发展中国家的截肢患者或许能为类人机器人提供真正实用的手部功能。
Amputees in developing nations may hold the key to giving humanoid robots hands that actually work.
总部位于圣地亚哥的初创公司 Psyonic 希望将这些仿生手提供给人们，并向他们支付报酬，以便收集有关他们如何抓握、举重以及使用这些手部动作的数据。其最终目标是完善这种机器人手的设计——这一直是物理人工智能领域长期面临的挑战。
Psyonic, a San Diego startup, wants to give people its bionic hands — and pay them — so it can capture data on how they grip, lift, and fold, with the ultimate goal of perfecting the robotic hand, a longstanding challenge in physical AI.
该公司成立于 2015 年，最初专注于为人类制造假肢手。但首席执行官 Aadeel Akhtar 表示，近年来许多机器人研发机构纷纷联系他们，希望将这些仿生手用于机器人技术中。
Founded in 2015, the company initially focused on making prosthetic hands for humans, but CEO Aadeel Akhtar said robotics developers in recent years came knocking at his door, interested in using the hands for robots.
Psyonic 目前正处于将这些仿生手推广到全球的早期阶段，重点关注巴基斯坦（Akhtar 的父母来自那里）和印度。用户在使用这些仿生手完成日常任务时，所产生的数据将被用于训练机器人模型，从而使这些假肢使用者成为数据收集的“劳动力”。Akhtar 希望这种方式能够降低那些需要这些仿生手的人的成本。
Psyonic is now in the early stages of an effort to deploy thousands of bionic hands globally — it is specifically eyeing Pakistan, where Akhtar’s parents are from, and India. Users would go about their everyday tasks; the data from every grip would then be used to train a robotics model, making the prosthetic wearers a data-collection workforce. That in turn, Akhtar hopes, can subsidize the cost of the hands for those who need them.
“Adeel Akhtar 对 Semafor 说：‘机器人技术面临数据收集的难题，尤其是在手部动作方面。’许多公司都在寻找新的数据来源来训练机器人：有些公司提供免费的家庭清洁服务以换取用户拍摄相关视频；还有一些公司在印度为工厂工人安装头戴式摄像头。’”Akhtar 解释说，Psyonic 的仿生手能够记录每次抓握动作所产生的力、扭矩以及手指的位置信息，这些数据可以与可穿戴摄像头的视频同步记录下来，从而使机器人模型既能了解动作的具体过程，也能感知到动作时的手感。
“Robotics has a data problem,” especially around hands, Akhtar told Semafor. Companies are scouring for new data sources that can be used to train robots: Some are offering free home cleanings in exchange for filming the work; others are strapping head-mounted cameras on factory workers in India. Akhtar said Psyonic’s hands log the forces, torques, and finger positions of every grip, which can be synced to video from a wearable camera, so a robotics model learns both what the task looks like and what it feels like.
Akhtar 表示，Psyonic 的计划是将训练出的机器人模型出售给相关公司，而不是直接出售原始数据。
Psyonic’s current plan is to sell the resulting models, rather than the data itself, to robotics companies, Akhtar said.
他认为，这同时也是为那些因肢体残缺而失业的人提供经济来源的一种方式；全球有数百万人失去了上肢，其中大多数生活在发展中国家。Psyonic公司正与总部位于伦敦的Koalaa公司合作，开发这种新型“假肢手”——这种假肢手采用柔软的织物材质制成，无需经过多次临床医生的安装调试即可使用。
He also sees it as a way to provide an economic livelihood to people who are unemployed because they are missing limbs; millions of people worldwide have upper-limb amputations, most of them in developing countries. Psyonic is working with Koalaa, a London-based company, to deploy the hands using a soft fabric sleeve — replacing the custom-molded socket that would ordinarily take several clinician visits to set up.
对于机器人工程师来说，人类的手是一个极具挑战性的研究对象，因为它们的结构既紧凑又复杂。工程师通常用“自由度”（即手可以独立运动的种类）来衡量手的复杂性：人类的手大约有27个自由度，而一些最先进的机器人手也只有22个自由度。Barclays的研究员Zornitsa Todorova指出：“人类的手在制造仿生机器人时往往是最昂贵的部分。”
Hands are particularly puzzling for robotics engineers in part because of how compact and complex they are. Engineers count a hand’s complexity in “degrees of freedom” — the number of independent ways it can move. A human hand has around 27; some of the most advanced robotic hands have about 22. Hands are often the biggest single cost in building a humanoid, Barclays researcher Zornitsa Todorova said.
“人类的手内部装有大量的电机、稀土材料以及执行器，”她解释道，“它就像强化版的iPhone一样，所有部件都被紧密地集成在一起。”除了这些技术上的挑战外，人类的手还需要以一种机器人腿部通常无法实现的方式与物理世界进行互动。例如，机器人做后空翻时落在混凝土地上并不会造成太大伤害；但人类的手必须能够“轻轻抓住一颗树莓而不会将其压坏，同时还能举起重达100磅的壶铃并进行挥动。”
“It’s loaded with motors and rare earths and actuators,” she said. “It’s like an iPhone on steroids. Everything is really packed.”On top of those engineering challenges, hands also have to interact with the physical world in a way that robotic legs often don’t. It doesn’t matter whether a robot doing a backflip, for example, lands hard or soft on a concrete floor. But a hand has to be skilled at “grabbing a raspberry without crushing it, but also lifting a 100-pound kettlebell and swinging it,” Akhtar said.
谷歌DeepMind机器人团队的研究总监Kanishka Rao表示：“这种基于触觉的复杂操控能力至今仍未被完全解决。”虽然DeepMind的机器人模型运行在第三方公司的硬件平台上，但他们已经在灵巧性方面取得了进展。今年7月，DeepMind发布了Gemini Robotics 2的演示视频，该机器人能够完成系垃圾袋这样的任务——即使是人类有时也会觉得这项任务很困难。
That kind of rich, tactile-based manipulation — essentially instilling a sense of touch and feeling in a robot hand — remains “generally unsolved,” said Kanishka Rao, a research director on Google DeepMind’s robotics team. The lab, whose models are built to run on other companies’ robot hardware, has made advances in dexterity: In rolling out Gemini Robotics 2 in July, DeepMind showed a video of a humanoid tying a trash bag — a task that even humans sometimes struggle with.
关键观点：饶将未来的挑战比作在看不见包内物品的情况下试图取出钥匙：人们可以判断物品是塑料的还是金属的、柔软的还是锋利的，但“这确实是一项非常困难的任务，目前的机器人很难完成。”机器人行业的一些人士认为，并不需要具备类似人类手的功能。目前一些在医院中使用的机器人（如 Cobot 和 Diligent Robotics 公司的产品）依靠轮子移动，并配备简单的夹持装置，用于运送物资、实验室样本和推车等物品。
Rao likened future challenges to pulling keys out of a purse when you can’t see what’s inside: A person can deduce what’s plastic versus metal, soft versus sharp, but “this is a very difficult task that robots would have a lot of trouble doing today.”Some in the robotics industry argue a human-like hand isn’t needed. Some bots that are currently in use in hospitals, from companies like Cobot and Diligent Robotics, move on wheels and have simple grippers to transport supplies, lab samples, and carts through hospital corridors.
7 月份，德国研究人员发表的一项研究回顾了 100 多篇科学论文，指出“在需要精细操作的情况下，仿人手的设计并无明显优势，因为更简单的机械结构就足够满足需求”。研究人员认为，机器人制造商应该重点模仿人类手的“功能”，而不仅仅是其外观。
A study from German researchers published in July, which reviewed more than 100 scientific papers, found that “in-hand manipulation does not benefit from anthropomorphic hand design, as simpler mechanisms are sufficient.”Robot-makers should try to copy the function, rather than just the form, of the human hand, they wrote.
许多中国初创企业也在竞相攻克这一技术难题；分析人士指出，由于中国在硬件制造方面的优势，这些企业在成本和规模上具有明显优势。中国领先的机器人公司 Unitree Robotics 上周宣布开发出了一款具有 22 个自由度（即可以灵活移动的关节）的仿人机器人手。
A host of Chinese startups are also racing to solve the hand hurdle, and analysts say they have an advantage in cost and scale, given China’s hardware-manufacturing prowess. Leading Chinese firm Unitree Robotics last week announced a new human-sized robotic hand with 22 degrees of freedom.
“在美国，开发这类硬件产品几乎是不可能的，因为供应链问题限制了生产进度，”一家位于深圳的机器人手公司的创始人对《卫报》表示。
“It was really impossible to do hardware in the United States because the supply chain problem is just so constraining,” the founder of a Shenzhen-based robo-hands company told The Guardian.
随着华盛顿在 7 月份禁止进口新型外国制造的机器人，各国的物理人工智能（AI）技术发展正逐渐变得孤立化。
The superpowers’ physical AI sectors are becoming increasingly isolated, after Washington in July banned new models of foreign-made robots.