作者： 悦峰

Will KnightThe Big StoryApr 29, 2026 6:00 AMI’ve Covered Robots for Years. This One Is DifferentFrom sorting chicken nuggets to screwing in lightbulbs, Eka’s robotic claw feels like we’re approaching a ChatGPT moment for the physical world.Photograph: Tony LuongCommentLoaderSave StorySave this storyCommentLoaderSave StorySave this storyA robot’s claw hurtles toward a light bulb on a table. I wince, waiting for the crunch. But suddenly the claw decelerates. It starts gingerly pawing around the table, as if searching for its glasses on the nightstand. It gently positions the bulb between its two pincers. The bulb rolls away. The claw goes chasing it across the table. After a few nips, the bulb is back in its grasp. The robot swiftly screws the bulb into a nearby socket, illuminating its work area.In more than a decade of writing about robots, I have never seen one move so naturally. Most are ham-fisted klutzes, even when remotely controlled by a person. Of the few dozen robot arms on the market today, not one can screw in a light bulb.I have come to visit Eka, a startup located in Kendall Square, Cambridge, Massachusetts, a short walk from MIT and a slightly longer bike ride from my home. The company’s office is a few floors above one of my favorite restaurants, called Shy Bird, a place I often come to work with my own pincers-typing out stories for WIRED.Eka’s testing facilities in Cambridge, Massachusetts.Photograph: Tony LuongEka’s office is small, and it’s packed with different robot arms, assorted grippers and hands, and tables covered with odd knicknacks of different shapes, sizes, and textures-gloves, small boxes of earplugs, hairbrushes, key rings, and so on.I try putting a few things beneath the robot. First the earplugs box, then a hairbrush, and finally-in an attempt to trip it up-my own jumble of keys, which have a plush key ring. Each time, the robot swoops down and nips gently at the item a few times before grasping and lifting it up. When I try to take my keys back from Eka’s machine, the robot resists for just a moment, then lets go and instantly turns its attention back to the table, hunting for something else to pick up. Its dedication to picking is impressive. It is also kind of freaky.Watching Eka’s robot in action reminds me of the first time I tried talking to ChatGPT. The robots are so fluid, so natural-seeming, that I can’t help but feel there’s something genuinely intelligent, if not quite human, behind them.In a conference room not far from the robots, Eka’s cofounders, Pulkit Agrawal, a professor at MIT, and Tuomas Haarnoja, an ex-Google DeepMind robotics researcher, lay out their vision for the curious new machine. “A couple of years ago, we realized that dexterity can finally be cracked,” Agrawal says. Eka’s robot demos suggest that the company’s approach should enable real robot dexterity with further training. If that’s true, it could revolutionize how robots are used-not only in factories and warehouses but also in shops, restaurants, even households. “Trillions of dollars flow through the human hand,” Agrawal says. “To me, this is the biggest problem in the world to be solved.”The two men believe they are halfway there. Solving dexterity, they say, is now just a question of scaling up the approach.The fastest humans can solve a Rubik’s Cube in about three seconds. In those same three seconds, a computer with a virtual Rubik’s Cube could solve thousands of variations of the puzzle. As the Austrian computer scientist Hans Moravec famously noted in the late 1980s, the tasks that often seem hardest to us humans are child’s play for a machine; the things a child does without thinking are often a struggle for machines. Moravec suggested that the ability to interact with the physical realm evolved so long ago that for us it’s innate, more so than “higher-level” reasoning. The question has been: Can we impart that embodied intelligence to machines?One of Eka’s newer machines with a three-point hand.Photograph: Tony LuongBack in October 2018, about four years before launching ChatGPT, OpenAI created Dactyl, a robotic hand that later used AI to solve a Rubik’s Cube. The company took an off-the-shelf hand from Shadow Robot and created a detailed simulation of its joints, servos, motors, and more-a virtual hand holding a virtual cube. Using reinforcement learning, which combines experimentation with positive and negative feedback, OpenAI trained an artificial neural network to manipulate the digital cube over and over. After many thousands of repetitions of wiggling its virtual fingers, Dactyl had figured out how to move the facets of the real thing.In a press release, OpenAI suggested that Dactyl had achieved “close to human-level dexterity.” In fact, the robot lacked elements of physical intelligence that we take for granted. If the cube began to slip from its grasp, it couldn’t recover. If its hands weren’t placed at a precise angle, it couldn’t manipulate the cube at all. Even under perfect conditions, the only object it could handle was a Rubik’s Cube. And that Rubik’s Cube wasn’t even a standard one-it had sensors that tracked the movement of the squares to feed back to Dactyl.A few years later, OpenAI gave up on its robotics work to focus on large language models and chatbots. (The company has since restarted work on robotics.) Agrawal, who has remained in touch with a couple members of the Dactyl team, says the project’s simulation approach was considered a dead end because of the so-called sim-to-real gap. But both he and Haarnoja, working at separate labs, remained convinced that they could close that gap by making the sim closer to the real.At Google DeepMind, Haarnoja was on a project that used virtual reinforcement learning to train small humanoid robots to play soccer. (If this sounds more complicated than training a robotic hand to screw in a light bulb, consider that the soccer field doesn’t roll around beneath the players’ feet.) At MIT, Agrawal was researching how to train a robotic hand to grasp objects from above, not just hold them in its palm. Where Dactyl had simply moved its unfeeling pincers until the sensors in the Rubik’s cube showed its squares shifting to the desired state, Agrawal’s system would need to know what its fingers were doing and how the cube was reacting at any given moment-while accounting for the pull of gravity. When he told someone who used to work on Dactyl about the project, he says, “I got a one-hour lecture from them saying, ‘This will never work.’”Eka cofounders Pulkit Agrawal (left) and Tuomas Haarnoja at the startup’s office in Cambridge.Photograph: Tony LuongAgrawal persevered. “Pulkit is a very creative thinker,” says Ken Goldberg, a professor at UC Berkeley who has known Agrawal since his student days and is currently an adviser to his company. “He’s always pushing in a direction that other people aren’t.” (I first met him in 2017 at a big AI conference in Long Beach, California. Then a graduate student, he had just published a paper outlining a new way for computers to learn to play video games.)By late 2021, Agrawal had created a virtual hand capable of manipulating 2,000 objects upside down. Yet simulation was continuing to lose favor among roboticists, and ChatGPT fever was taking hold. If vast amounts of human-written text could yield a remarkably general linguistic intelligence, then perhaps showing robots enough examples of humans using their hands could give them physical intelligence, too.Eka uses objects of varying sizes, shapes, and weights to test its robots.Photograph: Tony LuongA handful of well-funded startups are pursuing this vision, training what are called vision-language-action (VLA) models. To build one, you show the model videos of, say, humans folding T-shirts, or humans controlling T-shirt-folding robots. The hope is that with enough data, new robotic skills will emerge. Plenty of video is already available online, but a small industry has now emerged to generate more of this data. Companies pay people to spend hours doing routine tasks with their hands while wearing cameras and motion-capture gloves.Agrawal and Haarnoja, who originally met as graduate students at UC Berkeley, teamed up to pursue a different approach with Eka. Rather than having humans provide training data, the company wants robots to learn how to do things for themselves. They spend thousands of computer hours practicing movements inside simulated worlds and inventing their own solutions. In this sense, Eka’s bot is more like AlphaZero, the Google DeepMind program that learned to play different board games with superhuman skill, and which discovered, for itself, entirely new strategies in games like chess.Eka’s founders say their robots can transfer learning from a simulator to the real world more reliably than anyone else’s-though they won’t say exactly how. Agrawal seems optimistic that their approach could lead to greater and greater capabilities. “Some people want robots to be human-level,” Agrawal says. “For us the goal is superhuman.”Engineers look on as a robot screws in a light bulb.Photograph: Tony LuongAgrawal and Haarnoja declined to give details of how they train their robots since this is their commercial edge. But they reveal that they have created custom robot grippers that incorporate a sense of touch. Agrawal and Haarnoja also say they have developed a new kind of AI algorithm called a vision-force-action model. This model learns from a simulation that incorporates not just realistic joints and motors but principles of physics like mass and inertia. It learns both how moving affects the pixels on a screen and how the weight and speed of its movement interact with the objects in its grasp.Perhaps the most interesting Eka demo involves chicken nuggets.The company’s engineers have set up a station where chicken nuggets are strewn across a table. A conveyor belt carries plastic containers along one side. Eka’s robot has to grab the nuggets and place them into the boxes. It does this with not only impressive speed but also human-like improvisation, sometimes placing nuggets carefully, but other times-if a container is moving out of reach-almost tossing them in from a short distance.An Eka robot practices placing chicken nuggets into take-out containers.Photograph: Tony LuongFood handling is an area of work that still relies heavily on humans. Fruit, vegetables, meat, and other foods need to be handled quickly but gently. It is also hard to automate because no two pieces of fruit, vegetables, or chicken nuggets look exactly the same.Eka’s demos suggest that the company may be onto something big. I found myself mentally comparing their robots to GPT-1, OpenAI’s first large language model, developed four years before ChatGPT. GPT-1 was often incoherent but showed glimmers of general linguistic intelligence.The robots I saw seem to have a similar kind of nascent physical intelligence. When I watched a video of one reaching for a set of keys in slow motion, I noticed it did something that seemed remarkably human: It touched the tips of its grippers to the table and slid them along the surface before making contact with the keys and securing them between its digits. Eka’s algorithms seem to know instinctively how to recover from a fumble. This kind of thing is difficult for other robots to learn, unless the humans training them deliberately make a wide range of mistakes.Unlike with any other robot I can think to, it’s almost possible to imagine what the world is like for the robot. Its sensors seem to feel the weight of its arm, the inertia as it sweeps toward the keys and slows down. Once it has the keys in its grasp, it seems to sense the weight of them dangling from its claw.I don’t know if Eka’s approach really is the route to a ChatGPT-like breakthrough in robotics. Some very smart experts believe that mixing human demonstration with simulation will yield better results than simulation alone. Maybe some combination of the two approaches will ultimately be necessary? But it does seem clear that robots will eventually need to have the kind of tactile, physical intelligence that Eka is working on if they are to obtain humanlike dexterity.Agrawal tells me that the same general approach should work for finer manipulation. The fiddly dexterity required to build an iPhone, for instance, could be achieved by building different actuators and sensors and practicing the task in simulation.After spending a few hours at Eka, I decide to stop by the restaurant downstairs. I watch from the counter as the staff prepare food and make coffee. A descendant of the machine upstairs may be able to do these things just as well, if not better. But given how much I enjoy chatting with the people who work there, I think I would pay extra to keep humans around. Unless, that is, my hands get automated away too.What Say You?Let us know what you think about this article in the comments below. Alternatively, you can submit a letter to the editor at [email protected]

传统AI记忆多采用检索方式，但生产环境需要更可靠的结构化记忆系统。研究提出模式感知的迭代写入路径，将记忆摄入分解为对象检测、字段检测和字段值提取。通过验证门控、本地重试和状态提示控制，提升记忆的准确性和可靠性。在提取基准测试中达到90.42%对象级准确率和62.67%输出准确率，端到端记忆基准F1分数达97.10%，应用任务准确率达95.2%。

树莓派基金会成立于2008年，首款产品2012年发布，现已成为全球第三畅销计算机。创始人Eben Upton旨在让年轻人重新体验编程乐趣，解决计算机科学招生下降问题。从最初的5入门级电脑，发展到配备神经网络加速器的AI开发平台。目前80%收入来自工业应用，包括希思罗机场数字显示、电梯控制、国际空间站等。预计5年内可在口袋中实现Sonnet级别性能，推动边缘AI计算发展。

杭州中级人民法院裁定，公司不能仅为了用成本更低的AI替代员工而合法终止劳动合同。员工周某在质检岗位工作，因AI影响被降薪60%并转岗，拒绝后遭解雇。法院认定AI节省成本不属于法定解雇理由，降薪调岗要求不合理，判定解雇违法需赔偿。此判决确立劳动权益保护先例，明确AI应用不能成为随意解雇员工的借口。

Claude Code现在拒绝处理提到’OpenClaw’的提交请求，或者对这些请求收取额外费用。这表明AI工具提供商开始关注用户使用的具体AI助手，可能涉及版权或品牌使用限制问题。

发现一个新的Linux本地提权漏洞（CVE-2026-31431），仅732字节大小的漏洞利用程序。无需竞态条件，无各发行版偏移差异，通过页面缓存写入绕过磁盘完整性工具，可跨越容器边界，影响Ubuntu、Amazon Linux、RHEL、SUSE等系统。

这是一个用shell脚本构建的完整编程代理框架，仅400行代码。使用7个工具（bash, read, write, edit, grep, find, ls），支持Anthropic和OpenAI API，具备REPL、自动压缩、检查点恢复等功能。强调便携性，不依赖外部库，只使用系统原语。

## AI驱动的重大生物学发现\n\n### 研究突破\n- **研究机构**: Gladstone研究所\n- **研究方法**: 人工智能分析技术\n- **发现意义**: 挑战传统分子生物学认知

## Claude Opus 4.7的身份识别能力\n\n### 技术突破\n- **AI模型**: Claude Opus 4.7能够从文本片段中识别作者身份\n- **识别精度**: 即使是不相关的文本类型也能准确识别\n- **应用范围**: 政治文章、教育报告、电影评论、小说创作等

## 供应链攻击影响AI开发环境\n\n### 主要发现\n- **受影响项目**: PyPI包’lightning’（深度学习框架）版本2.6.2和2.6.3\n- **发布时间**: 2026年4月30日\n- **影响范围**: 图像分类、LLM微调、扩散模型、时间序列预测等项目