一条广为转发的帖子把它读成了讣告。帖子写道 “OpenAI has officially disbanded its robotics team, just five months after relaunching it”,接着是 “it’s over”,浏览量过了一百万。OpenAI 员工当天下午就出来反驳。工程师 David Schnurr 写道 “robotics at openai is alive and well.",并说有 15 个职位在招。简介为 “sora, robot researcher @ openai, cs @ stanford” 的 Welton Wang 回复:“this is completely wrong (read the article), we are investing more than ever in robotics.”
5 月 31 日,Sam Altman 写道:“Our world simulation research program, led by Aditya Ramesh (@model_mechanic), has evolved over the past year into OpenAI Robotics. Progress is rapid, and based on a foundation of co-design between robotics hardware and ML research.” 也就是世界模拟研究项目在过去一年演变成了 OpenAI Robotics,进展建立在机器人硬件和机器学习研究的协同设计之上。五个半月后,按 Fortune 的说法,机器学习研究向一位高管汇报,硬件向另一位高管汇报。
这是一个说得通的押注,不是撤退。如果通用模型真的能驱动机器人,机器人就成了这个模型的下游应用,和写代码、做法律工作一样。卡内基梅隆大学的 Guanya Shi 告诉 Fortune,他把这次调整理解为 OpenAI 认为机器人 “no longer a special problem”(不再是一个特殊问题)。在这个前提下,单独做机器人大脑的团队大多是在重复造通用模型,还要跟它抢算力。
速度和联网。 Shi 还对 Fortune 说,大语言模型能力强但慢,跑在云上,要求机器人一直联网。他认为仍需要能跑在机器人本体计算机上的小型专用模型。一个模型会做这件事,但每次都要连回云端再等结果,还算不上控制器。
现场数据在已经部署了机器人的人手里。 MIT 教授、Walden Robotics 联合创始人兼 CEO Russ Tedrake 对 Fortune 说,机器人公司有一个实验室没有的优势:它们的机器人已经在真实场所积累工作经验;基础模型越好,这些特定现场的数据越值钱,因为机器人能更快地从中学会。OpenAI 自己的机器人还没公开亮相;据 Fortune 报道,Altman 的目标是 2027 年首次公开演示自家人形机器人。
物理工程依然难。 这一点最好的证据是 OpenAI 自己 11 月 3 日的文章。文里的 Astra 做的是一台用来跟 Codex 对话的设备,不是机器人。两名成员把它当作副业,用 Astra 在一周内设计出来,代工厂不到一个月做出了一小批样机。OpenAI 写道:“Getting the device ready for manufacturing still required several rounds of review and iteration. Astra did not consistently account for the manufacturer’s production limits, tolerances, or details required for assembly. These gaps are helping us identify what future models need to do better.” 也就是可量产仍要好几轮评审和迭代,因为 Astra 没有始终考虑到厂商的生产限制、公差和装配细节;这些缺口在帮他们看清未来的模型还要在哪些地方做得更好。我的判断:据 Fortune,硬件团队现在归 Applied;这篇文章没说这台设备是哪个部门做的;机器人硬件也有类似的制造细节,所以我把它看成协同设计问题。
值得认真对待的反驳。“团队和汇报线不重要,同一批人还在做同样的事。“有时确实如此,CoRL 的场景也支持这点:Fortune 看到前团队成员站在空展位旁,照样有人围观。但算力紧张时,汇报线决定谁的路线图优先,而这正是 Fortune 描述的处境。“机器人只是模型的又一个应用。“有可能,但做通用机器人模型的创业公司 Generalist 的联合创始人 Pete Florence 对 Fortune 说:“There may be different starting points for how to grow intelligence.” 而且现在还看不出哪个已有模型会主导机器人领域。
Exclusive: OpenAI disbands its stand-alone robotics team as its general models start running robots (Fortune) — 13 Nov 2026, updated same day. OpenAI confirmed the reorganization to Fortune: robotics research team moved to the main research team under Chief Research Officer Mark Chen; applied robotics and hardware teams moved to the Applied organization under President Greg Brockman. According to Fortune's paraphrase, a spokesperson said that robotics remains a core part of OpenAI's long-term vision. At the start of the year researchers assumed building on Sora-style world models was the best path; since then general frontier models including GPT-6 Astra have shown they can be applied directly to robotics. OpenAI booth #208 at CoRL in Austin sat empty. Robotics team wound down in 2020, rebuilt in 2024 with a Figure partnership. Fortune's sentence: "That partnership ended in February 2025, when Figure said it would develop its own robot AI." That "would develop" clause is Fortune's wording, not Adcock's post. Altman targeting a first public demonstration of OpenAI's own humanoid in 2027. Cost context Fortune presents as background, not as OpenAI's stated reason: Altman's earlier call to abandon side quests, Sora dropped, operating losses reportedly $10.9 billion last quarter, possible IPO in early 2027. Guanya Shi (CMU) told Fortune he would interpret the move as a signal that OpenAI sees robotics as "no longer a special problem". LLMs slow, cloud-hosted, need a constant connection; still need smaller onboard models. Russ Tedrake (MIT, Walden Robotics): deployed robots gain work experience and site data, which better foundation models make more valuable. Pete Florence (Generalist): "There may be different starting points for how to grow intelligence." No headcount given. Fortune does not say a partner walks away with the data loop.
Sam Altman on X, "OpenAI Robotics is hiring" (31 May 2026) — "Our world simulation research program, led by Aditya Ramesh (@model_mechanic), has evolved over the past year into OpenAI Robotics. Progress is rapid, and based on a foundation of co-design between robotics hardware and ML research." "In the short term, we are focused on robots to support skilled workers to build our future infrastructure; in the long term, we imagine everyone having a personal robot doing anything they need."
David Schnurr on X, "robotics at openai is alive and well" (13 Nov 2026) — OpenAI engineer. "robotics at openai is alive and well." "15 open roles – come work with us!" The public list behind that link was checked on 17 Nov 2026; see the Ashby job-list source. Exactly 15 Robotics-team postings, and none is an ML research role.
Welton Wang on X, "this is completely wrong (read the article)" (13 Nov 2026) — Profile: "sora, robot researcher @ openai, cs @ stanford". Quote-posting the post that ends "it’s over": "this is completely wrong (read the article), we are investing more than ever in robotics." Also, keeping his wording: "one thing that makes OpenAI special and enduring is the ability to consistent make strategic changes (and organizational ones that reflect this) in a world where model capabilities continually surprise everyone."
Accelerating engineering for the physical world (OpenAI) — 3 Nov 2026. Not a robot, and the post does not say which group built it, and does not use the word handheld. "To do this, we used GPT-6 to define the functional requirements and design the enclosure, printed circuit board (PCB), and software integrations needed to create a device users can use to talk to Codex." "Two team members designed this test device with Astra in under a week while working on it as a side project. A contract manufacturer built a small batch of prototypes in under a month." "Getting the device ready for manufacturing still required several rounds of review and iteration. Astra did not consistently account for the manufacturer's production limits, tolerances, or details required for assembly. These gaps are helping us identify what future models need to do better." Hardware and infrastructure teams use Codex and ChatGPT for CAD, manufacturing drawings, simulations and design reviews, spanning robotics and compute infrastructure; hardware-team token usage up 23-fold since February.
Brett Adcock on X, leaving the Collaboration Agreement with OpenAI (4 Feb 2025) — "Today, I made the decision to leave our Collaboration Agreement with OpenAI." "Figure made a major breakthrough on fully end-to-end robot AI, built entirely in-house." He does not say Figure would build that system. The post also does not say a partner takes the data loop with it.
OpenAI public job list (Ashby) — Checked 17 Nov 2026 via https://api.ashbyhq.com/posting-api/job-board/openai (839 jobs). Exactly 15 have team and department Robotics, all listed, all San Francisco, full time. No ML research posting. Eleven are mechanical, electrical, firmware, manufacturing, or test: Compute Electrical Engineer, Robotics; Electrical Engineer, Robotics; Engineering Technician, Actuators; Manufacturing Engineer, Motors & Actuators; Mechanical Design Engineer, Motor Design & Prototyping; PCB Layout Engineer, Robotics; Firmware Engineer, Robotics; Electrical Engineer, PCBA Test Infrastructure, Robotics; Mechanical Engineer, Dynamometer and Actuator Testing Infrastructure; Mechanical Engineer, PCBA Test Infrastructure, Robotics; Robotics Test Engineer. The other four: Control Systems Software Engineer, Robotics (real-time control; the posting says the hire will work with machine learning engineers to implement reinforcement learning models); Robotics Software Engineer (data collection and evaluation); Technical Program Manager, Actuators; Lead Safety Engineer, Robotics.