原文链接：The Man Who Made Robots Dance Now Wants Them to Think for Themselves | WIRED
原文：
Anyone currently worrying about artificial intelligence taking over the world may want to swing by the Boston Dynamics AI Institute in Cambridge, Massachusetts. While walking around, they’d see that the robots that might lead a future uprising are still trying to tie their shoelaces, metaphorically speaking.
The Institute’s founder and executive director, Marc Raibert, has built some of the world’s most famous robots at his previous venture, Boston Dynamics. The company, acquired by Hyundai in 2020, has developed legged machines capable of running, leaping, and of course dancing with spryness that can veer into the uncanny.
Raibert’s creations include the four-legged, pony-sized Big Dog; its smaller dog-like buddy, Spot; and an acrobatic humanoid called Atlas. They have racked up influencer-levels of YouTube views and likes, found their way into comedy skits, and even inspired dystopian Black Mirror episodes.
The future shock inspired by Boston Dynamics’ robots can obscure the fact that off-camera it is humans providing most of the intelligence needed for their most impressive and daring stunts. Raibert’s AI institute, launched in August 2022 with Hyundai’s support, is working on ways to take humans out of the loop. It will research ways for robots to comprehend and tackle complex and unpredictable situations with little or no human help. Raibert sat down with WIRED at the institute’s headquarters to discuss his new venture.
Will Knight: When did you decide to pivot from focusing on robots’ physical capabilities to working on their intelligence?
Marc Raibert: It’s been a while that I’ve been frustrated, if you want to call it that, with how much work it’s been to get the robot to do each next thing. You need substantial resources, and it’s going to take years to accomplish at the level I’d love to see. The athletic part of robotics is really doing well, but we need the cognitive part.
And the parkour that we see Boston Dynamics robots doing is an example of that painstaking programming and engineering work?
Oh yeah, there’s a lot of work that goes into that.
There have recently been some big leaps in AI thanks to large language models and systems like ChatGPT. Can this technology help your mission?
We have a significant effort here looking at the role they can play in robotics. I’m an enthusiast for using what you know in combination with what you learn. One of the interesting things about language models is that the language comes from humans, who are embodied creatures. It’s not focused on physicality, but it’s also not devoid of embodiment.
Language models became much more powerful thanks to scaling up the training data and computing power involved. Could something similar occur in robotics?
I think that's starting to happen. Marco Hutter at ETH Zurich is a visitor here, and we're going to use some of his work. He’s been working on reinforcement learning, largely developed in simulation but then applied to physical robots. It’s in the same neighborhood as large language models, in that you're letting machines discover data and then putting the data together rather than someone hand-designing a solution. He's got robots climbing on things very impressively, using different parts of their bodies to climb without ever having encountered that particular obstacle before, because in simulation they've encountered so many different environments.
A key question for AI researchers these days is whether it is possible to move beyond the capabilities of large language models without giving machines some sort of physical form. Could the work you are doing help other forms of AI advance?
I think this is a great question—it's not like ChatGPT does everything. A child doesn’t have anything like the data that a large language model does and yet their ability to generalize solutions is just remarkable. And you’d certainly like to burn the kind of energy that a kid does rather than what a language model does. The physical and cognitive fit together.
Several companies seem to be following Boston Dynamics’ lead by developing humanoids—including Tesla. What do you think of that trend?
I used to think all that matters is functionality—mobility, dexterity, agility, the ability to see things in the environment, and some kind of intelligence—and it shouldn't matter what the robot looks like.
But if you look at the reaction to our robots, humanoids get 10 times the reaction to anything else. So if you care about people responding, you have to care about that. At Boston Dynamics we got a fantastic reaction to the “Do You Love Me” video, and contrary to what some people think, we did it for pure fun.
I would never count Elon Musk out. He has a big voice that's helping make it feasible for other people to do humanoids, even though I think Boston Dynamics blazed the trail.
Marc RaibertCourtesy of Boston Dynamics AI Institute
Some robots, including from Boston Dynamics, are already doing warehouse work. Do you expect to see that humanoid workforce growing in the coming years?
Warehouse work is one of the great applications at the moment because there's such a need for people. Employers are really frustrated with the difficulty of getting humans and training them and the turnover. You can organize the environment, but then that’s less interesting for people like us who want to, you know, make the real future happen.
What are some of the more challenging environments that the institute will try to have robots master?
One group is working on robots that repair bicycles. The dream is to be able to fix any bicycle, which isn't as hard as fixing a car, but it's also not just stacking boxes. There’s a bicycle out there in the lab, and the robot is doing some stuff, but it’s early days.
Another thing we're doing is building an ultra-mobile two-wheeled vehicle—a bicycle with a jumping and bouncing mechanism on it. It will also have a vision system and will plan. If you watch a parkour cyclist, they are jumping from object to object, which requires a lot of understanding of the environment and planning out a sequence.
The third large-scale project is dexterous manipulation. When you look at the world of robotics, this might be a biased opinion, but I feel that locomotion and mobility has come a long way, but manipulation has a much longer way to go, even though we've been working on it for 50 years. Intelligence is a key part—integrating cognitive intelligence and perception.
People sometimes feel creeped out by the robots you’ve built. Do you think that could get worse now that systems like ChatGPT have prompted more discussion around AI potentially becoming dangerous?
Most bad things that happen in the world aren’t coming from robots. And the question I have is, are people really afraid of robots? My experience of going around with Spot is that people love to engage with it, they love to have selfies with it, and they will pet the robot. Even in individuals, depending on how you pose the question, you could get people saying, “Yeah, I’m afraid of that robot but not afraid of this other robot.” I’d just like to know what the real story is.
Are you at all worried that the leaps we’ve seen in AI could bring new risks?
I'm not afraid at all. I'm a little surprised that Musk and Sam Altman are so out there saying we need to be careful. People worry about robots taking every job, but we're really trying to get them to do one job in most cases, or a couple of jobs. Will we get there someday? You know, maybe, but it’s going to be a long haul.
译文：
现在担心人工智能统治世界的人，不妨到位于马萨诸塞州剑桥市的波士顿动力人工智能研究所（Boston Dynamics AI Institute）转转。走一走，他们就会发现，未来可能领导起义的机器人还在努力系鞋带，比喻一下。
该研究所的创始人兼执行董事马克-雷伯特（Marc Raibert）曾在他之前的企业波士顿动力公司（Boston Dynamics）制造了一些世界上最著名的机器人。这家于 2020 年被现代汽车收购的公司开发出了能够奔跑、跳跃的腿部机器，当然也能跳舞，其敏捷程度令人匪夷所思。
雷伯特的作品包括四条腿、小马般大小的 “大狗”、体型更小的狗伙伴 “斑点”，以及一个名叫 “阿特拉斯 ”的杂技人形机器人。它们在 YouTube 上的浏览量和点赞量都达到了有影响力的水平，还被搬上了喜剧短剧的舞台，甚至激发了《黑镜》中乌托邦式情节的灵感。
波士顿动力公司（Boston Dynamics）的机器人所带来的未来震撼可能会掩盖这样一个事实，即在镜头之外，是人类提供了它们最令人印象深刻和最大胆的特技所需的大部分智能。在现代汽车公司的支持下，雷伯特的人工智能研究所于 2022 年 8 月成立，该研究所正在研究如何让人类脱离这个圈子。它将研究如何让机器人在很少或没有人类帮助的情况下理解和处理复杂而不可预测的情况。雷伯特在该研究所的总部与《WIRED》记者进行了座谈，讨论了他的新事业。
威尔-奈特 你是什么时候决定从关注机器人的物理能力转向研究它们的智能的？
马克-雷伯：有一段时间我一直很沮丧--如果你愿意这么说的话--要让机器人做下一件事需要做很多工作。你需要大量的资源，而且要花费数年时间才能达到我希望看到的水平。机器人技术的运动部分确实做得很好，但我们需要认知部分。
我们看到波士顿动力公司的机器人所做的跑酷就是这种艰苦编程和工程工作的一个例子？
是的，这需要大量的工作。
最近，由于大型语言模型和 ChatGPT 等系统的出现，人工智能有了很大的飞跃。这项技术能帮助你们完成任务吗？
我们在这里做了大量工作，研究它们在机器人技术中可以发挥的作用。我热衷于将你所知道的与你所学的结合起来使用。语言模型的一个有趣之处在于，语言来自于人类，而人类是具身的生物。它并不注重物理性，但也并非没有体现。
由于训练数据和计算能力的提升，语言模型变得更加强大。机器人技术中会出现类似的情况吗？
我想这已经开始了。苏黎世联邦理工学院的马可-胡特（Marco Hutter）是这里的访客，我们将使用他的一些研究成果。他一直在研究强化学习，主要是在模拟中开发的，但后来应用到了实体机器人上。这与大型语言模型属于同一领域，即让机器发现数据，然后将数据整合在一起，而不是由人工设计解决方案。他让机器人在各种物体上攀爬，令人印象深刻，它们使用身体的不同部位进行攀爬，而之前从未遇到过特定的障碍物，因为在模拟过程中，它们遇到过许多不同的环境。
如今，人工智能研究人员面临的一个关键问题是，在不赋予机器某种物理形态的情况下，是否有可能超越大型语言模型的能力。你们所做的工作能否帮助其他形式的人工智能取得进步？
我认为这是一个很好的问题--ChatGPT 并非无所不能。一个孩子所拥有的数据远不及一个大型语言模型，但他们归纳解决方案的能力却非常出色。你当然也希望像孩子那样消耗能量，而不是像语言模型那样。物理和认知的结合
多家公司似乎都在效仿波士顿动力公司开发仿人机器人，包括特斯拉。你如何看待这一趋势？
我曾经认为，最重要的是机器人的功能--移动性、灵巧性、敏捷性、观察环境的能力以及某种智能--机器人长什么样并不重要。
但如果你看看人们对机器人的反应，人形机器人的反应是其他任何东西的十倍。因此，如果你关心人们的反应，你就必须关心这一点。在波士顿动力公司，我们的 “你爱我吗 ”视频引起了巨大反响，与某些人的想法相反，我们这样做纯粹是为了好玩。
我永远不会把埃隆-马斯克排除在外。尽管我认为波士顿动力公司开辟了一条新路，但他的声音很大，帮助其他人实现了人形机器人的可行性。
马克-雷伯波士顿动力人工智能研究所提供
包括波士顿动力公司（Boston Dynamics）在内的一些机器人已经开始从事仓储工作。您预计未来几年仿人机器人的数量会增长吗？
仓库工作是目前最重要的应用之一，因为对人的需求量非常大。雇主们对难以获得人力、难以培训人力以及人员流动感到非常沮丧。你可以组织环境，但对于像我们这样想要实现真正未来的人来说，这就不那么有趣了。
研究所将尝试让机器人掌握哪些更具挑战性的环境？
有一个小组正在研究修理自行车的机器人。他们的梦想是能够修理任何自行车，这不像修理汽车那么难，但也不仅仅是堆放箱子。实验室里有一辆自行车，机器人正在做一些事情，但现在还为时尚早。
我们正在做的另一件事是制造一辆超移动双轮车--一辆装有跳跃和弹跳装置的自行车。它还会有一个视觉系统，并会进行规划。如果你看一个跑酷单车手，他们会从一个物体跳到另一个物体，这需要对环境有很多了解，并规划出一个顺序。
第三个大型项目是灵巧操纵。纵观机器人世界，我的观点可能有失偏颇，但我觉得运动和移动能力已经取得了长足进步，但操纵能力还有很长的路要走，尽管我们已经在这方面努力了 50 年。智能是一个关键部分，它整合了认知智能和感知智能。
人们有时会对你制造的机器人感到毛骨悚然。你是否认为，现在像 ChatGPT 这样的系统引发了更多关于人工智能可能变得危险的讨论，这种情况会变得更糟？
世界上发生的大多数坏事都不是来自机器人。我的问题是，人们真的害怕机器人吗？我和 Spot 一起四处走动的经验是，人们喜欢和它接触，喜欢和它自拍，还会抚摸机器人。即使是个人，也会有人说：“是的，我怕那个机器人，但不怕另一个。” 这取决于你如何提问。我只是想知道真实情况是怎样的。
您是否担心人工智能的飞跃会带来新的风险？
我一点也不担心。马斯克和萨姆-奥特曼如此直言不讳地表示我们需要小心谨慎，这让我有点惊讶。人们担心机器人会抢走所有工作，但我们其实是想让它们在大多数情况下只做一项工作，或者几项工作。有一天我们能做到吗？也许吧，但这将是一个漫长的过程。