内容来源：https://techcrunch.com/2026/01/30/physical-intelligence-stripe-veteran-lachy-grooms-latest-bet-is-building-silicon-valleys-buzziest-robot-brains/

内容总结：

旧金山探访Physical Intelligence：一家拒绝商业化诱惑的机器人“大脑”公司

在旧金山街头，一扇门上颜色稍浅的π符号，是唯一能指引你找到Physical Intelligence总部的标识。推门而入，没有前台，没有炫目的发光Logo，眼前是一个巨大的混凝土空间，被随意摆放的浅色长桌稍减冷峻。桌上散落着饼干盒、Vegemite酱料瓶，以及更多显示器、机器人零部件和缠绕的黑线——这里正在进行一场关于机器人通用智能的“头脑”革命。

机器人版的“ChatGPT”

公司联合创始人、加州大学伯克利分校副教授谢尔盖·莱文将眼前的景象比作“机器人版的ChatGPT”。在这个测试场里，多只机械臂正笨拙地执行任务：一条手臂努力折叠黑色裤子，另一条试图将衬衫内外翻转，第三条则熟练地为西葫芦削皮。这些看似简单的日常动作，正是公司核心技术的试验场。

莱文解释道，他们正在构建一个持续循环：从遍布各处的机器人工作站（包括仓库、家庭等）收集数据，用以训练通用的机器人基础模型。新模型训练完成后，再回到这样的工作站进行评估。削皮机械臂或许正在测试模型能否将“剥皮”这一基本动作泛化到从未见过的苹果或土豆上。公司甚至配备了昂贵的意式咖啡机，目的并非犒劳工程师，而是为机器人提供学习制作拿铁的数据。

“硬件平庸，智能补足”

公司的硬件策略 deliberately “去光环化”。他们使用的机械臂单价约3500美元，莱文坦言其中包含供应商的“巨额溢价”。若自行制造，材料成本可降至1000美元以下。“关键在于，”莱文强调，“优秀的智能可以弥补平庸的硬件。”几年前，机器人专家可能难以相信这些廉价设备能完成任何任务，但如今这正成为现实。

“不承诺商业化”的罕见融资

公司另一核心人物是31岁的拉奇·格鲁姆。这位年少成名的澳大利亚创业者（办公室里的Vegemite酱正是他的口味）在离开Stripe后，一直在寻找值得全身心投入的事业。他被莱文和斯坦福大学教授切尔西·芬恩的学术成果吸引，最终联合创办了Physical Intelligence。

成立仅两年，公司已融资超10亿美元，估值达56亿美元。但格鲁姆向包括Khosla、红杉资本在内的投资者做出了不寻常的承诺：他不提供明确的时间表。 “我不给投资者关于商业化的答案，”他说，“这有点奇怪，但人们容忍了这一点。”他清楚这种容忍并非无限，因此公司选择在当下充分融资。

战略：跨平台学习与数据多样性

如果不急于商业化，战略是什么？另一位来自Google DeepMind的联合创始人关·翁解释，核心在于“跨实体学习”和多样数据源。他们的目标是，无论未来出现何种新的硬件平台，都能将模型已掌握的知识快速迁移，大幅降低为新平台赋予自主能力的边际成本。公司已与物流、杂货乃至街对面的巧克力制造商等少量合作伙伴测试系统，关·翁称在某些领域已达到实用自动化水平。

路线之争：纯研究 vs. 商业飞轮

这条赛道并非只有Physical Intelligence。成立于2023年的匹兹堡公司Skild AI本月刚以140亿美元估值融资14亿美元，并已将其“全能体”Skild Brain投入商用，去年在安保、仓储和制造领域创收3000万美元。两家公司代表了鲜明的路线分歧：Skild AI相信商业部署能形成数据飞轮，反哺模型；而Physical Intelligence则赌注于抵制短期商业化诱惑，以产出更优越的通用智能。谁“更正确”可能需要数年才能见分晓。

格鲁姆将公司描述为一种“纯粹的”运营模式：“研究人员有需求，我们就去收集数据或获取硬件来支持，一切由内驱动。”公司原有一个5到10年的技术路线图，但在第18个月就被团队超越了。

挑战与信念

公司目前约有80名员工，格鲁姆希望“尽可能慢地”增长。最大挑战来自硬件本身：易损坏、交付慢、安全考量复杂。“硬件就是非常困难。我们所做的一切都比软件公司难得多。”

尽管外界对机器人进入家庭厨房的必要性、安全性以及巨额投入是否解决真问题存疑，但格鲁姆毫无动摇。他与数十年深耕此领域的专家共事，并相信时机已然成熟。硅谷也一如既往地给予像他这样的人充分的信任与空间——即使没有清晰的商业化路径、没有时间表、甚至不确定最终的市场形态。历史表明，并非所有赌注都会成功，但一旦成功，便足以证明此前所有尝试的价值。

此刻，访客离开时，机械臂仍在练习：裤子仍未叠好，衬衫依然正面朝外，只有西葫芦皮在持续堆积。一场关于机器人通用智能的漫长实验，仍在继续。

中文翻译：

从街上看去，Physical Intelligence公司在旧金山的总部只有一个不起眼的标识——门上那个颜色略异于门板的π符号。走进室内，眼前立刻被忙碌景象填满。这里没有接待台，没有荧光灯下闪耀的企业标识。

内部空间是个巨大的混凝土盒子，随意摆放的浅色原木长桌稍稍缓和了空间的冷峻感。部分桌子显然是午餐区，散落着女童子军饼干盒、维吉麦酱罐（公司里肯定有澳大利亚人），以及塞满各种调味料的小铁丝篮。而其他桌子则呈现完全不同的场景——大多数桌面上堆满显示器、备用机器人零件、纠缠的黑线，以及处于不同训练阶段的完整机械臂，它们正尝试掌握各种平凡技能。

参观期间，一只机械臂正在折叠黑色裤子，或者说试图折叠——进展并不顺利。另一只机械臂正以"不达目的不罢休"的架势试图将衬衫里外翻转，虽然今天未必能成功。第三只机械臂似乎找到了天赋所在，正快速削着西葫芦皮，随后要将削下的皮屑投入指定容器——至少削皮环节相当出色。

"可以把它想象成机器人的ChatGPT。"谢尔盖·莱文指着满屋正在运转的机械臂对我说。这位加州大学伯克利分校副教授兼Physical Intelligence联合创始人戴着眼镜，神态亲和，显然常年习惯于向非专业人士解释复杂概念。

他解释道，眼前所见是持续循环的测试阶段：数据从这里的机器人工作站及其他场所（仓库、住宅等任何团队能设立站点的地方）收集，用于训练通用机器人基础模型。研究人员训练出新模型后，会送回此类工作站评估。叠裤子和翻衬衫都是实验项目，西葫芦削皮测试则可能意在验证模型能否将技能迁移到不同蔬菜，掌握削皮的基本动作以处理从未见过的苹果或土豆。

公司在这栋建筑及其他地点设有测试厨房，使用现成硬件让机器人接触不同环境与挑战。旁边有台精密意式咖啡机，我原以为是员工福利，直到莱文澄清那是给机器人学习的工具。所有打出的奶泡拿铁都是数据样本，而非现场数十名工程师的福利——他们大多正盯着电脑屏幕或俯身观察机械实验。

硬件本身刻意保持朴素。这些机械臂售价约3500美元，莱文称供应商加了"巨额溢价"。若自主生产，材料成本可降至千元以下。他说若在几年前，机器人专家根本不敢相信这些设备能完成任何任务——而这正是关键所在：卓越的智能可以弥补硬件的不足。

当莱文暂时离开时，拉奇·格鲁姆朝我走来。他步履匆匆，仿佛同时处理着五六件事。31岁的格鲁姆仍保持着硅谷神童特有的青春气质——这个称号他当之无愧：13岁在澳大利亚创立首家公司，九个月后便成功出售（这也解释了公司里的维吉麦酱）。

早前当我首次提出采访请求时，他正接待一群穿连帽衫的访客，当即回绝："绝对不行，我要开会。"现在他或许能挤出十分钟。

格鲁姆在关注莱文实验室及其学生切尔西·芬恩（现任斯坦福机器人学习实验室负责人）的学术成果时，找到了追寻的方向。机器人领域所有有趣进展中不断出现他们的名字。听闻他们可能创业的传闻后，他联系了同样在斯坦福任教、参与该项目的谷歌DeepMind研究员卡罗尔·豪斯曼。"那是一场让你走出会议室时顿悟'就是它了'的会面。"

格鲁姆告诉我，尽管有人疑惑他为何不专事投资，但他从未打算成为全职投资人。作为Stripe早期员工离职后，他做了五年天使投资人，早期投资了Figma、Notion、Ramp、Lattice等公司，同时寻找适合自己创立或加入的企业。2021年对Standard Bots的首个机器人投资，让他重拾童年搭建乐高机器人时的热爱。他戏称"当投资人时度假更多"，但投资只是保持活跃、结识伙伴的方式，并非终极目标。"离开Stripe后我花了五年寻找创业方向，"他说，"天时地利人和的创意极其罕见。执行力固然重要，但糟糕的创意即便拼命执行仍是糟糕的创意。"

这家成立两年的公司已融资超10亿美元。当我问及资金消耗时，他立即澄清实际开支不大，主要成本集中在算力上。随即他又承认，若有合适条款与合作伙伴，公司会继续融资。"投入资金的额度没有上限，"他说，"总有更多算力可以用于解决问题。"

这种安排的特别之处在于格鲁姆未向投资者承诺盈利时间表。"我不向投资者回答商业化问题，"谈及包括Khosla Ventures、红杉资本、Thrive Capital等将公司估值推至56亿美元的投资方时他说，"人们能容忍这点其实挺奇怪。"但容忍确实存在，且未必持久，因此公司需要当下充分融资。

若不追求商业化，战略是什么？来自谷歌DeepMind的另一位联合创始人关·翁解释，战略核心在于跨实体学习与多元数据源。即使未来出现新硬件平台，也无需从头收集数据——所有已有模型知识均可迁移。"为新机器人平台加载自主功能的边际成本，无论何种平台，都会大幅降低。"他说。

公司已与物流、杂货、对街巧克力制造商等不同垂直领域少数企业合作，测试系统能否满足实际自动化需求。关·翁称某些场景已达标。通过"任意平台、任意任务"的模式，成功覆盖面已足够广泛，可以开始逐项攻克当前适合自动化的任务。

追逐这一愿景的不止Physical Intelligence。通用机器人智能的竞赛正在升温——正如三年前震撼世界的LLM模型，这将成为专业化应用的基础。2023年成立于匹兹堡的Skild AI本月刚以140亿美元估值融资14亿美元，其路径截然不同：Physical Intelligence专注纯研究，Skild AI则已商业部署"全能体"Skild Brain系统，去年短短数月内在安防、仓储、制造领域创收3000万美元。

Skild甚至公开抨击竞争对手，在其博客中指出多数"机器人基础模型"只是"伪装"的视觉语言模型，因过度依赖互联网规模预训练而非物理模拟与真实机器人数据，缺乏"真正的物理常识"。

这是尖锐的理念分歧。Skild AI赌注商业部署能形成数据飞轮，通过每个实际用例改进模型；Physical Intelligence则赌抵制短期商业化能催生更卓越的通用智能。孰是孰非需数年验证。

在此期间，Physical Intelligence以格鲁姆所说的"异常清晰"模式运作。"这是家非常纯粹的公司。研究人员有需求，我们就去收集数据支持——或是添置新硬件等——然后执行。不受外部驱动。"团队曾制定5-10年发展路线图，但他说第18个月时就已超越所有预期。

公司现有约80名员工并计划扩张，尽管格鲁姆希望"尽可能慢速增长"。他表示最大挑战在于硬件："硬件确实艰难。我们所有工作都比软件公司困难得多。"硬件会损坏、交付迟缓拖累测试、安全考量使一切复杂化。

当格鲁姆匆匆赶赴下一个行程时，我继续观察机器人训练。裤子仍未叠好，衬衫顽固地保持原样，西葫芦皮屑已堆积成小山。

存在许多显而易见的问题——包括我自己的疑问：是否真有人需要厨房削菜机器人？安全性如何？家中宠物会对机械入侵者发狂吗？投入的所有时间金钱究竟在解决足够重大的问题，还是在制造新问题？外界也质疑公司进展、愿景可行性，以及押注通用智能而非具体应用是否明智。

即便格鲁姆心存疑虑，也未曾显露。他与深耕该领域数十载的伙伴共事，这些人相信时机终于成熟——对他而言，知道这点便已足够。

况且硅谷自产业诞生之初就始终支持格鲁姆这类人，给予充分试错空间。资本深知，即便没有清晰商业化路径、没有时间表、无法确定抵达时的市场形态，他们终会找到出路。并非每次都能成功。但当成功降临时，往往足以弥补过往所有失败。

英文来源：

From the street, the only indication I’ve found Physical Intelligence’s headquarters in San Francisco is a pi symbol that’s a slightly different color than the rest of the door. When I walk in, I’m immediately confronted with activity. There’s no reception desk, no gleaming logo in fluorescent lights.
Inside, the space is a giant concrete box made slightly less austere by a haphazard sprawl of long blonde-wood tables. Some are clearly meant for lunch, dotted with Girl Scout cookie boxes, jars of Vegemite (someone here is Australian), and small wire baskets stuffed with one too many condiments. The rest of the tables tell a different story entirely. Many more of them are laden with monitors, spare robotics parts, tangles of black wire, and fully assembled robotic arms in various states of attempting to master the mundane.
During my visit, one arm is folding a pair of black pants, or trying to. It’s not going well. Another is attempting to turn a shirt inside out with the kind of determination that suggests it will eventually succeed, just not today. A third — this one seems to have found its calling — is quickly peeling a zucchini, after which it is supposed to deposit the shavings into a separate container. The shavings are going well, at least.
“Think of it like ChatGPT, but for robots,” Sergey Levine tells me, gesturing toward the motorized ballet unfolding across the room. Levine, an associate professor at UC Berkeley and one of Physical Intelligence’s co-founders, has the amiable, bespectacled demeanor of someone who has spent considerable time explaining complex concepts to people who don’t immediately grasp them.
What I’m watching, he explains, is the testing phase of a continuous loop: data gets collected on robot stations here and at other locations — warehouses, homes, wherever the team can set up shop — and that data trains general-purpose robotic foundation models. When researchers train a new model, it comes back to stations like these for evaluation. The pants-folder is someone’s experiment. So is the shirt-turner. The zucchini-peeler might be testing whether the model can generalize across different vegetables, learning the fundamental motions of peeling well enough to handle an apple or a potato it’s never encountered.
The company also operates a test kitchen in this building and elsewhere using off-the-shelf hardware to expose the robots to different environments and challenges. There’s a sophisticated espresso machine nearby, and I assume it’s for the staff until Levine clarifies that no, it’s there for the robots to learn. Any foamed lattes are data, not a perk for the dozens of engineers on the scene who are mostly peering into their computers or hovering over their mechanized experiments.
The hardware itself is deliberately unglamorous. These arms sell for about $3,500, and that’s with what Levine describes as “an enormous markup” from the vendor. If they manufactured them in-house, the material cost would drop below $1,000. A few years ago, he says, a roboticist would have been shocked these things could do anything at all. But that’s the point — good intelligence compensates for bad hardware.
TechCrunch Founder Summit 2026: Tickets Live
On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately
Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more.
TechCrunch Founder Summit: Tickets Live
On June 23 in Boston, more than 1,100 founders come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately
Save up to $300 on your pass or save up to 30% with group tickets for teams of four or more.
As Levine excuses himself, I’m approached by Lachy Groom, moving through the space with the purposefulness of someone who has half a dozen things happening at once. At 31, Groom still has the fresh-faced quality of Silicon Valley’s boy wonder, a designation he earned early, having sold his first company nine months after starting it at age 13 in his native Australia (this explains the Vegemite).
When I first approached him earlier, as he welcomed a small gaggle of sweatshirt-wearing visitors into the building, his response to my request for time with him was immediate: “Absolutely not, I’ve got meetings.” Now he has 10 minutes, maybe.
Groom found what he was looking for when he started following the academic work coming out of the labs of Levine and Chelsea Finn, a former Berkeley PhD student of Levine’s who now runs her own lab at Stanford focused on robotic learning. Their names kept appearing in everything interesting happening in robotics. When he heard rumors they might be starting something, he tracked down Karol Hausman, a Google DeepMind researcher who also taught at Stanford and who Groom had learned was involved. “It was just one of those meetings where you walk out and it’s like, This is it.”
Groom never intended to become a full-time investor, he tells me, even though some might wonder why not given his track record. After leaving Stripe, where he was an early employee, he spent roughly five years as an angel investor, making early bets on companies like Figma, Notion, Ramp, and Lattice while searching for the right company to start or join himself. His first robotics investment, Standard Bots, came in 2021 and reintroduced him to a field he’d loved as a kid building Lego Mindstorms. As he jokes, he was “on vacation much more as an investor.” But investing was just a way to stay active and meet people, not the endgame. “I was looking for five years for the company to go start post-Stripe,” he says. “Good ideas at a good time with a good team — [that’s] extremely rare. It’s all execution, but you can execute like hell on a bad idea, and it’s still a bad idea.”
The two-year-old company has now raised over $1 billion, and when I ask about its runway, he’s quick to clarify it doesn’t actually burn that much. Most of its spending goes toward compute. A moment later, he acknowledges that under the right terms, with the right partners, he’d raise more. “There’s no limit to how much money we can really put to work,” he says. “There’s always more compute you can throw at the problem.”
What makes this arrangement particularly unusual is what Groom doesn’t give his backers: a timeline for turning Physical Intelligence into a money-making endeavor. “I don’t give investors answers on commercialization,” he says of backers that include Khosla Ventures, Sequoia Capital, and Thrive Capital among others that have valued the company at $5.6 billion. “That’s sort of a weird thing, that people tolerate that.” But tolerate it they do, and they may not always, which is why it behooves the company to be well-capitalized now.
So what’s the strategy, if not commercialization? Quan Vuong, another co-founder who came from Google DeepMind, explains that it revolves around cross-embodiment learning and diverse data sources. If someone builds a new hardware platform tomorrow, they won’t need to start data collection from scratch — they can transfer all the knowledge the model already has. “The marginal cost of onboarding autonomy to a new robot platform, whatever that platform might be, it’s just a lot lower,” he says.
The company is already working with a small number of companies in different verticals — logistics, grocery, a chocolate maker across the street — to test whether their systems are good enough for real-world automation. Vuong claims that in some cases, they already are. With their “any platform, any task” approach, the surface area for success is large enough to start checking off tasks that are ready for automation today.
Physical Intelligence isn’t alone in chasing this vision. The race to build general-purpose robotic intelligence — the foundation on which more specialized applications can be built, much like the LLM models that captivated the world three years ago — is heating up. Pittsburgh-based Skild AI, founded in 2023, just this month raised $1.4 billion at a $14 billion valuation and is taking a notably different approach. While Physical Intelligence remains focused on pure research, Skild AI has already deployed its “omni-bodied” Skild Brain commercially, saying it generated $30 million in revenue in just a few months last year across security, warehouses, and manufacturing.
Skild has even taken public shots at competitors, arguing on its blog that most “robotics foundation models” are just vision-language models “in disguise” that lack “true physical common sense” because they rely too heavily on internet-scale pretraining rather than physics-based simulation and real robotics data.
It’s a pretty sharp philosophical divide. Skild AI is betting that commercial deployment creates a data flywheel that improves the model with each real-world use case. Physical Intelligence is betting that resisting the pull of near-term commercialization will enable it to produce superior general intelligence. Who’s “more right” will take years to resolve.
In the meantime, Physical Intelligence operates with what Groom describes as unusual clarity. “It’s such a pure company. A researcher has a need, we go and collect data to support that need — or new hardware or whatever it is — and then we do it. It’s not externally driven.” The company had a 5- to 10-year roadmap of what the team thought would be possible. By month 18, they’d blown through it, he says.
The company has about 80 employees and plans to grow, though Groom says hopefully “as slowly as possible.” What’s the most challenging, he says, is hardware. “Hardware is just really hard. Everything we do is so much harder than a software company.” Hardware breaks. It arrives slowly, delaying tests. Safety considerations complicate everything.
As Groom springs up to rush to his next commitment, I’m left watching the robots continue their practice. The pants are still not quite folded. The shirt remains stubbornly right-side-out. The zucchini shavings are piling up nicely.
There are obvious questions, including my own, about whether anyone actually wants a robot in their kitchen peeling vegetables, about safety, about dogs going crazy at mechanical intruders in their homes, about whether all of the time and money being invested here solves big enough problems or creates new ones. Meanwhile, outsiders question the company’s progress, whether its vision is achievable, and if betting on general intelligence rather than specific applications makes sense.
If Groom has any doubts, he doesn’t show it. He’s working with people who’ve been working on this problem for decades and who believe the timing is finally right, which is all he needs to know.
Besides, Silicon Valley has been backing people like Groom and giving them a lot of rope since the beginning of the industry, knowing there’s a good chance that even without a clear path to commercialization, even without a timeline, even without certainty about what the market will look like when they get there, they’ll figure it out. It doesn’t always work out. But when it does, it tends to justify a lot of the times it didn’t.