通用智能和具身智能之间的那条线正在变得模糊。过去两年，机器人公司们共享着一个前提：通用大模型长于理解数字世界，而物理世界需要一套从头训练的基座模型。这个前提撑起了行业的融资叙事，也撑起了一批公司的技术路线。9月初，它第一次受到来自大模型侧的正面挑战。
9月3日，OpenAI发布GPT-6 Astra，发布的第二天，就有团队把它接上机械臂，结果发现：Astra不只能理解任务、拆解任务，还能直接输出机器人末端执行器的位姿，完成了本该属于VLA等策略模型的那部分工作，在部分测试里成功率甚至超过专门为机器人任务训练的具身模型。
很快，通用大模型即将覆盖具身模型的说法在社交平台上快速扩散，而同一时间公开表态的具身公司创始人们，大多站在另一侧。他们的观点基本可以概括为：通用模型的能力来自各自领域的大量数据，而物理交互这一侧还没有同样的积累，所以复杂交互仍是具身公司的阵地。但也有创业者判断，留给具身和世界模型公司兑现预期的时间可能只有一到两年。
星尘智能创始人来杰并不认同这种具身模型和通用模型之间的简单对立。他认为有第三条路，AI是一波大势，具身公司肯定要做自己的模型，星尘的AI团队很快会发第三代基座模型，又要想办法吃下GPT-6 Astra之类通用模型带来的红利，而不是“拧着来”，“AOS会是模型之间最好的桥梁。黑客松已经持续验证了这一点”。
“我从来没有认为我们和通用大模型之间是对抗关系，也不认为它们是简单的分层关系。我更认可，它们通用智能和专用智能之间是一种伙伴关系，有着同等价值，需要各司其职，携手共进。”来杰表示，“星尘觉得对待AI的合理姿态，就是既要从第一天起就自己研发，也要尽早搭建这些智能所需要的‘脚手架’。当你搭建好这些底层架构之后会发现，两个伙伴协作，对机器人这些底层能力的发挥效果，是远超于一个人单干的。”
早在2024年训练自己的具身基模时，来杰曾突发奇想地给团队出了一道题：他用三支笔在桌上摆了一个箭头，让机器人把东西抓到箭头前面。结果，通用大模型理解“箭头”并放对位置，具身模型却理解不了。
“箭头这类抽象概念在专用模型的训练数据里几乎没有。”来杰的判断是，“我们要去相信大模型的发展，并且结合大模型的能力，从一开始训模型的时候就要找到分工。而并不是说我们一定要训出一个取代大模型的模型。”也是从那时起，Design for AI 成为星尘搭建操作系统的一条主线。
10月24日至25日，星尘智能在深圳举办机器人黑客松总决赛，这是该系列黑客松的第三场活动。“一句话生成机器人App”是该系列黑客松的核心表达，创作者大多是来自全国各地的学生、产品经理、独立开发者等。决赛现场的17支队伍在36小时里将想法推进到真机测试，16个App跑通真机，上真机率达94%。更值得注意的是，其中不少参赛者此前没有机器人开发经验。
之所以能做到这一点，正得益于星尘OS操作系统（Astribot OS，简称AOS）。这套系统将感知、移动、操作和交互等复杂能力封装为可调用、可组合的Meta能力包和Skill，并与最新的AI模型和Agent适配。创作者结合GPT、DeepSeek等大模型和AI编程工具，无需从零编写底层代码，便能从一句话需求出发，完成能力组合、流程设计、仿真测试和真机验证。星尘从2024年开始搭建这套系统，直到今年8月第一届黑客松启动时才定下AOS这个名字；此前，团队也考虑过AIOS。
写给模型看的接口
来杰曾在多个公开场合以苹果的模式解释星尘的战略布局：苹果通过硬件、操作系统和应用生态的协同，让更多人参与开发，也让用户更方便地使用产品。星尘希望在本体、AOS与未来的应用生态之间建立类似的联系，而自研Lumo系列模型则与AOS、本体共同构成星尘的三位一体架构。
“苹果图形界面把开发者从会写代码的那一小圈人，扩张到设计师、科学家和各行业的普通人，也让使用门槛从读懂命令行降到会使用鼠标就行。”来杰认为。而在今天的机器人行业，开发者仍主要集中在会写代码的专业圈层，App少，使用者也少。在来杰的设想中，AOS希望承担的，正是类似的角色。
来杰曾给AOS开发负责人提过一个要求：设计调用接口时，要优先考虑GPT能不能看懂。要让普通人参与机器人编程，就必须把自然语言翻译成能力要求、约束和执行规划的工作交给AI，接口因此也必须写成AI能读懂的样子。这就是AOS“Design for AI”的核心思路。
为了让GPT能够看懂接口，星尘的工程师需要多做一层工作。“程序员写接口的活变得像文字工作者，写完还得把这份能力讲给模型听。”来杰笑着说，“这些描述并非附在文档里就够了，而要写进接口本身，因为模型不会去查文档；描述不到位，这个能力在模型那里就等于不存在。它不知道有这个工具，自然不会调用。”
另一方面，星尘的工程师还必须为模型补足一部分能力。“模型算不准的地方要一项项包成工具，六自由度估计、握力估计、打滑检测，每一样都得先知道模型在哪里会失手，再把它做成模型能调用的形态。”来杰说，“大量的工作反而不是在于怎么去把硬件的接口写好，而是怎么能让智能真正理解它的工作。”
AOS的效果已经在星尘内部显现。工具链接通之后，星尘内部的开发速度提升了3到5倍，新来的应届生和实习生在开发速度上也能接近老手，因为他们不必先花几年熟悉机器人硬件。普通人开发应用也看到了可能性：第一届黑客松的参赛者中就有两名幼儿园老师，她们全程用自然语言互动，成功做出了逗小朋友玩的应用。
在来杰看来，硬件那一侧的接口都非常明确，而AOS后续要继续加强的，还是跟AI对接的接口。例如AI应该如何读懂失败：机器人出现抓不住、打滑的情况，返回给模型的信息得让它明白发生了什么、下一步问谁，而这条链路怎么解决，现在还没定论。另外，仿真到真机的落差、真正用一句话生成应用等问题，也都是AOS后续进一步优化的方向。
创作者生态
AOS尚未解决的，是模型与机器人如何更稳定地协作。而黑客松试图回答的，则是另一个问题：当底层能力被封装、开发门槛被降低之后，谁会成为机器人应用的创作者，他们又会做出什么？
总决赛的作品给出了一些初步答案。冠军团队“远程分身”把手机端与机器人连接起来，将铲猫砂、浇花、倒水、取快递和聊天等需求组合进同一个应用。还有团队让机器人写毛笔字、跟拍运镜，或接手3D打印结束后的收料工作。这些想法并非来自星尘预设的任务清单，而是来自创作者各自的生活经验。
“有创意的人没有掌握工具，有工具的人又没什么创意，还好中间有AI。”来杰的这句话，解释了星尘为何如此看重普通创作者。程序员往往更关心功能怎样实现，产品经理、设计师和行业从业者则可能更在意表达、互动和体验。机器人打完麻将后的“嘚瑟”、书写春联背后的生活记忆，未必是底层技术上的突破，却可能决定一个应用是否真正能够走进生活。
黑客松展示的是应用创作的入口，支撑这一入口的，则是星尘模型、操作系统与本体的三位一体架构。Lumo系列具身基座模型负责物理世界中的理解、预测和行动，并在需要时与通用大模型配合。AOS将感知、移动、操作和交互等能力组织成模型与开发者都能调用的工具。绳驱机器人本体则在真实环境中完成动作，呈现执行中的限制与问题。
这三者又并非简单的全栈拼接。模型能力的变化，会影响AOS需要提供怎样的接口；本体在抓取、受力和移动中遇到的问题，又会成为模型训练、工具补充和系统迭代的依据。AOS因此既连接通用智能与具身智能，也连接机器人的智能与身体。
在这一结构中，应用不是三位一体之外的第四块拼图，而是三者能力面向用户的出口，也是新需求进入系统的入口。创作者不必掌握全部底层技术，可以把精力放在需求、交互和体验上。应用在真实场景中暴露的问题，则有望推动星尘补充新的Skill、改进异常处理，甚至重新调整模型与本体之间的分工。星尘所设想的生态，因而不只是应用数量的增加，而是应用、模型、系统和本体能否共同进化。
不过，这个循环尚未完全闭合。“一句话生成App”目前仍需要人与AI反复沟通，应用也要经过仿真和真机调试。一个作品在比赛中跑通，更不等于它已经能够进入应用生态并产生收入。对普通消费者而言，稳定的使用体验、足够大的终端规模和可持续的开发者回报，仍然是必须补齐的条件。
商业化的距离
AOS的部分商业化已经开始，但主要发生在专业用户一侧。星尘已经销售面向科研和二次开发用户的SDK、遥操作等软件和工具，商业场景的二次开发者也在利用这些能力完成交付。硬件、软件与工具的组合，构成了现阶段更现实的收入来源，也为AOS继续迭代提供了真实需求。
星尘为生态设定的扩展顺序同样从已有买方出发：先连接商业场景中的二次开发和交付需求，再连接科研用户之间的模型、数据与工具需求，最后才走向更多真正购买和使用机器人的普通消费者。黑客松的价值，不只是选出几个优胜作品，也在于让星尘观察创作者怎样表达需求、AI怎样介入开发，以及哪些能力有可能成为下一次产品迭代的起点。
来杰在最近的一次公开采访中还提到一个观点：目前约200家具身公司，与未来最终上市的企业，可能只有10%重合。这并非看衰具身智能的长期价值，而是认为不同技术路线和商业选择不会全部成立，行业必然经历收敛。届时，一些公司会转向小而专或其他经营方式，也会有新的公司走出来。
在他看来，这些公司未必都来自今天的本体或模型厂商，基于AOS开发应用、深耕体验与创意的团队，同样可能成为下一批创业者。这一判断也让星尘的生态布局有了另一层含义：它不仅要降低机器人应用的开发门槛，也希望在行业价值向应用层扩展时，提前建立连接需求与创作者的入口。
AOS为通用大模型与具身模型之争提供了另一种答案：通用智能未必只是专用模型的替代者，也可能迅速拉低机器人应用的创作门槛。AOS真正要验证的，是机器人产业的价值，能否从厂商自己开发功能，扩展到更多人在其上创造功能。在应用稳定运行、创作者获得回报之前，星尘距离机器人时代的iOS和App Store仍有距离，但三次黑客松，已经让这条路径有了可被观察的样本。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The line between universal intelligence and intelligence is becoming blurred. Over the past two years, robotic companies have shared the premise that universal LLM is longer than understanding the digital world, and that the physical world needs a base model of fresh training. This premise has fueled the financial narrative of the industry and the technical routes of companies. In early September, for the first time, it faced a positive challenge from the side of LLM.
On 3 September, on the second day of OpenAIpublishGPT-6 Astra, publish, a team attached it to a mechanical arm, found that Astra could not only understand the mission, dismantle the mission, but also directly export the position of the robotic end implementer, and complete part of the work that would have belonged to a strategy model such as VLA, which, in part of the tests, even exceeded the body model that had been trained for robotic missions.
Soon, the claim that General LLM was about to cover the body model spread rapidly on social platforms, while the founders of the body company, who had spoken publicly at the same time, were mostly on the other side. Their views can largely be summarized as follows: the capability of the generic model is derived from a large amount of data in their respective fields, and the physical interaction has not yet accumulated on the same side, so complex interaction remains the position of the company. However, there is also the judgement of entrepreneurs that only one to two years may be left for physical and world model companies to deliver on their expectations.
The founders of stardust intelligence did not agree with the simple contradiction between this body model and the generic model. He believed that there was a third path, that AI was a wave of power, that the company would have to model itself, that the team of Stardust would soon send out a third-generation base model, and that it would find a way to reap the dividends of a generic model such as GPT-6 Astra, rather than “crush it” and that “AOS would be the best bridge between models”. Hackerston has continuously verified this.”
“I never thought that our relationship with General LLM was confrontational, nor that they were simple hierarchical relationships. I would like to acknowledge that their universal and dedicated intelligence is a partnership of equal value and requires that they work together.” “Stardust feels the rational attitude of treating AI is to develop itself from day one and to build the scaffolds that are needed for this intelligence as soon as possible. When you build these sub-structures, you find that the two partners working together are far more effective than a single person doing these sub-capacitys.”
As early as 2024, when he trained his own body model, Jai came to ask the team an unexpected question: He used three pens to put an arrow on the table, so the robot would grab it in front of the arrow. As a result, General LLM understood the "Arrow" and placed it in the right place, but the body model did not.
“Escaping concepts such as arrows are virtually non-existent in training data for specialized models.” The judgment is, "We have to believe in the development of LLM and combine the capabilities of LLM to find a division of labour from the beginning of the training model. And it's not that we have to train a model to replace LLM." And it's also since then that Design for AI has become the main line for building the operational system of stardust.
From 24 to 25 October, the Stardust Intelligence organized the robot hackson final in Shenzhen, the third event in the series. The core expression of the series, “App for a single sentence”, is the collection of hackers, mostly students, product managers, independent developers from all over the country. Seventeen teams at the finals site pushed their ideas to the real machine test in 36 hours, and 16 Apps ran the real machine, with a 94 per cent chance. More notably, many of these contestants had no previous experience in robot development.
This was achieved thanks to the Stardust OS operating system (Astrobot OS, known as AOS). The system covers the complex capabilities of perception, movement, operation and interaction as callable, combinationable Meta capacity packages and Skill, and is compatible with the latest AI model and Agent. In combination with the GPT, DeepSeek and LLM and AI programming tools, the creators are able to complete a combination of capabilities, process design, simulation testing and real machine validation without having to produce a zero code. Stardust built the system from 2024 until August, when the first hacker pine was launched, the name AOS was not established; the team had previously considered AIOS.
An interface written to the model
In many public places, Raj explained the strategic layout of stardust in the form of apples: Apples involve a greater number of people in their development through the synergy of hardware, operating systems and applications, as well as making products more user-friendly. Stardust hopes to establish a similar link between its own body, AOS and its future applied ecology, while the self-study of Lumo series models together with AOS, the home body, forms the triple architecture of stardust.
"The Apple GUI expands the developers from a small circle of code-writers to designers, scientists and ordinary people in various industries, and allows the threshold to be used from reading command line to using the mouse." And today, in the robotics industry, developers are still concentrated in professional circles, with fewer Apps and fewer users. What AOS wants to assume is a similar role.
He asked the AOS development manager to give priority to the GPT's understanding when designing the call interface. For ordinary people to be involved in robotic programming, the translation of natural languages into competency requirements, constraints and implementation planning must be entrusted to AI, and therefore the interface must also be written in the form that AI can read. That's the idea at the heart of AOS “Design for AI”.
To enable GPT to understand the interface, Stardust engineers need to do an extra layer of work. “The work of the programmer's writing interface becomes like a writer, and the ability to write is given to the model.” “These descriptions are not attached to the document enough, but are written into the interface itself, because the model will not look at the document; if the description is not in place, this capability is not present at the model. It doesn't know there's this tool, and it won't call."
On the other hand, the stardust engineer must also supplement the model with some of its capabilities. "Where the model is wrong, a package of tools, six degrees of freedom estimation, grip estimation, slide testing, each of which must first know where the model will fail and then make it into a modelable form." He said, “There's a lot of work going on, not how to write the hardware interface, but how to make intelligence truly understand its work.”
The effect of AOS is already visible inside the stardust. The speed of development within the stardust has increased by three to five times after the tool link, and new students and interns are able to develop at a rate that is close to their seniors because they do not have to spend a few years first familiarizing themselves with robotic hardware. There are also possibilities for ordinary people to develop applications: Among the participants in the first hackers, there were two kindergarten teachers who interacted with each other in their natural language and successfully used to entertain their children.
It seems to Jay that the interface on the hardware side is very clear, and what AOS continues to strengthen is the interface with AI. How, for example, AI should read failures: robots are caught, slipped, returned to the model with information that makes them understand what happened, next step asks who, and how this link is resolved, is still open. In addition, such issues as simulations and genuine use of one sentence to generate applications are areas for further optimization by AOS.
Creator Ecology
What AOS has not resolved is how models and robots can work more steadily. And the other question that Hackerston is trying to answer is: who becomes the creator of robotic applications when the bottom capacity is sealed and the development threshold is lowered?
The finals give some preliminary answers. The champion team "tele-diverse" connects the end of a cell phone to a robot and combines the needs of shoveling cat sand, pouring flowers, pouring water, pick-up couriers and chats in one application. There are also teams that allow robots to write quills, take mirrors, or take over the feed after 3D printing. These ideas are not derived from the pre-set list of tasks, but from the individual life experiences of the creators.
"Innovative people don't have the tools. People with tools don't have any ideas. This statement by Jay explains why stardust values ordinary creators so much. Programmers tend to be more concerned about how functions are delivered, and product managers, designers and industry practitioners may be more concerned with expression, interaction and experience. The memory of life behind a mahjong after the robot has finished playing mahjong and writing the Spring Union is not necessarily a breakthrough at the bottom of the technology, but may determine whether an application can really enter life.
Hacker Zone presents an entry point for applied creative work, which is supported by a three-way structure for stardust models, operating systems and the home system. The Lumo series has base-hub models for understanding, prediction and action in the physical world and works with General LLM when needed. AOS organizes sensory, mobile, operational and interactive capabilities as models and tools that can be accessed by developers. The rope-driven robotics themselves perform their actions in a real environment, presenting constraints and problems in implementation.
These three are not simple collages. Changes in model capabilities will affect the interface that AOS needs to provide; problems encountered by the body in capture, receptivity and movement will also form the basis for model training, tool supplementation and systems over time. AOS is thus connected to both universal intelligence and the intelligence and body of robots.
In this structure, the application is not a fourth puzzle outside the triad, but a three-man capability oriented towards the exports of users and an entry point to the system for new demands. The creators do not have to master all the technologies at the bottom, and can concentrate on demand, interaction and experience. Applying the problems that are exposed in the real scene is expected to drive stardust to replenish the new Skill, improve the treatment of anomalies, and even reorder the division of labour between models and the body. The ecology envisaged by stardust is not just an increase in the number of applications, but the ability of applications, models, systems and organisms to evolve together.
However, this cycle is not yet completely closed. “App generation” still requires repeated communication with AI, and applications need to be simulated and debugged. A work runs through the competition, let alone being able to access applied ecology and generate income. For ordinary consumers, stable use experience, sufficiently large end-scale and sustainable developer returns remain conditions that must be filled.
Distance from commercialization
Partial commercialization of AOS has begun, but mainly on the side of professional users. Stardust has been sold software and tools such as SDK, remote operation for scientific research and secondary development users, and the secondary developers of commercial scenes are using these capabilities to complete delivery. The combination of hardware, software and tools constitutes a more realistic source of income at this stage and provides real demand for AOS to continue to evolve.
The ecological order of expansion set by stardust also begins with existing buyers: first linking the demand for secondary development and delivery in the business landscape, then linking the demand for models, data and tools among research users, and finally moving towards more ordinary consumers who actually buy and use robots. The value of hackers and pines lies not only in the selection of winning works, but also in how to allow the stardust viewers to express their needs, how AI is involved in development, and what capacity is likely to be the starting point for the next product.
In a recent public interview, Raj also referred to the view that at present about 200 furniture companies, possibly only 10 per cent of the companies that will eventually be listed, overlap. This does not depend on the long-term value of being smart, but rather on the perception that different technological routes and business choices will not be fully established and that industries will inevitably experience a contraction. At that time, some companies would turn to small, specialized or other business and new companies would emerge.
In his view, these companies may not all come from today's home-grown or model manufacturers, but also from teams based on AOS development applications, deep-farming experiences and creativity, and may be the next entrepreneurs. This judgment also gives another meaning to the ecological configuration of stardust: It not only lowers the development threshold for robotic applications, but also wants to establish early access to demand and creators as industry values expand to the application layer.
AOS provides another answer to the GLLM contest with the body model: General Intelligence is not necessarily a substitute for a dedicated model, but may also rapidly lower the creation threshold for robotic applications. What AOS really wants to prove is the value of the robotics industry, and the ability to develop functionality from manufacturers themselves to more people creating it. Until the application is stable and the creators get their return, there is still a distance between the iOS and App Store of the robotic age, but the three hackings have allowed this path to have an observed sample.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.