2026年10月18日，OpenAI创始人Sam Altman官宣，公司计划于2027年展示首个人形机器人样机。这是OpenAI首次给出人形机器人的公开时间表，也再次让具身智能圈绷紧神经。OpenAI曾在2017年展示机械手解魔方，却在2020年停止了机器人研究，又在2024年重启机器人研究的招聘。过去每次OpenAI说要做机器人，业内都会说一次“狼来了”。但这一次，情况有所不同。2026年9月，OpenAI推出GPT-6 Astra，能够以零样本控制机械臂抓放物体，展现出比过往模型强得多的物理空间理解能力。早前有消息称，OpenAI正通过自建具身数据工厂和外采数据两种方式，累计上千万小时的具身操作数据用于AI训练。
On October 18, 2026, OpenAI founder Sam Altman officially announced the company plans to showcase its first humanoid robot prototype in 2027. This is OpenAI's first public timetable for humanoid robots, once again tightening nerves in the embodied intelligence circle. OpenAI demonstrated a robotic hand solving a Rubik's cube in 2017 but stopped robotics research in 2020, then restarted robotics hiring in 2024. Every time OpenAI said it would do robots in the past, the industry said "the wolf is coming." But this time is different. In September 2026, OpenAI launched GPT-6 Astra, capable of zero-shot controlling robotic arms to grasp and place objects, showing physical spatial understanding far stronger than past models. Earlier reports said OpenAI is accumulating over ten million hours of embodied manipulation data for AI training through both self-built embodied data factories and external data procurement.
Astra公布后，极佳视界创始人朱政在2026年外滩大会上直言，这次狼真的来了，放弃幻想。朱政口中的狼，不单指OpenAI一家，而是以OpenAI为代表的大语言模型公司，既有国外Anthropic、Google，也有国内智谱、DeepSeek、字节等。它们极有可能在未来几个月公布在具身智能领域的新进展，并凭借更多的算力、更多的资金和更高的人才密度，与独立具身大脑公司开展竞争。
After Astra's release, Jiashijie founder Zhu Zheng said bluntly at the 2026 Bund Conference: "This time the wolf is really here—abandon illusions." The wolf in Zhu Zheng's words is not just OpenAI but large language model companies represented by OpenAI—both international ones including Anthropic and Google, and domestic ones including Zhipu, DeepSeek, and ByteDance. They are highly likely to announce new progress in embodied intelligence in the coming months and compete with independent embodied brain companies with more compute, more capital, and higher talent density.
机器人要完成动作，具身大脑负责感知、理解、任务规划和决策推理，相当于人的认知系统；具身小脑则负责更快的运动控制，让关节和夹爪稳定执行。而真正让机器人把决策转化为动作的，是专门训练的具身基座模型。目前具身模型技术路线尚未收敛，但呈现出一种相对明显的演进方向。VLA率先把视觉、语言和动作连在一起，让机器人能够看见指令就动手，但VLA模型的短板在于泛化性不足，导致机器人换个任务场景就容易失灵。世界模型补上的，正是这种动手前先想一步的能力，它根据当前状态和候选动作，预测环境与机器人自身接下来可能发生的变化，帮助系统比较方案、规划步骤和纠正偏差。
For robots to complete actions, the embodied brain handles perception, understanding, task planning, and decision reasoning—equivalent to a human's cognitive system; the embodied cerebellum handles faster motion control, letting joints and grippers execute stably. What truly converts decisions into actions is the specially trained embodied foundation model. Currently embodied model technical routes have not converged but show a relatively clear evolution direction. VLA first connected vision, language, and action, letting robots "see an instruction and act," but VLA models' shortcoming is insufficient generalization, causing robots to fail easily when switching task scenarios. World models fill in exactly this "think a step ahead before acting" capability—predicting likely changes in the environment and the robot itself based on current state and candidate actions, helping the system compare options, plan steps, and correct deviations.
当以大语言模型为基座的通用模型开始切入具身大脑的领域，留给创业公司的空间还有多少。悲观预判开始流行。一家具身仿真数据供应商的联合创始人王怀认为，大模型正在吃掉独立的具身大脑。从Astra在机械臂的部署效果看，他认为其已经具备空间理解能力，也具备了一定的物理理解能力。前者指物体在物理空间的位置，后者指理解施加作用力后产生的效果。而目前较薄弱的物理理解能力，有望在后续技术迭代下快速弥补，甚至实现超越。宇树科技创始人王兴兴近期也表达了类似看法。他把Astra部署到多种机器人上做各种任务测试，认为效果实际上非常不错。王兴兴表示，我感觉整个临界点的到来还是会相对比较快，快的话甚至一两年，因为明年GPT-7出来，可能直接把具身智能问题打穿。
When general models based on large language models begin entering the embodied brain field, how much space remains for startups. Pessimistic predictions are becoming popular. Wang Huai, co-founder of an embodied simulation data supplier, believes large models are eating independent embodied brains. From Astra's deployment results on robotic arms, he believes it already possesses spatial understanding and certain physical understanding. The former refers to object positions in physical space; the latter to understanding effects after applying force. The currently weaker physical understanding is expected to be rapidly弥补 through subsequent technology iteration, even surpassed. Unitree founder Wang Xingxing recently expressed a similar view. He deployed Astra to various robots for various task tests, finding results "actually very good." Wang Xingxing said: "I feel the critical point's arrival will still be relatively fast—possibly even one or two years, because next year GPT-7 comes out and may directly break through the embodied intelligence problem."
不过，大模型会吃掉具身大脑这一判断，建立在大模型持续高速迭代的设想上；落地到真实机器人，仍存在现实硬约束。最直接的证据，来自银河通用在Astra发布不久后，于GitHub上发布的一份30页技术报告。这份名为《GPT-6 Astra作为具身策略》的报告，在机器人仿真测试平台RoboDojo上，将Astra与开源具身基模π0.5放进同一套仿真双臂测试台，进行十类复杂操作任务共50次。结果显示，Astra单干成功率只有26%，接入π0.5后成功率升至48%；而Astra全程只修正了全部执行控制步的14.4%，剩下85.6%的动作仍由π0.5直接给出。这一实验结果一定程度上说明了一种暂时有效的分工，Astra充当具身大脑，用于判断目标、检查任务状态等慢思考功能；π0.5充当听从Astra指挥的具体执行系统，用于将指令、状态等转化成机器人本体、接触、运动的物理动作。
However, the judgment that large models will eat embodied brains is built on the assumption of continued rapid large model iteration; landing on real robots still faces hard constraints. The most direct evidence comes from a 30-page technical report Galbot published on GitHub shortly after Astra's release. Titled "GPT-6 Astra as Embodied Policy," the report placed Astra and open-source embodied foundation model π0.5 on the same simulated dual-arm test bench on robot simulation platform RoboDojo, conducting ten complex manipulation task types totaling 50 trials. Results showed Astra alone achieved only 26% success rate, rising to 48% after connecting to π0.5; Astra only corrected 14.4% of all execution control steps throughout, with the remaining 85.6% of actions still directly given by π0.5. This experimental result partly demonstrates a temporarily effective division of labor—Astra acts as the embodied brain for slow-thinking functions like judging goals and checking task states; π0.5 acts as the specific execution system following Astra's commands, converting instructions and states into physical actions of the robot body, contact, and motion.
这一报告说明，大模型已经能拿下认知层，却仍吃不动物理层。因此，当前的可行解是通用推理加专用操作的混合架构，而非用大模型替代具身模型。成本则是另一重约束。世界模型初创公司脸谱心智创始人陆弘远，在机器人仿真测试平台RoboDojo上发现，完整地用GPT-6 Astra跑完一套实验，需要消耗20万美元。这笔高昂的实验费用恰恰表明，在大规模、长链路的机器人任务评测中，Astra的推理和试错成本仍然很高。
This report indicates large models can already take the cognitive layer but still cannot eat the physical layer. Therefore, the current feasible solution is a hybrid architecture of general reasoning plus specialized operation rather than replacing embodied models with large models. Cost is another constraint. World model startup Lianpu Xinzhi founder Lu Hongyuan found on RoboDojo that fully running a complete experiment with GPT-6 Astra costs $200,000. This high experimental cost precisely indicates that in large-scale, long-chain robot task evaluation, Astra's reasoning and trial-and-error costs remain high.
目前来看，更大的冲击体现在独立具身大脑公司的投资估值上。据中泰证券统计，2026年上半年，以具身大脑为核心业务的公司吸纳融资222.53亿元，占具身智能赛道融资的50.8%，而机器人本体整机公司同期融资仅56亿元，占比12.8%。头部创业公司的估值同步走高，银河通用、星海图、智平方、千寻智能、自变量机器人等大脑派公司相继突破200亿元估值。过去一年，具身智能公司估值水涨船高的核心原因，在于独立具身大脑这一叙事，即认为物理AI领域会诞生独立的OpenAI。
Currently, the bigger impact appears in independent embodied brain companies' investment valuations. According to Zhongtai Securities statistics, in the first half of 2026, companies with embodied brains as core business absorbed 22.253 billion yuan in financing, accounting for 50.8% of embodied intelligence track financing, while robot body complete machine companies raised only 5.6 billion yuan in the same period, accounting for 12.8%. Leading startups' valuations rose synchronously—Galbot, Galaxea, Zhipingfang, Qianxun Intelligent, and AgiBot successively broke through 20 billion yuan valuations. Over the past year, the core reason embodied intelligence company valuations rose was the "independent embodied brain" narrative—the belief that an independent OpenAI would emerge in physical AI.
但这一叙事逻辑可能迎来180度大转弯，物理AI领域的OpenAI，可能就是OpenAI本身。投资人肖虹将今年上半年涌现的一批世界模型创业团队视为具身泡沫中后期的信号，一些团队技术路线仍在快速分化，尚未证明能稳定落地，但估值已迅速被推至数亿美元。他认为，Astra一出，以世界模型为代表的具身大脑公司估值势必打折，最受冲击的将是资源不足、入局较晚、且纯讲具身大脑或世界模型故事的中腰部公司；头部公司会有足够的资金和资源实现转型。
But this narrative logic may face a 180-degree turn—the OpenAI of physical AI may be OpenAI itself. Investor Xiao Hong sees the batch of world model startup teams emerging in the first half of this year as a signal of "mid-to-late embodied bubble"—some teams' technical routes are still rapidly diverging, not yet proven stable deployment, but valuations have rapidly been pushed to hundreds of millions of dollars. He believes once Astra appeared, embodied brain companies represented by world models will inevitably see valuation discounts, with the hardest hit being mid-tier companies lacking resources, entering late, and purely telling embodied brain or world model stories; leading companies will have enough capital and resources to transform.
Astra之后，创业公司的分化将会出现。受冲击最直接的是只讲纯大脑或纯世界模型故事的中腰部公司。没有模型、技术可行性未验证、缺少真实结果的公司较为脆弱；有硬件基础、研发积累、真实业务的公司更有韧性。还有一类公司，转向的是具身小脑或与大模型协作。王怀认为，当大模型成为具身大脑，原本独立做具身大脑的公司更胜任转型成具身小脑公司，根据机器人对当前物理空间及所处状态的感知，更轻、更快地决策、精细操作，甚至形成条件反射。头部公司资金和资源更丰富，也更容易转向与Astra等大模型协作。
After Astra, startup divergence will appear. The most directly impacted are mid-tier companies purely telling pure brain or pure world model stories. Companies without models, unverified technical feasibility, and lacking real results are more fragile; companies with hardware foundations, R&D accumulation, and real business are more resilient. Another type of company turns toward embodied cerebellum or collaboration with large models. Wang Huai believes when large models become embodied brains, companies originally doing independent embodied brains are better suited to transform into embodied cerebellum companies—based on robot perception of current physical space and state, making lighter, faster decisions, fine operations, even forming conditioned reflexes. Leading companies with richer capital and resources more easily turn toward collaboration with large models like Astra.
大模型入局具身智能，不仅改写具身大脑创业公司的命运，其影响也传导到上游的数据供应商。对做具身数据的采集与处理的公司而言，OpenAI等大模型公司进入机器人领域，很大程度上是好事，因为客户源拓宽了，收入结构也可能随之改变。星忆智能是做第一人称人类数据采集与处理的初创公司，原本主要客户是具身智能领域的公司，但创始人宋知珩发现，来自大模型公司的客户正在显著增加。他预计，大模型公司对公司产品的需求量，将大于具身模型公司。原因在于，相比具身基模，大模型的参数量更大，理解能力更强，对真实世界数据，尤其对人类操作与交互数据的需求会更明显。
Large models entering embodied intelligence not only rewrites embodied brain startups' fate but also transmits impact to upstream data suppliers. For companies doing embodied data collection and processing, OpenAI and other large model companies entering robotics is largely good news because customer sources broaden and revenue structures may change accordingly. Xingyi Intelligent is a startup doing first-person human data collection and processing, originally mainly serving embodied intelligence companies, but founder Song Zhiheng found customers from large model companies increasing significantly. He predicts large model companies' demand for the company's products will exceed that of embodied model companies. The reason is that compared with embodied foundation models, large models have larger parameter counts and stronger understanding, with more obvious demand for real-world data, especially human operation and interaction data.
但客户源拓宽，不等于数据生意的商业模式已经成立。投资人的质疑并未消失。早期基金Alphaist创始人陈哲曾表示，数据公司必须提供今天AI还不能解决的问题的数据，等这些问题被AI学会解决，就需要提供更复杂、更高精度的数据。即使数据公司能成为模型公司的上游卖水人，其生意养活一支团队没有问题，但投资人肖虹对于这门生意是否能长得足够大，仍然存疑。一是数据是否能跨客户二次销售，是否能避免过度定制化从而降低整体成本；二是数据生意作为一种软件服务生意，是否会和SaaS一样，存在同质化和低价竞争的问题。
But broadening customer sources does not equal the data business model being established. Investors' skepticism has not disappeared. Early fund Alphaist founder Chen Zhe once said data companies must provide data for problems AI cannot yet solve today; once AI learns to solve these problems, more complex, higher-precision data is needed. Even if data companies can become upstream "water sellers" to model companies, their business can support a team, but investor Xiao Hong remains doubtful whether this business can grow large enough. First, whether data can be resold across customers, whether over-customization can be avoided to reduce overall costs; second, whether the data business as a software service business will face homogenization and low-price competition like SaaS.
Astra拉开了大模型进军物理世界的序幕，也击碎了单一的具身大脑叙事。资本正在重新定价，行业迎来分化洗牌。模型能力的边界还会持续扩张，但物理交互的工程门槛、本体约束、实时性要求不会轻易消失。具身智能的终局不是谁消灭谁，而是通用认知能力与物理执行能力持续融合。对创业公司而言，窗口还在，但不会太长。真正稀缺的不再是大脑或世界模型这些概念，而是把认知转化为可靠动作、把数据转化为场景闭环的能力。谁能在大模型与物理世界之间占据不可替代的一环，谁才有资格走向未来。
Astra opened the curtain on large models entering the physical world and shattered the single embodied brain narrative. Capital is repricing, and the industry faces divergence and reshuffling. Model capability boundaries will continue expanding, but physical interaction engineering thresholds, body constraints, and real-time requirements will not easily disappear. Embodied intelligence's endgame is not who eliminates whom but continuous fusion of general cognitive capability and physical execution capability. For startups, the window remains but will not be too long. What is truly scarce is no longer concepts like brains or world models but the ability to convert cognition into reliable actions and data into scenario closed loops. Whoever occupies an irreplaceable link between large models and the physical world is qualified to move toward the future.
