机器人开发不再是专家们的专利。10 月 25 日，星尘智能主办的星尘 OS 机器人黑客松总决赛在深圳收官。在这场全国最大规模、面向多元背景创作者的机器人 App 黑客松中，17 支队伍奋战 36 小时，16 个 App 从仿真到跑通真机测试，上真机率达 94%。
远程铲猫砂和浇花、运镜跟拍、写书法……参赛者借助 AOS 丰富的工具包和 GPT-6、DeepSeek 等先进大模型，把想法做成 App、上真机验证，带来了护理、文娱、科研等领域的可运行应用，这不仅证明了“一句话生成 App”的可行性，更预示着星尘 AOS 正加速构建机器人时代的“iOS”和“App Store”的生态雏形。
星尘 OS 操作系统（Astribot OS，简称 AOS）把感知、移动、操作和交互等复杂能力封装成 Meta 包（机器人技能模块包），封装成可调用、可组合的 Meta 包和 Skill，并与最新的 AI 模型与 Agent 适配。创作者无需从零编写底层代码，从一句话需求出发，通过 AOS 组合能力、设计流程、仿真测试、真机验证，做成一个机器人 App。AOS 把复杂装进平台，简单留给创作者。
这是 Design for AI 的落点：AOS 既给人用，也给 AI 用。瞄准具身时代的苹果，星尘希望以 AOS 探索机器人时代的 iOS 与 App Store，让每个人都有机会成为机器人 App 的创造者和使用者。
从模型到本体，星尘全栈布局
今年以来，星尘智能陆续推进自研具身基座模型 Lumo-2（业界首个家庭隐式世界动作模型）和起售价 8.99 万元的第二代高精细操作机器人 T1，并布局万台级产能。随着 AOS 通过开发者活动进入更广泛的视野，公司围绕「AI 模型—具身操作系统—绳驱本体」构建的全栈架构持续升级。
从公司成立第一天开始，星尘就把模型、操作系统与本体放在一起。此次 AOS 面向更多开发者开放，是要让过去积累的能力被更多应用起来。
星尘认为物理 AI 的竞争不是单点技术的突破，而是全栈系统的竞争，在其技术体系中，Lumo 系列基座模型负责理解任务、预测并生成行动，并通过强化学习持续进化；AOS 围绕 AI 来组织任务、调用和组合机器人能力；绳驱机器人本体则在真实环境中执行动作。三者协同，将 AI 智能转化为物理世界中可调用、可组合、可扩展的任务执行能力，并通过实际部署与数据积累，持续优化机器人系统。
星尘坚持的 Design for AI（DFAI，面向 AI 设计），正是从设计之初就围绕 AI 的理解、决策和行动能力，协同设计模型、软件系统与机器人身体，而不是在传统机器人上简单叠加 AI。
“AOS 设计接口的时候，要考虑 GPT 能不能看懂，不能只考虑开发者能不能看懂。”星尘智能创始人、CEO 来杰在赛后采访中表示。AOS 从设计时就考虑了两类使用者：人可以用它开发应用并使用，AI 模型也可以通过 AOS 调用机器人的能力。
如果说模型决定机器人能够具备什么智能，本体决定它能够在物理世界中执行什么动作，那么操作系统所探索的，就是如何将这些能力组织起来，交给更多人使用和创造。
从一句话生成 App，到真实机器人执行任务
据主办方介绍，本次活动是国内规模最大的、面向多元背景创作者的机器人应用黑客松。北京站、深圳站及总决赛三站吸引 47 支队伍参与，创作者来自北京、上海、深圳、中国香港和中国澳门等地。学生、大厂产品经理、独立开发者等多元背景的创作者来到现场，其中最年轻的一批参赛者出生于 2006 年。他们带着各自对生活与工作的观察，尝试回答同一个问题：如果能为机器人开发一个应用，你希望它做什么？
通用大模型在理解需求、生成代码和任务规划上的进步，正在改变机器人应用的开发方式。不过，把任务变成实际动作，还需要机器人具备可调用的操作能力，并在执行中处理接触、受力和异常情况。模型越强，越需要一套系统把这些能力组织好。
与此同时，Vibe Coding 让更多人能够从描述需求开始开发应用，由 AI 辅助编写代码、调用能力，并在测试中持续迭代。
当这一变化进入 Physical AI，软件的输出便不再局限于屏幕，而有机会转化为现实世界中的行动。
不过，机器人应用开发仍存在一道门槛。一个看似简单的任务，往往涉及任务规划、接口适配、动作控制、异常处理与真机调试。即使熟悉软件开发，也未必掌握机器人控制技术；而专业工程团队也难以提前覆盖所有生活和生产需求。
围绕「一句话生成 App」原则，参赛者借助星尘 OS 提供的机器人能力模块（Meta 能力包）与开发工具，无需从零开始编写底层代码，即可调用已封装的视觉理解、语言交互以及移动、抓取等机器人能力；再结合 GPT、DeepSeek 等大模型和 AI 编程工具，搭建应用界面、后端代码和任务流程。应用在仿真环境中测试后，还需到人形机器人上调试、验证。这让创作者可以减少对底层控制、通信与硬件驱动的重复开发，把更多精力放在需求和应用体验上。
冠军团队「远程分身」小组深度调用 AOS 的 Meta 包，36 小时打通手机端↔机器人的双向实时控制链路，把图传、抓取、语音、移动底盘和操作能力捏在一起。让真机稳稳跑通铲猫砂、浇花、倒水、取快递、聊天等刚需活儿，还搭好了通用架构，让 T1 随时能变身家庭保安或差旅管家，具备极大的商业想象力。两支亚军团队 GowithFlow 和笔走龙蛇队，则分别探索机器人运镜跟拍摄影与毛笔写书法。
「远程分身」冠军队在赛后采访中提到，AOS 提供了“一套比较易用的开发者工具链”，从查资料、开发，到调试和部署，都有对应接口可使用。
生活里的具体烦恼，也成了应用的起点。季军团队「未定义」希望让机器人接手 3D 打印后的收料工作；最佳人气奖团队 CatDance·小猫跳舞的灵感，则来自旅行时家中猫砂盆故障、只能请朋友上门处理的经历。团队由此尝试开发「旅行管家」，探索离家期间的家庭照看。
另外两支季军团队中，过河小兵关注棋类对弈；Symbios 则用「光绘涂鸦」把脑洞画成光：用户与机器人一起构思图案，再由机器人挥动荧光棒，联动相机拍摄，把人物与光影留在同一张纪念照中。
这些案例的共同点是，需求并非来自机器人厂商预设的功能清单，而是来自创作者对真实生活的观察和趣味脑洞。随着机器人能力逐步封装，软件开发者、产品经理和行业从业者都将机会把熟悉的需求转化为机器人应用。
本次总决赛评委、北京大学科学智能学院院长助理、助理教授周大权在现场交流中，特别提到了毛笔书写和 3D 打印自动收料两个项目。在他看来，通用大模型与封装好的原子技能库配合，由 Agent 组织调用，是一种值得探索的机器人应用开发新范式。这让开发者有机会基于已有能力完成新任务，并进一步将应用分享给其他用户；其中的关键，是底层技能能否适应不同场景，具备足够的泛化能力。
此外，一句话生成 APP 特别强调真机测试，因为机器人必须面对真实环境中的物体、空间和安全约束，复杂应用需经过工程适配与真机验证。AOS 希望降低的是开发门槛，而不是绕过必要的工程流程。
机器人时代的“iOS”，想象力在应用生态
苹果的启示不只是打造优秀的硬件产品，更在于通过操作系统和开发工具，让第三方持续创造应用，拓展设备的用途。
星尘希望探索类似的路径，但机器人面临更复杂的挑战：应用不仅要在数字世界中运行，还必须在物理世界中完成任务。生态能否形成，取决于模型能力、本体执行、软件接口、应用供给和开发者回报能否共同发展。
对开发者而言，AOS 希望降低从需求到机器人应用的开发门槛；对用户而言，应用意味着有机会按需获得更多服务，而不必理解背后的模型与控制算法。
这也让 AOS 的战略价值超越单一软件工具：它连接星尘已有的模型与本体布局，并尝试为更多应用创造者提供入口。未来，机器人产品的竞争不只在于硬件参数或单项任务能力，也可能延伸至谁能提供更易用的开发工具、更多可复用的机器人能力，以及更丰富的应用选择。
但应用生态不会因为操作系统的出现而自动形成。星尘仍需要通过持续的产品迭代、稳定的能力接口、更多真机运行成果和真实用户需求，验证这条路线能否成立。
“我想要扩大价值，首先得扩大受众和用户。”来杰表示。星尘希望把应用的蛋糕做大，让更多软件开发者、产品经理和行业从业者带着需求参与。星尘过去解决过的问题，开发者不用重复再做一次，把更多精力放在自己的应用和用户体验上。
从 Lumo-2 到 T1，再到 AOS，星尘正在尝试连接 AI 的智能、机器人的身体与应用创造工具。如果说智能手机让人们拥有了可以不断扩展用途的个人计算设备，那么星尘希望探索的，是让更多人也能为个人机器人创造新的用途。
优胜团队与应用场景
本次总决赛的获奖应用覆盖远程管家、摄影、书法、对弈和家庭照看等场景。创作者从真实需求出发，借助 AOS 开发、调试各具特色的机器人应用。
冠军｜远程分身（手机远程管家）
人出门了，另一个「你」留在家。「远程分身」小组深度调用 AOS 的 Meta 包，36 小时打通手机端↔机器人的双向实时控制链路，把图传、抓取、语音、移动底盘和操作能力捏在一起。让真机稳稳跑通铲猫砂、浇花、倒水、取快递、聊天等刚需活儿，还搭好了通用架构，让 T1 随时能变身家庭保安或差旅管家，具备极大的商业想象力。
亚军｜GowithFlow（拍照大师）
面向景区游客和短视频创作者，探索人物锁定、自动构图和移动跟拍，尝试让机器人充当摄影助手。
笔走龙蛇队（书法应用「问墨」）
让机器人拿起毛笔，根据指定内容书写。团队中的双胞胎兄弟从小看爷爷写春联、到镇上售卖，这次把这段生活记忆带进比赛，探索机器人书写与赠送墨宝的应用。
季军｜过河小兵（机器人下棋）
把棋类对弈搬到机器人上，探索与用户面对面下棋的互动体验。
Symbios（光绘涂鸦）
「光绘涂鸦」让用户和机器人共同构思图案，再由机器人挥动荧光棒绘制，联动相机拍摄，把人物与光影留在同一张纪念照中。
未定义（3D 打印收菜）
「3D 打印收菜」瞄准打印结束后的人工收料环节，尝试让机器人取下成品、放回打印板并开启下一轮打印，接续自动化流程，让使用者少为收模型设一个半夜闹钟。
最佳人气奖 CatDance·小猫跳舞（旅行管家）
一次旅行中，家里的电动猫砂盆突然卡住，朋友只能打车上门处理。三位互联网大厂工作的成员，此前都没有机器人开发经验，却从自己的麻烦出发，提出「旅行管家」方案，围绕铲猫砂等离家照看任务开发应用。
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。
Robotic development is no longer a patent for experts. On October 25th, the Stardust OS robot hackson finals, sponsored by Stardust Intelligence, gathered in Shenzhen. Of the country ' s largest robot, App Hawthorne, 17 teams fought for 36 hours, 16 Apps, from simulation to running machine tests, with a 94% chance of success.
Long-range shoveling of cat sands and flowering, mirroring and filming, and writing of calligraphy... The contestants, using the AOS rich toolkit and advanced LLM, such as GPT-6 and DeepSeek, have developed ideas into App, real-time validation, bringing operational applications in such areas as nursing, recreation, scientific research, not only to prove the feasibility of “App to produce a single sentence”, but also to predict that Stardust AOS is accelerating the construction of the ecological prototypes of the robotic era of “iOS” and “App Store”.
The Stardust OS operating system (Astrobot OS, abbreviated by AOS) covers the complex capabilities of perception, movement, operation and interaction in Meta packages (robots skill modules) and in callable, combustible Meta packages and Skill, and corresponds to the latest AI model with Agent. The creators do not have to write the bottom code from zero, starting with one sentence and using AOS combination capabilities, design processes, simulation tests, genuine machine validation as a robot App. AOS puts complexity on the platform, leaving it simply to creators.
This is the drop point for Design for AI: AOS is for people as well as AI. Aim at the apples of the body age, the stardust hopes that AOS will explore the iOS and App Store of the robotic age, giving everyone a chance to be the creator and user of the robot App.
From the model to the body, the stardust whole stack layout
Since the beginning of this year, Stardust intelligence has been developing the Lumo-2 model and the start-up price of $899 million in second-generation high-precision robots, T1, with the deployment of 10,000 stage-grade capacity. As AOS enters a broader perspective through its developers' activities, the company continues to upgrade its entire structure around the "AI Model - A hands-on operating system - a rope-pulsor."
From the first day of the company ' s existence, stardust brought together models, operating systems and the physical body. This time, AOS is open to more developers to make the capabilities accumulated in the past more applicable.
Stardust believes that the competition for physical AI is not a breakthrough in single-point technology, but rather a competition for the whole-storey system, in which the Lumo series base model is responsible for understanding tasks, predicting and generating actions and for continuous evolution through enhanced learning; AOS organizes missions, calls and combinations of robotic capabilities around AI; and the tethered robotic body operates in the real environment. Three are working together to transform AI intelligence into an accessible, groupable, scalable mission implementation capability in the physical world and to continuously optimize robotic systems through actual deployment and data accumulation.
Design for AI (DFAI, design for AI) is the foundation of AI's ability to understand, make decisions and act together to design models, software systems and robotic bodies, rather than simply superimpose AI on traditional robots.
"When designing the AOS interface, consider whether GPT can read, not just the developers." The founder of Stardust Intelligence, CEO, said in a post-selection interview. AOS has been designed to consider two types of users: people can use it to develop applications and use them, and AI models can also call on robotics through AOS.
If models determine what intelligence robots can have and what actions they can perform in the physical world, then what the operating system explores is how to organize these capabilities and give them to more people for use and creation.
From a word to an App to a real robot.
According to the organizers, the event was the largest robotic application for creators of diverse backgrounds in the country. Beijing, Shenzhen and the three finals attracted 47 teams from Beijing, Shanghai, Shenzhen, Hong Kong, China and Macao, China. The scene was attended by creators from diverse backgrounds, such as students, product managers and independent developers, the youngest of whom was born in 2006. They try to answer the same question with their own observations of life and work: If you could develop an app for a robot, what would you like it to do?
General LLM advances in understanding demand, generating codes and mission planning are changing the way robotic applications are developed. However, operationalizing the mission will also require robotics to have ready operational capability and to deal with exposure, stress and anomalies during their implementation. The stronger the model, the greater the need for a system to organize these capabilities.
At the same time, Vibe Coding allows more people to develop applications from the description of needs, with AI-assisted code writing, callable capabilities and continuous iterativeity in testing.
When this change enters the Physical AI, the software output is no longer limited to screens, but has the opportunity to be translated into action in the real world.
However, there is still a threshold for robotic application development. A seemingly simple task often involves mission planning, interface fit, movement control, anomaly processing and debugging. Even when familiar with software development, robotic control techniques are not necessarily available; and professional engineering teams have difficulty covering all life and production needs in advance.
Around the "App" principle, the contestants use the Meta Capability Module (Meta Capability Package) and development tools provided by Stardust OS, without having to start from scratch, to call on enclosed visual understanding, language interaction and robotic capabilities such as moving, capturing, etc.; and to combine the GPT, DeepSeek and other LLM and AI programming tools with application interfaces, backend codes and task flow. After testing in a simulation environment, human robots need to be debugged and validated. This allows creators to reduce the duplication of bottom controls, communications and hardware-driven development and to focus more on demand and application experiences.
The team called AOS’s Meta package at depth, 36 hours to access the two-way real-time control chain of the mobile phone-end robots, and squeezed drawings, grabs, voice, moving chassis and operating capability. To keep a steady run through cat dusts, flowers, water, couriers, chats, etc. in order, a universal structure has been put in place that allows T1 to become family security or travel housekeepers at all times with great commercial imagination. Two sub-military teams, GowithFlow and the pen-and-breather team, explored robotic mirrors and photography and brush writing, respectively.
In a post-race interview with the Long-Range Detachment Champions, AOS mentioned that AOS provides a “more user-friendly development tool chain” with a corresponding interface from information checking, development, to commissioning and deployment.
Specific problems in life are also the starting point for application. “Undefined” the team wants the robot to take over the 3D printout; the best human-spirited team, CatDance Catcat, is inspired by the experience of a family cat’s pelvis failure during a trip, which only involves friends. As a result, the team tried to develop a "travel housekeeper" to explore home care during the absence.
Among the other two teams, the passers-by soldiers followed the chess game; Symbios used the word "photographic graffiti" to paint the brain hole: users worked with robots to frame it, and robots waved the fluorescent rods, linked the camera, and left the person with the light in the same photo.
The common denominator of these cases is that demand does not come from a list of functions pre-set by robot manufacturers, but rather from creators ' observations of real life and interesting brain holes. As robotic capabilities are gradually sealed, software developers, product managers and industry practitioners will have the opportunity to translate familiar needs into robotic applications.
In the course of the field exchange, Assistant Director and Assistant Professor Zhou Zhou of the Beijing University School of Science and Smart Science, the judges of this final competition made special mention of the items curry writing and 3D printing automatic intake. In his view, General LLM, in conjunction with the sealed atom skill pool, called by Agent, was a new paradigm for robotic application development worth exploring. This provides an opportunity for developers to build on their ability to perform new tasks and to further share applications with other users; the key is whether the bottom skills are adapted to different scenarios and have sufficient generalization capacity.
In addition, the first sentence generates an APP that places special emphasis on real machine testing, as robots have to face objects, space and safety constraints in the real environment and complex applications have to be validated by engineering and real machine. AOS wishes to lower the development threshold rather than bypass necessary engineering processes.
The iOS of the robotic age.
The inspiration of apples is not just the creation of good hardware products, but the continuous creation of applications by third parties and the expansion of the use of equipment through operating systems and the development of tools.
Stardust wants to explore similar paths, but robots face the more complex challenge of applying not only in the digital world but also in the physical world. Ecology depends on the co-development of modelling capacity, physical implementation, software interfaces, application supply and developer returns.
For developers, AOS wants to lower the development threshold from demand to robotic applications; for users, application means access to more services on demand without understanding the underlying models and control algorithms.
It also allows AOS to have a strategic value beyond a single software tool: it connects existing models and home layouts of stardust and attempts to provide access to more application creators. In the future, the competition for robotic products will not only be based on hardware parameters or individual mission capabilities, but may also extend to those who can provide more user-friendly development tools, more reusable robotic capabilities and more extensive application options.
However, applied ecology does not automatically result from the emergence of operating systems. Stardust still needs to verify the viability of this route through continuous product overlay, stable capability interfaces, more real machine performance and real user needs.
“I want to expand the value, first of all by expanding the audience and the users.” Jay said. Stardust hopes to make the cakes used larger, involving more software developers, product managers and industry practitioners with needs. Stardust has solved problems in the past, and developers need not repeat it and focus more on their applications and user experiences.
From Lumo-2 to T1, then to AOS, stardust is trying to connect AI with the smart, robotic body and application creation tool. If smartphones allow people to have personal computing devices that can expand their use, what Stardust hopes to explore is that more people can create new uses for individual robots.
The winning team and the application scene
The award for the finals will cover remote housekeepers, photography, calligraphy, games and home care. The creators started with real needs and developed, debugged and adapted robotic applications with AOS.
The winner's remote bilocation.
People left, and another "you" stayed at home. The team called AOS’s Meta package in depth, placing 36 hours into a two-way real-time control chain of mobile phone-side robots, squeezing maps, grabs, voice, moving chassis and operating capability. To keep a steady run through cat dusts, flowers, water, couriers, chats, etc. in order, a universal structure has been put in place that allows T1 to become family security or travel housekeepers at all times with great commercial imagination.
Gowith Flow.
Targeting viewers and short video creators, exploring the locking, automatic mapping and moving and filming of people and trying to use robots as camera assistants.
The Dragon Snakes.
Let the robot pick up the quill and write according to what's specified. The team's twin brothers, who went to the town from childhood to see Grandpa write the Spring Union and sell it, took this memory of life to the competition and explored the use of robot writing and gifted ink.
Queens across the river.
Move the chess match to the robot to explore the interaction with users.
Symbios
The "light-painted graffiti" allows users and robots to co-design a pattern, drawn by robots swinging fluorescent rods, linked cameras, and leave the person and the shadow in the same picture.
Undefined (3D print order)
The 3D Print Collection aims at the artificial feed chain after printing, trying to get robots to take off the finished product, put back on the printboard and open the next round of printing, to follow up the automated process and to keep users less able to set up a midnight alarm for the collection model.
CatDance Kitty Dance.
In one trip, the electric cat basin in the home was suddenly stuck and friends had to hit the door. Three members of the major Internet factory, who had no previous experience in robot development, initiated a travel housekeeper programme to develop applications around out-of-home care tasks such as shoveling of cat sands.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.