今年9月，OpenAI发布了被称为迄今能力最强的模型GPT-6 Astra。很快，社交媒体上就有人直接用Astra操控机械臂，完成从抓取到装配的各类任务。第三方机构Robocurve公布的对照测试显示，在把积木放进碗里这类简单任务上，GPT-6 Astra零样本条件下20次测试完成19次，成功率95%，同期Claude Fable 5.1的成绩是40%。这一下，行业坐不住了。此前的具身玩家，一部分是从小脑长出来的，做机械臂、底盘、人形本体与运动控制，靠执行能力吃饭，另一部分是具身模型公司出身，从数据、仿真和端到端算法里一点点磨做的本事。如今通用基座模型直接包揽了机器人的规划与粗抓取，再叠加大模型厂商更快的迭代节奏，大模型厂会吃掉所有具身模型厂的说法，一时传遍全行业。
In September this year, OpenAI released GPT-6 Astra, called its most capable model to date. Soon, people on social media were directly using Astra to control robotic arms, completing tasks from grasping to assembly. Third-party organization Robocurve's controlled tests showed that on simple tasks like putting blocks into a bowl, GPT-6 Astra completed 19 out of 20 tests under zero-shot conditions, a 95% success rate, while Claude Fable 5.1's score was 40%. This set the industry on edge. Previous embodied players were partly those who grew from the "cerebellum," making robotic arms, chassis, humanoid bodies, and motion control, living on execution capability; and partly embodied model companies, slowly honing "doing" skills from data, simulation, and end-to-end algorithms. Now general foundation models directly handle robot planning and coarse grasping, combined with large model makers' faster iteration rhythm—the saying "large model makers will eat all embodied model makers" spread through the industry.
恐慌之前，有个问题或许要先想清楚，OpenAI为什么能一步跨到这里。在原力灵机看来，答案不在模型架构，而在数据、算力和时间。具身智能和大语言模型一样，最终是一场资源战，谁的素材更多、算力更足、熬得更久，谁就先摸到下一个能力台阶。2026年是具身智能的上市大年，年内已有5家机器人相关企业登陆港股，梅卡曼德、优地机器人、仙工智能等先后敲钟，宇树科技也在8月登陆科创板。走过Demo时代，商业化怎么选、技术路径怎么走，都被摆上了台面。OpenAI的入场，给这场混战又添了一把火。
Before panicking, one question may need clarifying first—why could OpenAI step here in one move. In Dexmal's view, the answer is not in model architecture but in data, compute, and time. Embodied intelligence, like large language models, is ultimately a resource war—whoever has more material, more compute, and endures longer touches the next capability step first. 2026 is a "big IPO year" for embodied intelligence, with 5 robot-related enterprises listing on Hong Kong stocks this year, Mech-Mind, UDI Robot, and Seer Intelligent successively ringing the bell, and Unitree also listing on the STAR Market in August. Having passed the "Demo era," how to choose commercialization and which technical path to take are all on the table. OpenAI's entry adds fuel to this melee.
OpenAI并不是没做过机器人。恰恰相反，它做得很早，也失败得很早。2015年到2016年，OpenAI就认可模型规模加数据规模同步scale的路线，也尝试过具身方向，但最终没有做起来。过去，最核心的卡点始终是物理世界数据匮乏。当时机器人数据起不了量，而OpenAI又是沿着模型规模加数据规模同步scale的逻辑，这成了制约其发展的原因。而现在，OpenAI终于能把当年缺的东西补上了。2026年1月，Business Insider援引知情人士称，OpenAI在旧金山建立机器人实验室，约有100名数据采集人员，使用遥操作方式让机械臂完成叠衣服、把面包放入烤面包机等任务。
OpenAI did not avoid robots. On the contrary, it did them very early and "failed" very early. In 2015–2016, OpenAI already recognized the route of scaling model size plus data size synchronously and tried the embodied direction, but ultimately could not make it work. Previously, the core bottleneck was always the scarcity of physical world data. Robot data could not scale at the time, and OpenAI followed the logic of scaling model size plus data size synchronously—this became the constraint on its development. Now, OpenAI can finally fill what was missing then. In January 2026, Business Insider cited people familiar with the matter saying OpenAI established a robotics lab in San Francisco with about 100 data collectors using teleoperation to have robotic arms complete tasks like folding clothes and putting bread in a toaster.
行业内要实现具备通用能力的人形机器人大脑，至少需要一百万小时的高质量数据。而行业里积累的可用数据，目前也就刚到百万小时量级。根据报道，OpenAI目前所使用的具身数据超过千万小时级。GPT-6的预训练，就灌入了大几百万甚至上千万小时的ego数据。原力灵机联合创始人汪天才提到，十年前不具备的条件，今天具备了。所以，在原力灵机看来，大模型本质上是一场数据创新，OpenAI之所以形成断崖式领先，关键在于其庞大的数据优势。中国具身机器人行业若想追赶OpenAI，首先需要迈过千万小时级训练素材的量级门槛。
For the industry to achieve a humanoid robot brain with general capabilities, at least one million hours of high-quality data are needed. The accumulated usable data in the industry is currently just reaching the million-hour scale. According to reports, OpenAI's embodied data currently exceeds ten million hours. GPT-6's pretraining was infused with millions to tens of millions of hours of ego data. Dexmal co-founder Wang Tiancai mentioned that conditions unavailable ten years ago are available today. So in Dexmal's view, large models are essentially a data innovation—OpenAI's cliff-like lead key lies in its massive data advantage. If China's embodied robot industry wants to catch up with OpenAI, it first needs to cross the ten-million-hour training material threshold.
不过，把OpenAI神化同样不客观。具身智能除了需要大量的数据训练，也需要足够精准的数据。在GPT-6发布后，原力灵机也进行过测试，一些简单任务，比如在仿真器里抓方块这类操作问题不大，但换到精巧的操作类任务，短板就暴露出来。在测试中，一旦切换到精密拼图插入、精细接触操作，Astra迅速失效，和竞品同样成功率很低，仅有10%左右。模型能够理解任务、进行推理，并不意味着它已经掌握了物理世界的全部操作能力。如果缺少高质量物理世界交互数据，技术路线就无法快速收敛。
However, deifying OpenAI is equally objective. Embodied intelligence requires not only massive data training but also sufficiently precise data. After GPT-6's release, Dexmal also tested it—simple tasks like grasping blocks in a simulator were not a problem, but switching to delicate manipulation tasks exposed shortcomings. In tests, once switched to precision puzzle insertion and fine contact operations, Astra quickly failed, with success rates as low as competitors, only about 10%. A model understanding tasks and performing reasoning does not mean it has mastered all physical world operation capabilities. Without high-quality physical world interaction data, technical routes cannot converge quickly.
大语言模型之所以能爆发，是因为互联网文本数据近乎无限且标注成本低。但物理世界不同，每一次抓取、每一个关节运动，都是高成本、低频率的孤立事件。缺乏大规模、高质量的身体记忆，模型就像读过万卷书却从未下过田的农夫，理论完美，落地蹒跚。原力灵机联合创始人汪天才举了个例子，有人把篮球扔过来，你只识别出有个球飞过来了不够，还得预判它的动力学轨迹，才能接得住。数字世界完全不关心物理动力学，你不会在意一段代码按什么运动规律运行，但物理世界是真实的，处处都是动力学。
Why large language models could explode is because internet text data is nearly infinite and annotation costs are low. But the physical world differs—every grasp, every joint movement, is a high-cost, low-frequency isolated event. Lacking large-scale, high-quality "body memory," models are like farmers who have read ten thousand books but never gone to the field—perfect theory, stumbling deployment. Dexmal co-founder Wang Tiancai gave an example: someone throws a basketball at you; recognizing "a ball is flying over" is not enough—you must predict its dynamic trajectory to catch it. The digital world does not care about physical dynamics at all; you do not care what motion law a piece of code runs by. But the physical world is real—dynamics are everywhere.
不难预见，站在时间的维度上，作为世界级的大公司，OpenAI的财力雄厚，随着其持续投入发展，国内具身智能公司的竞争壁垒难免被缩小。但对于大模型厂威胁具身模型厂的言论，手握丰富真实场景数据的原力灵机内部，反而显得乐观。理由也很简单，头部大模型厂商这一轮下场，终于让行业看到，也重视了具身原生的价值。那么，这一波轮到具身智能赛道里，未来留下的将是哪些玩家。原力灵机仍然把追赶的关键归纳成上述提到的三点，即时间、数据、算力。
It is not hard to foresee that from the time dimension, as a world-class company with deep financial resources, OpenAI's continued investment will inevitably narrow domestic embodied intelligence companies' competitive barriers. But regarding the "large model makers threaten embodied model makers" talk, Dexmal, holding rich real-scenario data, appears optimistic internally. The reason is simple—leading large model makers entering this round finally let the industry see and value embodied-native worth. So in this wave of embodied intelligence, which players will remain in the future. Dexmal still summarizes the key to catching up as the three points mentioned above—time, data, compute.
至于怎么追OpenAI上千万小时级的数据量，原力灵机认为核心变量是资源整合能力。汪天才指出，国内行业最大的症结是资本市场过度分散，各方单打独斗，资金被摊薄到大量企业身上，行业内部自相消耗。与美国市场资源高度集中于OpenAI等少数头部不同，国内融资还比较分散。汪天才表示，与其如此，不如集中力量办大事，将训练数据渐进积累，资金、人才也同步整合。原力灵机认为，或许当前行业正处于三分天下之前的混战阶段，未来真正做模型、做大脑的企业可能收缩至四五家，其余中小厂商或被并购、被合并，或出清，资金向头部集中。
As for how to catch up with OpenAI's tens-of-millions-hour data volume, Dexmal believes the core variable is resource integration capability. Wang Tiancai pointed out that the biggest crux in the domestic industry is excessive capital market dispersion—all parties fighting alone, funds diluted across large numbers of enterprises, internal mutual consumption. Unlike the US market where resources are highly concentrated in a few leaders like OpenAI, domestic financing is still relatively dispersed. Wang Tiancai said that rather than this, it is better to concentrate forces for big things, gradually accumulating training data while integrating capital and talent. Dexmal believes the industry may currently be in a melee stage before "three kingdoms division," and future companies truly doing models and "brains" may shrink to four or five, with other small and medium manufacturers acquired, merged, or eliminated, with funds concentrating toward leaders.
最后，对于通用模型能否通吃所有垂类模型的探讨，原力灵机方面的回答是，其正在打造的通用具身模型，即具身机器人的基模、基础模型，相当于人的九年义务教育阶段，培养的是逻辑能力、对知识的理解能力与基础学科素养。完成这一阶段后，再学习修车、厨师、记者、科学家等细分场景，将事半功倍。所以，OpenAI的入场，到底改变了什么。它把具身智能从Demo叙事拽回了现实，证明了通用模型在高层规划和粗操作上的爆发力，也把物理世界精细交互的鸿沟，摊开给了所有人。未来的赛点，不在谁更像人形，而在谁能把数据、算力、模型与真实场景拧成闭环，对国内具身公司而言，真正的危机或许不是OpenAI太强，而是资源太散、投入太浅、耐心太少。
Finally, on whether general models can eat all vertical models, Dexmal's answer is that its general embodied model—the embodied robot's foundation model—is equivalent to a person's nine-year compulsory education stage, cultivating logical ability, knowledge understanding, and basic subject literacy. After completing this stage, learning subdivided scenarios like car repair, chef, journalist, or scientist will yield twice the result with half the effort. So what exactly has OpenAI's entry changed. It pulled embodied intelligence from "Demo narrative" back to reality, proving general models' explosive power in high-level planning and coarse manipulation, and exposing the physical world fine interaction gap to everyone. The future competition point is not who looks more humanoid but who can twist data, compute, models, and real scenarios into a closed loop. For domestic embodied companies, the real crisis may not be OpenAI being too strong but resources being too scattered, investment too shallow, and patience too little.
