, then gone dark, and I suspected that Neo was just more overpromoted vaporware.

我报道硅谷已有一段时间,这等于说,我这辈子听过的空话大话已经够多了。在造访 1X 之前,我也准备好对 Neo 嗤之以鼻。这家公司几个月前在社交媒体上发布了一些炫目的视频,随后便沉寂下来,我怀疑 Neo 不过是又一款过度炒作的雾件。

In fact, I was ready to call bullshit on the whole robotics industry. At the Las Vegas Consumer Electronics Show this January, there were numerous robots on display, from many manufacturers. They all seemed clunky and loud. They clanked when they walked. Sure, one robot landed a backflip—but then a piece of it broke off and flew toward the crowd. Nothing there seemed ready for consumer use.

事实上,我准备好对整个机器人行业嗤之以鼻。在今年一月拉斯维加斯消费电子展上,众多制造商展出了大量机器人。它们看起来都笨重而嘈杂,走路时哐当作响。没错,有一台机器人完成了一个后空翻——但随后它的一块零件断裂,飞向人群。那里似乎没有任何东西准备好供消费者使用。

But Neo was different. Neo was elegant. Neo was soft. Above all, Neo was quiet. With its padded feet and innovative tendons, it hummed along at twenty-two decibels, about the same volume as leaves rustling in a breeze. I felt a deep and horrible emotion triggered inside me, one familiar to any American—consumer lust. I wanted Neo in my house. I wanted it instantly, badly, immediately.

但 Neo 不一样。Neo 很优雅。Neo 很柔软。最重要的是,Neo 很安静。凭借带衬垫的双脚和创新的肌腱,它以二十二分贝的音量轻声运行,大约相当于微风吹拂树叶的沙沙声。我感到一种深沉而可怕的情绪在心中被触发,这是任何美国人都熟悉的——消费欲望。我想要 Neo 在我家里。我想要它,立刻、迫切、马上。

Sleeper took me to a demonstration kitchen, where Neo effortlessly slotted dishes into a wire rack. I was impressed by the robot’s dexterity and fluidity of motion. Then I noticed a tele-operator standing to the side, wearing a V.R. headset, holding controllers, and dictating Neo’s every move. The robot I was seeing was a marionette. 1X declined to show me a demonstration of the A.I. that will power Neo. Sleeper told me that previous versions of Neo had tended to fall over. 1X said that this problem had been resolved, but, when I asked Sleeper if the latest models always remained upright, he said, “To say it doesn’t fall is, like, a total stretch.” Still, when I asked if he believed 1X would meet its deadline for home delivery in 2026, he answered with no hesitation: “It’s a promise we’ve made to the world, and a promise we’ve gotta keep.”

斯利珀带我来到一个演示厨房,Neo 在那里毫不费力地将盘子插入沥水架。机器人的灵巧和动作的流畅给我留下了深刻印象。然后我注意到旁边站着一位远程操作员,戴着 VR 头显,握着控制器,指挥着 Neo 的每一个动作。我看到的机器人是一个提线木偶。1X 拒绝向我展示将驱动 Neo 的人工智能演示。斯利珀告诉我,之前版本的 Neo 容易摔倒。1X 表示这个问题已经解决,但当我问斯利珀最新型号是否总能保持直立时,他说:“说它从不摔倒,那也太夸张了。”尽管如此,当我问他是否相信 1X 能在 2026 年按期交付家用机器人时,他毫不犹豫地回答:“这是我们向世界做出的承诺,也是我们必须遵守的承诺。”

Neo is one of a dozen humanoid robots scheduled to hit the market in the next twelve months. Its rivals include Figure’s 03 Humanoid, Boston Dynamics’s Atlas, and Apptronik’s Apollo. In January, Musk announced that he was converting a portion of a Tesla automotive plant in Fremont, California, into a robot-production factory, and hoped eventually to make a million Optimus robots there annually. His ultimate objective is to have robots working on the production line, making more robots, which he has called an “infinite money glitch.”

Neo 是未来十二个月内计划上市的十几款人形机器人之一。它的竞争对手包括 Figure 的 03 Humanoid、波士顿动力(Boston Dynamics)的 Atlas 和 Apptronik 的 Apollo。今年一月,马斯克宣布他将把特斯拉位于加州弗里蒙特的一座汽车工厂的一部分改造成机器人生产工厂,并希望最终每年在那里制造一百万台 Optimus 机器人。他的最终目标是让机器人在生产线上工作,制造更多机器人,他称之为“无限金钱漏洞”。

In China, efforts are equally ambitious. Unitree, based in Hangzhou, shipped more than five thousand of its G1 humanoids last year, making the company one of the world’s leading suppliers. The G1 stands about four and a half feet tall and has a hole in its head where a face would go. Its motors are loud, it clomps when it walks—it basically has no charm at all. Starting at around fourteen thousand dollars, it is also one of the more affordable advanced robots on the market, and, because it can run open-source software, it has become a favorite of academic robotics researchers. The G1 has also gained popularity with home-robotics enthusiasts—during the recent N.B.A. Finals, a G1 in a Knicks jersey attended watch parties outside Madison Square Garden.

在中国,努力同样雄心勃勃。总部位于杭州的宇树科技(Unitree)去年出货超过五千台 G1 人形机器人,使该公司成为全球领先的供应商之一。G1 身高约四英尺半,头部本该是脸的地方有一个洞。它的电机噪音很大,走路时咚咚作响——基本上毫无魅力可言。起售价约一万四千美元,它也是市面上价格较亲民的先进机器人之一,而且由于能运行开源软件,它已成为学术机器人研究人员的宠儿。G1 也受到家用机器人爱好者的欢迎——在最近的 NBA 总决赛期间,一台穿着尼克斯队球衣的 G1 在麦迪逊广场花园外参加了观赛派对。

I first encountered the G1 at the robotics lab of Aaron Ames, a professor of mechanical engineering at Caltech. Wearing a V.R. headset and holding controllers, I moved the robot back and forth a bit, then had it do some light calisthenics. It seemed shaky, almost drunk. I tried to make it clap its hands, but the G1 resisted. “Try it vertically,” a graduate student at the lab suggested. I did, and the robot pantomimed a delicate golf clap, without ever quite touching its hands together.

我第一次见到 G1 是在加州理工学院机械工程教授亚伦·埃姆斯(Aaron Ames)的机器人实验室。我戴着 VR 头显,握着控制器,让机器人来回移动了一下,然后让它做了一些轻微的健美操。它看起来摇摇晃晃,几乎像喝醉了。我试图让它拍手,但 G1 抗拒了。“试试垂直方向拍,”实验室里一位研究生建议道。我照做了,机器人模仿出一个优雅的高尔夫式轻拍,双手却始终没有真正碰到一起。

Later, in his office, Ames spoke for more than ten minutes, at a cadence of roughly two hundred words per minute, delivering a comprehensive takedown of 1X’s efforts. Ames believes that, no matter how sophisticated the engineering is, reliable A.I. for autonomous robots is years away. “I don’t know how 1X is actually going to get away with it,” Ames said. “I would worry about the legal ramifications when one of those robots falls on a person.”

后来,在埃姆斯的办公室里,他以每分钟大约两百个词的语速,滔滔不绝地讲了十多分钟,对 1X 的努力进行了一番全面的批驳。埃姆斯认为,无论工程设计多么精妙,用于自主机器人的可靠人工智能仍需多年时间。“我不知道 1X 到底打算怎么蒙混过关,”埃姆斯说。“我会担心当其中一台机器人砸到人身上时,会带来什么样的法律后果。”

Many industrial roboticists I spoke with were suspicious of the humanoid push. They told me that there are no standardized benchmarks for progress, and that videos of remarkable feats were often culled from hundreds of takes. In March, I talked with Modar Alaoui, a venture capitalist who invests in humanoid startups. He enthused about Iron, a robot from the Chinese manufacturer Xpeng, whose catwalk-inspired strut is so realistic that exhibitors have removed panels from its body to prove that it’s not a human in a robot suit. “Because of Xpeng’s work, I tell people not to work on locomotion anymore,” Alaoui said. “Locomotion is solved.” The day after I spoke to Alaoui, an Iron robot at a public demonstration at a mall in Shenzhen suddenly seized and fell over—it looked like it was having a stroke. The robot, which weighs more than a hundred and fifty pounds, could not recover, and three human handlers had to drag it away.

我交谈过的许多工业机器人专家对人形机器人的这股热潮持怀疑态度。他们告诉我,目前没有衡量进展的标准化基准,那些展示惊人壮举的视频往往是从数百次拍摄中挑选出来的。今年三月,我与投资人形机器人初创公司的风险投资家莫达尔·阿拉维(Modar Alaoui)交谈。他对中国制造商小鹏(Xpeng)的机器人 Iron 赞不绝口,其受 T 台走秀启发的步态逼真到参展商不得不拆掉它身上的面板,以证明它不是穿着机器人套装的人类。“因为小鹏的工作,我告诉人们不要再研究运动行走了,”阿拉维说。“运动行走问题已经解决了。”在我与阿拉维交谈的第二天,深圳一家商场公开演示中的一台 Iron 机器人突然僵住并摔倒——看起来像是中风了。这台重逾一百五十磅的机器人无法恢复,三名人类操作员不得不把它拖走。
Image may contain Adult Person Clothing Pants Machine Wheel Escooter Transportation Vehicle and Tire
The founders of Skild are trying to make an A.I.-powered brain that can be put into any robot body.
Skild的创始人正试图打造一个能装入任何机器人躯体的AI大脑。
Photograph by Ian Allen for The New Yorker

To celebrate Lunar New Year this February, Unitree staged a display of dozens of G1s doing a choreographed kung-fu routine. The performance was real but misleading: the robots were executing a preprogrammed script, most likely derived from true wushu masters in motion-capture suits. And robotic martial artists can be dangerous to bystanders, as was recently demonstrated at a Chinese amusement park, where a G1 wearing a clown wig kicked a small child in the stomach. Outside of programmed scripts, robots struggle to perform autonomous actions in uncontrolled environments. “The same robot that can land a backflip might not be able to walk up a flight of stairs,” Parada, from Google, told me.

为庆祝今年二月的农历新年,宇树科技安排了一场数十台 G1 表演编排好的功夫套路。表演是真实的,但具有误导性:机器人执行的是预编程脚本,很可能源自真正的武术大师在动作捕捉服中的演示。而且,机器人武术家对旁观者可能很危险,正如最近在中国一家游乐园所展示的那样,一台戴着假发小丑的 G1 踢了一个小孩的肚子。在编程脚本之外,机器人在不受控的环境中难以执行自主动作。“同一台能完成后空翻的机器人,可能无法走上一段楼梯,”谷歌的帕拉达告诉我。

Accidents are not the only concern; Neo might also obey a malicious command. Imagine a young child asking Neo to bash its robotic head against the dining-room table. Guardrails installed in the A.I. would almost certainly prevent the robot from following the command, at least initially. But research shows that such guardrails can be circumvented. If the child were persistent enough (in the way that small children tend to be) and creative enough (in the way that small children tend to be) he or she might succeed in getting the robot to comply.

事故并非唯一的担忧;Neo 也可能服从恶意指令。想象一个小孩让 Neo 用它的机器脑袋猛撞餐桌。人工智能中安装的护栏几乎肯定能阻止机器人服从这一指令,至少最初是这样。但研究表明,这类护栏可以被绕过。如果孩子足够执着(就像小孩子常有的那样)且足够有创意(也像小孩子常有的那样),他或她或许能成功让机器人照做。

Robots can also be hacked. Last year, the security researchers Andreas Makris and Kevin Finisterre found a vulnerability in Bluetooth that would enable them to take control of a fleet of Unitree G1s, infecting one after another and “creating a robot botnet that spreads without user intervention,” they wrote. Humanoid robots need cameras and microphones to make their way through the world, which creates privacy concerns. Plus, their actions are unpredictable. “Babysitting these robots demands an entire operation team,” Jim Fan, a research scientist at Nvidia, wrote at the end of last year. “Mistakes are irreversible and unforgiving.”

机器人也可能被黑客入侵。去年,安全研究人员安德烈亚斯·马克里斯(Andreas Makris)和凯文·菲尼斯特雷(Kevin Finisterre)发现了一个蓝牙漏洞,能让他们控制一整队宇树 G1,一台接一台地感染,“创建一个无需用户干预就能传播的机器人僵尸网络,”他们写道。人形机器人需要摄像头和麦克风来感知世界,这引发了隐私担忧。此外,它们的行为不可预测。“照看这些机器人需要一整个运营团队,”英伟达研究科学家范麟熙(Jim Fan)在去年年底写道。“错误是不可逆的,也是无情的。”

Still, Parada conceded, putting robots in the home is “the ultimate dream.” Growing up in Venezuela, she fantasized about having a robot do her chores. Several roboticists I spoke with seemed a little disappointed by the current state of the field, compared with the expectations they’d had as children—nothing inanimate one day just started moving, nothing ever magically came to life. Sleeper recalled moments as a child when, he said, he “looked at stuffed animals, and got bummed out about how fucking lame they were.” Technology, he told me, sometimes made him sad: “It’s all just going into a screen, instead of into the world.”

尽管如此,帕拉达承认,将机器人送入家庭是“终极梦想”。在委内瑞拉长大时,她就幻想有机器人帮她做家务。我交谈过的几位机器人专家似乎对当前领域的现状有些失望,这与他们儿时的期望相去甚远——没有哪个无生命的东西某天突然动起来,没有什么东西神奇地活过来。斯利珀回忆起儿时的时刻,他说,他“看着毛绒玩具,对它们那副怂样感到沮丧。”他告诉我,科技有时让他难过:“一切都只是进入屏幕,而不是进入世界。”

After I left 1X’s offices, I took my first ride with the autonomous ride-sharing service Waymo. The Waymo was clean and quiet, and avoided a congested highway by navigating through side streets and frontage roads. But, after delivering me to my destination, the car turned in to a crowded In-N-Out Burger drive-through lane, which can take more than an hour to pass through. I laughed out loud, preparing to report on yet another epic A.I. fail—until the car executed a crisp three-point turn, escaped the line, and drove off.

离开 1X 办公室后,我第一次乘坐了自动驾驶共享出行服务 Waymo。这辆 Waymo 干净安静,它避开了一条拥堵的高速公路,穿行于小街和辅路。但是,在把我送到目的地后,这辆车拐进了一条拥挤的 In-N-Out 汉堡得来速车道,通过那里可能需要一个多小时。我笑出声来,准备报道又一次史诗般的人工智能失败——直到这辆车执行了一个干脆的三点掉头,逃离了队伍,开走了。

I later learned that, in uncertain situations like this, a Waymo will sometimes request assistance from a human operator. Professional tele-operation “pilots,” working from the Philippines, oversee galleries of vehicle-camera feeds, alert for construction zones and lines of hungry Californians. Such pilots are also at work in other industries: in Japan, 7-Elevens do “autonomous” stocking using robot arms, which can be controlled from afar by workers wearing V.R. headsets. Human pilots also provide training data for A.I., and Waymo has used at least half a million hours of real-world data to fuel its autonomous-driving model.

我后来了解到,在这种不确定的情况下,Waymo 有时会请求人类操作员协助。专业的远程操作“飞行员”在菲律宾工作,监控着一排排车载摄像头画面,警惕施工区域和饥饿的加州人排起的长队。这类飞行员也在其他行业工作:在日本,7-Eleven 便利店使用机械臂进行“自主”补货,这些机械臂可以由戴着 VR 头显的工人远程控制。人类飞行员也为人工智能提供训练数据,Waymo 已使用至少五十万小时的真实世界数据来驱动其自动驾驶模型。

A car merely has to avoid hitting things. A robot has to manipulate them without breaking them. 1X previously emphasized tele-operation in its marketing, but this generated pushback from potential customers. Now the company emphasizes A.I., but it hasn’t abandoned tele-operation: part of 1X’s gamble is that its customers will come to terms with the fact that a human might be peering into their home through Neo’s eye cameras. (DePue, the early adopter in San Francisco, told me that he would feel safer with a stranger tele-operating his robot than with a stranger cleaning his home.) Tele-operators will sit next to the A.I. team at 1X’s campus in Silicon Valley, and when the lighted rings on Neo’s earpieces change color customers will know that it is being remotely operated. “If you want a robot to bartend your party, we’ll get one of our operators,” Børnich said. “It’s also useful data for us.”

汽车只需避免撞到东西。机器人则必须操控物体而不弄坏它们。1X 此前在营销中强调远程操作,但这引起了潜在客户的反感。现在公司强调人工智能,但并未放弃远程操作:1X 赌注的一部分是,其客户将接受这样一个事实:可能有人正通过 Neo 的眼部摄像头窥视他们的家。(旧金山的早期尝鲜者德普告诉我,他觉得让陌生人远程操作他的机器人,比让陌生人打扫他的家更安全。)远程操作员将坐在 1X 硅谷园区的人工智能团队旁边,当 Neo 耳机上的光圈变色时,客户就会知道它正被远程操控。“如果你想在派对上让机器人当酒保,我们会派一名操作员来,”伯尼奇说。“这对我们来说也是有用的数据。”

Data is to the A.I. revolution what coal was to the industrial one. A.I. communicates in discrete units of data called tokens. In language processing, a token might represent a few letters. In robotics, a token might represent a trajectory associated with a finger joint. But language models have the text of the entire open internet to learn from, not to mention the (uncompensated) copyrighted work of human authors, lifted from the dark web. There exists no equivalent repository—legal or otherwise—of motion trajectories for joints.

数据之于人工智能革命,犹如煤炭之于工业革命。人工智能以称为令牌(token)的离散数据单元进行交流。在语言处理中,一个令牌可能代表几个字母。在机器人技术中,一个令牌可能代表与手指关节相关的运动轨迹。但语言模型有整个开放互联网的文本可供学习,更不用说从暗网攫取的、(未获补偿的)人类作者受版权保护的作品。对于关节运动轨迹,不存在同等的数据库——无论是合法的还是非法的。
Headstone in cemetery listing all of the healthy foods the man had eaten in his lifetime.
Cartoon by Edward Koren
Copy link to cartoon
Shop

One approach to this problem is to radically increase motion-capture efforts. The German robot developer Neura has put more than a thousand industrial workers in motion-capture suits, and is using these data to train humanoids. “Everyone is scared of a shitstorm, because it’s, like, ‘Oh, they’re replacing your work,’ ” David Reger, Neura’s C.E.O., told me. “It’s not! We’re getting, like, a lot of data.” This didn’t seem to address the job-loss question, so I pressed Reger on what would happen to the labor market if his efforts succeeded. He straightened his posture and insisted that, in the next few years, the U.S. would need these robots. “You will have a decrease of working labor, and then you will feel it,” he said. “Right now, you still don’t feel it, but in Europe, we feel it a lot. We think everyone disappeared somehow.”

解决这个问题的一种方法是大幅增加动作捕捉的力度。德国机器人开发商 Neura 已让一千多名产业工人穿上动作捕捉服,并利用这些数据来训练人形机器人。“每个人都害怕引发众怒,因为那就像是,‘哦,他们在取代你的工作,’”Neura 首席执行官大卫·雷格(David Reger)告诉我。“不是这样的!我们是在获取大量数据。”这似乎没有回应失业的问题,于是我追问雷格,如果他的努力成功,劳动力市场会发生什么。他挺直了身板,坚称未来几年美国将需要这些机器人。“你将看到劳动人口减少,然后你就会感受到它,”他说。“现在,你还没感觉到,但在欧洲,我们感受很深。我们觉得人好像都凭空消失了。”

Even if the entire population of Earth donned motion-capture suits, it would take decades to generate the amount of data that was used to train ChatGPT. Another approach is to train robots using videos from head-mounted cameras which have been posted on YouTube. With a little engineering magic, these “egocentric” videos can be turned into joint-trajectory data sets. More advanced video models might be able to extrapolate data from third-person views of real-world experiences such as sports games and cooking shows.

即使地球上所有人都穿上动作捕捉服,也需要几十年才能生成用于训练 ChatGPT 的数据量。另一种方法是利用 YouTube 上头戴摄像头拍摄的视频来训练机器人。通过一些工程魔法,这些“以自我为中心”的视频可以转化为关节轨迹数据集。更先进的视频模型或许能从体育比赛和烹饪节目等真实世界体验的第三人称视角中推断出数据。

Some robotics companies don’t build hardware at all. Instead, they are focussed on helping existing robots navigate the physical world. About two and a half miles north of 1X’s headquarters, I visited the startup Skild AI. The two firms could not be more different. 1X’s robots are exquisite objects of high design. Skild’s lobby looked like an emergency room, with deactivated Unitree robots hanging from cranes and lying on tables. One was on a gurney; another sat in a wheelchair.

一些机器人公司根本不制造硬件。相反,它们专注于帮助现有机器人在物理世界中导航。在 1X 总部以北约两英里半的地方,我拜访了初创公司 Skild AI。这两家公司截然不同。1X 的机器人是高端设计的精美物件。Skild 的大厅看起来像一间急诊室,停用的宇树机器人有的吊在起重机上,有的躺在桌子上。一台在轮床上;另一台坐在轮椅上。

There, I met with Skild’s co-founders, Deepak Pathak and Abhinav Gupta, both professors at Carnegie Mellon University. If Sleeper and Børnich are confident marketers, Pathak and Gupta represent a different Silicon Valley archetype: the entrepreneur whose technological convictions lead him to ignore the pressure of public opinion. This was not the type of company that was going to plant a forest on campus.

在那里,我会见了 Skild 的联合创始人迪帕克·帕塔克(Deepak Pathak)和阿比纳夫·古普塔(Abhinav Gupta),两人都是卡内基梅隆大学的教授。如果说斯利珀和伯尼奇是自信的营销者,那么帕塔克和古普塔则代表了另一种硅谷原型:那种因技术信念而无视舆论压力的企业家。这不是那种会在园区里种一片森林的公司。

Skild is attempting to build a “general purpose” physical A.I.: a single brain that can be slotted into any body. This brain should be able to control not just a humanoid but also a set of detached arms, a robot animal, or a wheeled cart. Pathak and Gupta have to stress-test these robots, so they spend a lot of time kicking them, punching them, and generally harassing them—at one point, they showed me a video of a man using a chainsaw to cut off the legs of a robotic dog. Within seconds, the dog began walking on the stumps.

Skild 试图构建一个“通用”物理人工智能:一个可以插入任何身体的单一大脑。这个大脑不仅能控制人形机器人,还能控制一组分离的机械臂、一只机器动物或一辆轮式推车。帕塔克和古普塔必须对这些机器人进行压力测试,所以他们花大量时间踢它们、打它们,总之就是不断骚扰它们——有一次,他们给我看了一段视频,一个男人用电锯锯掉了一只机器狗的腿。几秒钟内,这只狗就开始用残肢行走。

I did not initially understand why Skild was treating its robots this way. (Nor did I understand why the chainsaw video was scored to hair metal.) Pathak explained that the ability to compensate for injury is one of the most important aspects of generalized physical intelligence. In Skild’s workshop were dozens of partially assembled robots, from a variety of manufacturers; one was wandering around without a head. Midway through the tour, we stopped in front of a robotic dog. One of its eyes was disabled, and its wiring was exposed; an employee was kicking it repeatedly. I was invited to join in.

我起初不明白 Skild 为何如此对待它的机器人。(我也不明白为什么那段电锯视频要配上华丽金属乐。)帕塔克解释说,补偿损伤的能力是通用物理智能最重要的方面之一。在 Skild 的车间里,有几十台来自不同制造商的半组装机器人;其中一台没有头,正在四处游荡。参观中途,我们在一只机器狗面前停下。它的一只眼睛坏了,线路外露;一名员工正反复踢它。我被邀请加入。

“Kick it like how you would kick a human,” Pathak said.

“就像你踢一个人那样踢它,”帕塔克说。

“I wouldn’t kick a human,” I said.

“我不会踢人,”我说。

“Don’t overthink it,” Pathak said.

“别想太多,”帕塔克说。

After a moment’s hesitation, I planted my heel on the robot dog’s body and shoved it with force.The dog barrel-rolled across the showroom floor but was upright within half a second of coming to a stop.

犹豫片刻后,我用脚后跟抵住机器狗的身体,用力推了出去。机器狗在展厅地板上翻滚了一圈,但在停下的半秒内就重新站了起来。

Later, in a conference room, Pathak and Gupta explained that their seemingly cavalier attitude toward robot welfare was actually motivated by concern for human safety. (Better that a human kick a robot a thousand times than that a robot kick a human once.) Right now, they don’t think that humanoids are safe enough to use in residential settings—in fact, Pathak refuses to have one in his home. “People are using appearance as a way to misguide the public,” Pathak said. “If you make a robot humanlike, you expect it to have humanlike capabilities. But technology is far behind that.’ ”

后来,在一间会议室里,帕塔克和古普塔解释说,他们看似对机器人福利漫不经心的态度,实际上是出于对人类安全的关切。(人类踢机器人一千次,总好过机器人踢人类一次。)目前,他们认为人形机器人还不够安全,不能用于居住环境——事实上,帕塔克拒绝在自己家里放一台。“人们正在利用外观来误导公众,”帕塔克说。“如果你把机器人做得像人,你就会期望它具备像人一样的能力。但技术还远远落后于此。”

Skild is one of the best-funded robotics startups, but much of its capital is spent on buying computing power from data centers. This is another way of saying that most of its money eventually goes to Nvidia. Nvidia, which has a five-trillion-dollar market capitalization, is the most valuable company in the world. In fact, adjusting for inflation, Nvidia may be the most valuable company in human history; its rival for the title is the Dutch East India Company. Jensen Huang, Nvidia’s C.E.O., has cosmic ambitions. “Wherever the universe has structure, we can turn that into A.I.,” he recently said.

Skild 是资金最充裕的机器人初创公司之一,但其大部分资金都花在了从数据中心购买算力上。这等于说,它的大部分钱最终都流向了英伟达。英伟达市值五万亿美元,是全球最有价值的公司。事实上,经通胀调整后,英伟达可能是人类历史上最有价值的公司;与之争夺这一头衔的是荷兰东印度公司。英伟达首席执行官黄仁勋(Jensen Huang)怀有宇宙级的雄心。“宇宙中任何有结构的地方,我们都能将其转化为人工智能,”他最近说道。

Nvidia moved into robotics in 2014. Around this time, Huang recalls, he watched a neural network teach a robotic arm to shoot a hockey puck; he immediately calculated a rough estimate of the eventual size of the robotics A.I. market. Soon, Nvidia was touting virtual-reality simulators to train robots, and microchips to deploy in robot brains. (The simulators allow developers to experiment with surface friction, mimic faulty joints, and even alter gravity.) These early initiatives struggled at first, but Nvidia was able to establish what a former employee called “vender lock,” making robotics manufacturers dependent on it. Every single robot I encountered while reporting this article ran on an Nvidia microchip, and every single one had trained in an Nvidia digital gym.

英伟达于 2014 年进入机器人领域。黄仁勋回忆说,大约在那个时候,他看到一个神经网络教一只机械臂击打冰球;他立刻粗略估算出机器人人工智能市场的最终规模。很快,英伟达就开始推广用于训练机器人的虚拟现实模拟器,以及部署在机器人大脑中的微芯片。(这些模拟器允许开发者试验表面摩擦力、模拟故障关节,甚至改变重力。)这些早期举措起初举步维艰,但英伟达成功建立了前员工所称的“供应商锁定”,使机器人制造商对其产生依赖。我报道本文时遇到的每一台机器人,都运行在英伟达的微芯片上,每一台都在英伟达的数字健身房里训练过。

Nvidia has also invested in many startups, including Figure, Neura, and Skild. This has produced accusations of “circular” dealmaking from skeptical investors, since Nvidia is providing funds to its own customers, who then use that money to purchase its hardware. Michael Burry, the investor made famous in “The Big Short” for betting against the housing market in the late two-thousands, today warns of “suspicious” revenues at many A.I. companies, including Nvidia. Huang has called such claims “ridiculous,” and defended the investments, saying that they are necessary to expand Nvidia’s “ecosystem.”

英伟达还投资了许多初创公司,包括 Figure、Neura 和 Skild。这引发了持怀疑态度的投资者对“循环”交易的指责,因为英伟达是在向自己的客户提供资金,而这些客户又用这笔钱购买它的硬件。因在 2000 年代末做空房地产市场而闻名于《大空头》的投资人迈克尔·伯里(Michael Burry),如今警告许多人工智能公司(包括英伟达)存在“可疑”收入。黄仁勋称此类说法“荒谬”,并为这些投资辩护,称它们对于扩展英伟达的“生态系统”是必要的。

Nvidia’s robotics microchips are different from the ones they sell for A.I. training. They are “edge” chips, meaning that they do their processing in the robot instead of in a remote data center. The human brain uses about twenty per cent of the energy that the body generates; an Nvidia edge chip can drain as much as sixty per cent of a robot’s electricity supply. “It’s not the motors that take up most of the battery,” Cardenas, the C.E.O. of Apptronik, told me. “It’s actually the compute.”

英伟达的机器人微芯片不同于他们为人工智能训练出售的芯片。它们是“边缘”芯片,意味着它们在机器人内部而非远程数据中心进行处理。人脑消耗身体产生能量的约百分之二十;一块英伟达边缘芯片可能消耗机器人高达百分之六十的电力供应。“消耗电池最多的不是电机,”Apptronik 首席执行官卡德纳斯告诉我。“实际上是计算。”

But, if a robot needs more brainpower, it can draw it from a local server via Wi-Fi. A robot’s A.I., like most computer software, is regularly upgraded via the internet; robots get smarter every week. Most importantly, they have a hive mind. “Whatever one robot learns, the others all learn at the same time,” Børnich said.

但是,如果机器人需要更多脑力,它可以通过 Wi-Fi 从本地服务器获取。机器人的 AI,如同大多数计算机软件,通过互联网定期升级;机器人每周都在变得更聪明。最重要的是,它们拥有蜂巢思维。“一台机器人学到的东西,其他所有机器人都会同时学到,”伯尼奇说。

This means that the wider the release of humanoids, the more skilled they become. If a thousand humanoid robots are sent to fold towels, they’ll learn from one another’s mistakes, and will adapt far faster than a single robot would. This network effect may account for 1X’s eagerness to get robots in the home. “There is definitely a data flywheel there,” Børnich said.

这意味着人形机器人投放得越广,它们就变得越熟练。如果一千台人形机器人被派去叠毛巾,它们会从彼此的错误中学习,适应速度将远超单台机器人。这种网络效应或许可以解释 1X 为何急于让机器人进入家庭。“那里肯定存在一个数据飞轮,”伯尼奇说。

“Robot” is one of a small number of words that have entered the English language via Czech. (“Polka” and “howitzer” are others.) The term was first used in Karel Čapek’s play “Rossum’s Universal Robots,” from 1920. In the play, an industrial robotics conglomerate has transformed the world economy: robots do most of the work, and they are also deployed as soldiers. At the same time, the human birth rate declines to zero. Eventually, the robots take over, and kill almost everyone on Earth.

“Robot”(机器人)是少数几个经由捷克语进入英语的词汇之一。(“Polka”波尔卡舞和“howitzer”榴弹炮是另外两个。)这个词首次使用是在卡雷尔·恰佩克(Karel Čapek)1920 年的剧作《罗素姆的万能机器人》中。剧中,一家工业机器人集团改变了世界经济:机器人承担了大部分工作,还被部署为士兵。与此同时,人类出生率降至零。最终,机器人接管了一切,杀死了地球上几乎所有人。

Any extended discussion of robots, I have found, will eventually reference the apocalypse. Even in more positive treatments, such as Isaac Asimov’s “I, Robot,” the androids end up taking over. When I broached the subject of the robot revolution with Børnich, I sensed a certain fatigue. “Look, they won’t seek revenge on us,” he said. “We can wipe their minds.”

我发现,任何关于机器人的深入讨论,最终都会提及末日。即使在更积极的处理方式中,如艾萨克·阿西莫夫(Isaac Asimov)的《我,机器人》,机器人最终也接管了世界。当我向伯尼奇提起机器人革命这个话题时,我感觉到一丝倦怠。“听着,它们不会向我们复仇的,”他说。“我们可以抹去它们的记忆。”

Perhaps the fearful visions of Čapek—not to mention those of James Cameron, in his “Terminator” movies—have tainted the West’s cultural perception of robots. David Reger, Neura’s C.E.O., told me that American journalists ask him anxious questions about robots, but South Korean journalists just seem impatient. “They’re tired of waiting!” he said. “They ask, ‘When is it coming?’ ” Reger’s observation is supported by data. A Stanford University analysis of opinion polls across three dozen countries found that respondents in China, South Korea, Thailand, and Indonesia were the most excited about consumer applications of A.I. Respondents in the U.S. and Canada were the least.

或许恰佩克——更不用说詹姆斯·卡梅隆(James Cameron)在《终结者》系列电影中——那些令人恐惧的想象,已经玷污了西方对机器人的文化认知。Neura 首席执行官雷格告诉我,美国记者会问他关于机器人的焦虑问题,但韩国记者似乎只是不耐烦。“他们等得不耐烦了!”他说。“他们问,‘什么时候能来?’”雷格的观察有数据支持。斯坦福大学一项针对三十多个国家民意调查的分析发现,中国、韩国、泰国和印度尼西亚的受访者对人工智能的消费应用最为兴奋。美国和加拿大的受访者则最不兴奋。

Most of the roboticists I spoke with seemed more excited by the industrial applications of humanoid robots than by the prospect of their doing domestic chores. (“You’ll see them in McDonald’s first,” Rev Lebaredian, who runs physical A.I. simulation at Nvidia, said.) Apptronik’s Apollo humanoid—porcelain white, with knobby limbs and thick, sturdy legs—has been used in production at European automotive factories owned by Mercedes-Benz. For now, the robot is limited to activities such as picking out parts and loading conveyors. It needs to be fairly strong to complete this task, which means that a malfunction might hurt someone. For this reason, the robots are fenced off from the rest of the workers.

我交谈过的大多数机器人专家,似乎对人形机器人的工业应用比对它们做家务的前景更感兴趣。(“你会先在麦当劳看到它们,”英伟达负责物理人工智能模拟的雷夫·勒巴雷迪安(Rev Lebaredian)说。)Apptronik 的 Apollo 人形机器人——瓷白色,四肢有节,双腿粗壮结实——已在梅赛特斯-奔驰旗下的欧洲汽车工厂投入生产使用。目前,该机器人仅限于拣选零件和装载传送带等活动。完成这项任务需要它相当强壮,这意味着一旦发生故障可能会伤人。因此,这些机器人被围栏隔开,与其他工人保持距离。

Humanoids make up only a tiny fraction of the five million or so robots currently deployed in factories worldwide. Mike Haley, a senior researcher at Autodesk, a maker of industrial-design software, has spent much of his career programming these specialized, non-humanoid factory bots. He told me that he has never seen a humanoid do anything useful in an industrial setting. Simpler robotics solutions, like detached arms for spraying paint, or autonomous forklifts for moving pallets, are more efficient, cost less, have fewer moving parts, and require less maintenance. “I truly believe that we have a moment a few years down the road from now where we look back at this humanoid thing and go, ‘Boy, that was bloody stupid,’ ” Haley said.

人形机器人仅占全球工厂目前部署的约五百万台机器人的极小部分。工业设计软件制造商 Autodesk 的高级研究员迈克·黑利(Mike Haley)职业生涯的大部分时间都在为这些专用的非人形工厂机器人编程。他告诉我,他从未见过人形机器人在工业环境中做过任何有用的事。更简单的机器人解决方案,如用于喷漆的分离式机械臂,或用于搬运托盘板的自主叉车,效率更高,成本更低,活动部件更少,所需维护也更少。“我真心相信,几年后的某个时刻,我们会回看这股人形机器人热潮,然后说,‘天哪,那真是蠢透了,’”黑利说。

Commercial service robots are already commonplace in China, but these are cheap, special-purpose devices that deliver food or fold laundry. Expensive humanoids may not be needed for these tasks. At Skild, Pathak’s solution is to isolate a robot’s parts to fit the task at hand. “Let’s say you’re putting a robot in a factory like this, and you’re assembling something—should you really spend energy on the legs?” he asked. “I can separate the torso and put it on the table, then use the legs to carry something else.”

商用服务机器人在中国已经司空见惯,但这些是廉价、专用的设备,用于送餐或叠衣服。这些任务可能不需要昂贵的人形机器人。在 Skild,帕塔克的解决方案是分离机器人的部件以适应手头的任务。“比方说,你把一个机器人像这样放进工厂,用来组装东西——你真的应该把能量花在腿上吗?”他问道。“我可以把躯干拆下来放在桌子上,然后用腿去搬运别的东西。”

It seems that 1X is taking the Apple approach, designing beautiful products from scratch and controlling the entire manufacturing process. Skild is taking the Android approach, writing cross-platform software and running it on hardware made by other companies. Apptronik, Unitree, Tesla, and others—maybe they will be like Nokia, or Blackberry. Or maybe one of them will win.

看起来,1X 走的是苹果路线,从零开始设计精美产品,并控制整个制造过程。Skild 走的是安卓路线,编写跨平台软件,并在其他公司制造的硬件上运行。Apptronik、宇树、特斯拉等——也许它们会像诺基亚或黑莓。又或许其中一家会胜出。

About seven or eight million years ago, a branch of primates broke off and started developing superior brains, dexterous hands, and the capacity for language. That branch diversified into Australopithecus, and Ardipithecus, and Homo erectus. Eventually, Homo sapiens emerged and took over. Surveying the diversity of humanoid forms at this early stage of robotics, I could see something similar occurring. The essential form—two arms, two legs, a torso, and a head—was consistent, but there were many variations. Agility’s Digit has a square head and inverted kneecaps, like a grasshopper. Apptronik’s Apollo uses a classic sci-fi design, with digital readouts on its mouth and chest. Boston Dynamics’s Atlas has a toylike carapace and a glowing, hollow circle for a head. Tesla’s Optimus, Neura’s 4NE-1, and Figure’s 03 have all converged on similarly anonymous and ever so slightly sinister designs: white bodies with smooth black faceplates. One of these is likely the evolutionary ancestor of all that is to come.

大约七八百万年前,灵长类动物的一个分支分化出来,开始发育出更优越的大脑、灵巧的双手和语言能力。那个分支多样化为南方古猿、地猿和直立人。最终,智人出现并接管了一切。审视机器人技术早期阶段人形形态的多样性,我能看到类似的事情正在发生。基本形态——两条手臂、两条腿、一个躯干和一个头——是一致的,但存在许多变体。Agility 的 Digit 有一个方形的头和像蚱蜢一样倒置的膝盖骨。Apptronik 的 Apollo 采用经典的科幻设计,嘴部和胸部有数字读数器。波士顿动力的 Atlas 有一个玩具般的外壳和一个发光的空心圆环作为头部。特斯拉的 Optimus、Neura 的 4NE-1 和 Figure 的 03 都趋同于类似匿名且略带阴森的设计:白色机身配光滑的黑色面板。其中一款很可能就是未来一切事物的进化祖先。

The missing link might be the hand. The human hand is one of three pinnacles of hominid evolution, along with the brain and the voice box. It is by far the most capable manipulator in the animal kingdom, and can perform twenty-seven independent motions. Robotic hands lag far behind: we are years away from a robot that can both tie its shoes and shuffle a deck of cards.

缺失的一环可能是手。人手是原始人类进化的三大巅峰之一,与大脑和喉头并列。它是动物界中迄今为止能力最强的操作器,能完成二十七种独立动作。机器人手则远远落后:我们距离一台既能系鞋带又能洗牌的机器人,还有数年之遥。

At 1X, I saw rows of hands and forearms, denuded of housing and grasping toward the ceiling. Technicians seated along an assembly line were fastening artificial tendons to the fingers to link them with actuators in the wrist. Others attached electrical wire to sensors on the fingertips; the wire would eventually be routed to the robot’s brain. Nearby, a finger was being tested for durability. It pointed, curled back, then repeated the movements. A counter on a monitor noted that this action had been performed 2,860,631 times.

在 1X,我看到一排排手和前臂,外壳被剥去,朝天花板抓握着。坐在装配线上的技术人员正将人造肌腱固定在手指上,将它们与手腕中的致动器连接起来。其他人则将电线连接到指尖的传感器上;电线最终将被引向机器人的大脑。附近,一根手指正在接受耐久性测试。它指向,蜷起,然后重复这些动作。显示器上的计数器显示,这个动作已经执行了 2,860,631 次。

But here, again, the mechanical engineering was far ahead of the A.I. This point was confirmed during a demonstration at Skild, where I watched a Unitree robot carefully grab a white ceramic coffee cup from a table and set it upright in a bin. I moved the coffee cup back to the tabletop, and the robot’s eyelike cameras seemed to follow. Then I turned the cup on its side. The robot seemed perplexed. It moved its hand about six inches from the cup and began to grab uselessly in the air.

但在这里,机械工程再次远远领先于人工智能。这一点在 Skild 的一次演示中得到了证实,我观看了一台宇树机器人小心地从桌上抓起一个白色陶瓷咖啡杯,并将其直立放入一个箱子里。我把咖啡杯移回桌面,机器人眼睛般的摄像头似乎在跟随。然后我把杯子侧放。机器人似乎困惑了。它将手移到离杯子约六英寸的地方,开始徒劳地在空中抓握。

With a disapproving look, a technician set the cup upright, and the robot immediately returned it to the container. I took a blue plastic bowl from the bin and placed it upside down. The robot identified the object but couldn’t get a grip on it. As its hand pushed the bowl around fruitlessly, I gave a little smirk of human superiority. But then the robot stopped. And—I swear to this—it started to think. After about ten seconds, it perceived that the overturned bowl had a raised circular foot. Pinching the foot of the bowl, the robot picked it up and placed it in the bin. “See?” the technician said. “It figured it out.” ♦

一位技术人员带着不以为然的表情,把杯子扶正,机器人立刻将它放回了容器。我从箱子里拿出一个蓝色塑料碗,把它倒扣着放下。机器人识别出了这个物体,但无法抓住它。当它的手徒劳地把碗推来推去时,我露出了一丝人类优越感的得意微笑。但随后机器人停了下来。而且——我对此发誓——它开始思考了。大约十秒钟后,它察觉到倒扣的碗有一个凸起的圆形底足。机器人捏住碗的底足,把它拿起来放进了箱子。“看到了吗?”技术人员说。“它自己想出办法了。” ♦

2 free previews left — register to get 30 translations a month