您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语： 具身智能的资本筛选标准，正在从"讲得动人"转向"兑现得动"。
作者丨齐铖湧
编辑丨岑 峰
AI 科技评论获悉，11 月 2 日，RoboScience 机器科学完成 A2 轮融资。
据悉，本轮由琥珀资本、鸿瑞达投资、上海半导体产投 联合领投，360 战投、爽资本、国联民生资本、金鼎资本及多家上市公司平台跟投，融资金额达到了数亿元。
这意味着，这家 2025 年 3 月才正式开始运营的具身公司，先后已完成五轮融资。
据 AI 科技评论了解，近期具身行业融资市场并不乐观。过去一年，这个赛道的融资叙事换了好几轮：最早讲世界模型就能拿到钱；后来光讲技术不够了，得讲落地，讲进工厂的 ROI ；再后来，得有"具身 RSI "（递归自我改进）的想象空间。
故事通胀的背后，是大多数团队的底子撑不起叙事的升级，每次风口转向，就得重新攒一套说辞。资本也随之变得挑剔，同样的 BP，去年能融到，今年连会都约不上。（雷峰网(公众号：雷峰网)）
但 RoboScience 机器科学的节奏完全不同，逆势不断受到资本青睐，也因此成为具身领域里一个值得研究的样本公司。
回看机器科学过去五轮融资的时间点，几乎每一轮前后，都有可验证的产品技术落地：ICRA 最佳论文、仿真平台 RoboMirage、通用具身大模型 Visics、与腾讯云共创的云端具身智能服务 EaaS、首款轮式仿人形通用机器人 REX G1 以及真实的项目落地。
某种意义上，这意味着，具身市场的融资，不该是技术跟着融资的故事走，反过来，融资应该跟着技术的兑现走。
在 AI 科技评论看来，这是本轮融资最值得关注的信号：具身智能的资本筛选标准，正在从"讲得动人"转向"兑现得动"。
一家公司的融资节奏如果能与技术节奏对齐，说明它手里握着可以持续拿出来的东西。在 AI 科技评论看来，这跟创始人团队有着分不开的关系。
01
斯坦福双子星，
如何撑起一家中国具身头部公司
在中国具身行业，从占据头部公司的创始人学术履历已经能看出几条清晰的派系：清华系、伯克利圈、斯坦福圈等等。
其中影响最深的，是斯坦福圈。
银河通用王鹤、穹彻智能卢策吾、Hillbot 苏昊，RoboScience 机器科学的（）这些公司和创始人路线各不相同，但往上追，大多能追到同一个实验室，美国三院院士、几何计算学者 Leonidas J. Guibas 在斯坦福的组。同一条脉络上还有 OpenAI 研究员严梦媛、清华大学助理教授弋力。
这不算偶然。
斯坦福几乎是全球具身智能研究的策源地之一：李飞飞，作为 ImageNet 发起人，视觉与数据驱动范式的奠基者，Chelsea Finn 的 REAL Lab ，以及几何计算学者 Leonidas J. Guibas，以及国外头部具身公司 Pi 创始成员都来自斯坦福。
2016 年前后，斯坦福 Leonidas J. Guibas 教授（美国三院院士 ）的几何计算组，聚集了一批华人学生，其中就包括王鹤、邵林和他们的师兄苏昊等。
Guibas 教授组有一条相当一致的研究主线：3D 视觉出身，重点钻研操作泛化，相信抽象表征和仿真。邵林的个人轨迹几乎是这条主线的完整样本。
2014 年邵林进斯坦福，他先在几何计算泰斗、美国三院院士 Leonidas J. Guibas 组里做 3D 视觉，后转向机器人操作，师从 Jeannette Bohg，现任新加坡国立大学助理教授。
学术方面，邵林的一作论文 UniGrasp 是数据驱动灵巧手抓取的基准性工作；2025 年，团队凭借 D(R,O) Grasp 从四千多篇投稿中斩获 ICRA 机器人操作与运动方向最佳论文奖，成为近五年亚洲机构首次以第一单位身份拿下这一奖项的团队，打破了欧美的长期垄断；2026 年，Bi-Adapt 再度入围该奖项提名，使其成为亚洲唯一连续两年站上这个位置的队伍。同场角逐的是伯克利、MIT、斯坦福、清华。
此外，斯坦福系还有另外两条支线，分别是李飞飞线和吴恩达线。李飞飞线走出了卢策吾、非夕王世全，吴恩达线走出了田野。
2014 年，田野以中科大物理系专业第一的身份进入斯坦福 AI Lab，师从斯坦福 AI 教父吴恩达，与邵林同年入校。之后他在苹果待了七年，做到 Apple AI Platform 技术负责人，是苹果总部最年轻的主任工程师之一，主导打造了被业内称为“苹果的 PyTorch 与 CUDA ”的 AI 平台以及全球首个端云协同大模型推理系统的建设，后者要服务的不是实验室里的测试机，而是超过 10 亿用户、20 亿设备，包括 iPhone、Vision Pro 和 Apple Intelligence。
这段经历留给他的核心判断是：苹果的壁垒从来不是某个单点技术，而是软硬一体的系统能力。
在 AI 科技评论看来，这段经历塑造了田野的一个基本判断：真正难的不是单点技术突破，而是让技术跨设备、跨场景稳定运行的系统能力。
邵林和田野为旧相识，图片为早期合影
2024 年底，一个从苹果出来管工程，一个从学界出来管学术，两条支线汇合成 RoboScience机器科学。放在斯坦福系的谱系里看，这可能是“学术纵深+工程纵深”配置最完整的一家创业公司：邵林那条从 UniGrasp 到 D(R,O) Grasp 再到 Bi-Adapt 的路线，对应的是模型的跨本体泛化能力；田野在苹果做端云协同、服务十亿级设备的经验，对应的是把能力做成可部署、可迭代的系统。两人的组合，就是 RoboScience 技术路线图的另一种写法。
02
三个重大节点的具身闭环
从 2025 年 9 月发布 RoboMirage 仿真引擎，到 2026 年 6 月发布 Visics 通用具身大模型，再到 8 月发布 REX G1 轮式仿人形机器人，RoboScience 用约一年时间，把“数据、模型、本体”三个环节全部自研并且打通。这种技术组合，不是 RoboScience 独有，但在所有做全栈的公司里，AI 科技评论认为 RoboScience 的技术选择意图，是最明确的。
数据层，先解决"吃得起"的问题。 具身智能最大的瓶颈不是算法，而是数据和环境，真机遥操采集又贵又慢，行业普遍困在"数据规模撑不起 Scaling Law"的循环里。
RoboMirage 从底层求解器做起，完成了对具身世界中全物体空间的覆盖，效率和并发性也达到了业界先进水平。刚性柔性材料耦合、极薄物体的接触力，甚至"洗牌""拉扯橡皮筋"这类遥操都难以完成的动作，都能高精度模拟。配合全自动的视频标注管线，这套"仿真+视频"体系把单条数据成本压到 0.05 元，约为传统方案的 1/20 到 1/200，数据以每周数十万小时的速度扩张。2026 年初到 10 月，视频训练数据从 100 万涨到 400 万。现在业界的共识是，当一个任务有足够高效、全面的验证环境，往往这个任务就能被AI解决。RoboScience 的数据引擎为模型迭代提供了足够的动能。
模型层，核心是"学什么"的选择。 Visics 采用 VLOA（Vision-Language-Object-Action）架构，内部双引擎：上层世界模型用海量互联网视频预训练，负责理解任务、预演物体的 3D 运动轨迹；下层操作模型负责把轨迹翻译成具体机器人的控制信号。中间的接口是物体 3D 点云轨迹。
这个设计的精髓，可以用一个问题概括：教机器人抓杯子，是让它记"杯子该怎么动"，还是记"这台机器人的关节该怎么转"？前者是物理世界的通用语言，后者是特定硬件的方言。学前者的模型，天然具备跨本体能力，这正是邵林从 UniGrasp 到 D(R,O) Grasp 一路研究的产业化：一个模型驱动 LeapHand、Shadow Hand 等多种主流形态各异的灵巧手。
能力是否经得起检验，要看真实任务。
2025 年 5 月，RoboScience 基于 Visics 通用具身大模型 完成了复杂度、精度和步骤数都极高的拼家具任务，全程一镜到底。几个月前的 WAIC 2026 现场，AI 科技评论观察到，RoboScience 的展台演示的是多款不同品牌灵巧手跑同一套 Visics通用具身大模型，30 秒识别新硬件、无需微调，真正做到换手即用，观众随机递上的物品，模型全程自主抓取，做到真正的泛化性。
这些演示有展示的成分，但"随机物品"和"现场换手"两个设置，确实是对泛化能力的公开检验。
因此，这次大会上，RoboScience 机器科学和苏度，是两家留给我们印象最深刻的公司。
本体层，闭环的最后一环。 今年 8 月，RoboScience 机器科学发布的 REX G1 整合移动、感知与双臂操作。
自研本体对 RoboScience 的意义不是多卖一款硬件，而是让"仿真训模型、模型驱动本体、真实数据回流"的飞轮有载体可跑，形成真正的数据飞轮。
另一个值得注意的外部变量，是通用大模型的快速迭代。大模型向物理世界的渗透，比多数人预想的更快。GPT Astra 这类模型已经能理解空间关系、规划任务步骤，行业里也有一些团队开始直接用通用大模型充当机器人的"大脑"。
但如果你看过一些关于大模型来调动机械臂的直播或演示，就会发现成功率很低，这是落到真实场景，短板很快暴露。让一台机器人把桌上的杯子挪到左边，大模型可以把任务拆得很漂亮：找到杯子、规划路径、移动机械臂。可杯子是玻璃的还是纸的、该用多大力气、打滑了怎么办。接触、摩擦、弹性，这些物理属性不在文字和视频数据里，模型没见过，它也不知道。
这正是当下具身智能最难的一环：不是"想明白"，而是"做成功"。操作的泛化能力、执行速度、任务成功率，决定了机器人能不能走出演示视频。这些需要 AI 学习到真实世界的各种物理属性，对应的各种数据，需要有公司去采集和搭建相关能力。
RoboScience 机器科学的选择，是把这层能力自己做扎实。
它的 Visics 大模型和 RoboMirage 仿真引擎，用分层解耦的架构，加上高精度、高多样性的物理数据，训练出一个跨本体、跨物体、跨任务的统一操作模型。机器人先"看懂"物体怎么动，再通过同一个接口，驱动不同的身体完成任务，不管这个身体是什么样。
03
聊聊 RoboScience 的商业化思路
从某种意义上，我们认为 RoboScience 机器科学的商业化是软硬双线。
REX G1 轮式仿人形机器人面向具体场景做项目交付，同时由于 Visics 通用具身大模型采用的VLOA技术路线，独特的分层解耦的设计，使得了 Visics 成为了行业首个云端具身大模型，RoboScience 机器科学与腾讯云合作，共同推出了首个云端具身智能服务 EaaS（Embodied-AI-as-a-Service），用户像调用 API 一样获取操作能力。
两条路径一重一轻，一起来验证一件事：模型能力能否复用，而不是每个项目从头开发一遍。
同时，在商业模式设计 RoboScience 机器科学的商业思路设计了从云端 API 到整机交付的多元化商业图景，不难看出，RoboScience 的这套商业化思路覆盖了不同的群体，
离模型最近的一层，是云端 API。
面向本体厂商、集成商和开发者，Visics 被封装成极简的接口，随取随用，类似云服务商，按 API 调用或 License 收费。对一家做灵巧手或人形机器人的公司来说，最缺的从来不是硬件，而是让硬件"会干活"的大脑。现在它不需要自己训练模型，调一个 API 就能获得跨物体的操作能力。他们最近与腾讯云合作的 EaaS 在做的事，正是最轻、最容易起量的一层。
往下一层，是给第三方硬件装上"大脑"——即为第三方硬件提供即插即用的“通用操作大脑”。
面向物流、零售、酒店、泛工业自动化、设备商等场景的集成商，RoboScience 机器科学提供集成了模型、算力和控制接口的控制器，主打即插即用。它瞄准的是传统自动化啃不动的硬骨头：无序分拣、混合码垛，货箱随手堆在一起，大小不一、材质各异，传统方案需要逐一建模、逐一定制，而基于物体轨迹的 Visics 天然不管"抓的是什么"。这一层按控制器加 License 收费，支持私有化部署，实际上是在延展自动化设备厂商的能力边界：以前接不了的非标单子，现在能接了。
最重的一层，是自己下场，面向即时零售前置仓、商超、仓储分拣这类终端场景。
RoboScience 机器科学用自研本体 REX G1 做整机交付，感知、续航、移动、操作全部针对场景需求设计，配合通用末端执行器接口，整机销售加部署运维。
为什么要把最重的一层留给自己？
在我们看来，答案藏在真实场景的数据价值里。零售和物流场景 SKU 庞杂、环境多变，既是检验模型泛化能力的考场，也是积累真实物理交互数据的入口。
搭载 Visics 通用具身大模型的 REX G1，承担着打通这一闭环的角色。Visics 以物体的 3D 点云轨迹作为中间表征，降低操作能力对特定货品和机器人硬件的依赖，让模型面对未见过的 SKU 时，也具备跨本体、跨物体的操作经验。据了解，RoboScience 机器科学正与包括国内大型零售、供应链企业在内的客户推进多个 POC 。对其而言，这些项目既是商业化验证，也是模型持续迭代的数据入口：真实作业中的成功经验与失败案例，可以回流至训练与评测体系，推动模型能力提升。
由此看，自研本体的投入，有望转化为整个技术体系的复利。场景验证越深入，数据积累越充分；模型能力越强，API 与控制器的交付价值就越高。三层拼在一起，逻辑就清楚了：整机交付在最复杂的场景里打磨模型、沉淀数据；控制器把验证过的能力带进千行百业；API 则把它变成人人可调的共享能力，听起来是不是有点类似具身版的 OpenAI 。
04
一年三融，产业资本持续加注
最后，提醒大家关注本轮 RoboScience 的投资方构成，穿透这轮融资架构可发现，除财务机构外，还出现了琥珀资本、爽资本、 360 战投基金及多家上市公司平台，产业资本的身影非常显著。
产业资本离应用现场更近，问题也更具体，设备能跑多久，要不要人看守，新增一类商品要调试几天。他们的入场，可以理解为产业端对"能力可复用"这条路线的下注，但投资不等于订单，最终仍要现场说话。从结果来看， RoboScience机器科学 已经完成了产业资本的尽调考验。
回到融资本身：具身智能行业的故事通胀，本质上是因为多数团队缺乏可持续兑现的筹码。RoboScience 机器科学的连续融资，说明其在研究、工程、产品之间搭好了一条完整的链条。A2 轮之后，评判标准会更朴素：机器人进场后能不能稳定干活，换一种物品、换一台设备额外花多少钱，项目越多交付是否越轻。
在 2026 年下半年，具身能讲的融资新故事该讲的都讲完了，接下来，到了需要用硬核技术和真实的落地场景说话的新阶段了。（雷峰网）
雷峰网原创文章，未经授权禁止转载。详情见转载须知。
为了您的账户安全，请验证邮箱
请验证您的邮箱
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。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: Smart capital screening criteria are moving from "spoken" to "realized."
Author
Edit Peak
AI Science and Technology Review was informed that on 2 November, RoboScience machine science completed the A2 round of financing.
According to the information received, the round was co-financed by Amber Capital, Hung Rida Investment, Shanghai Semiconductor Investments, 360 war investments, discomfort capital, Union of Nations Living Capital, Kim Ding Capital and a number of listed company platforms and investments, with financing amounting to hundreds of millions of yuan.
This means that this self-employed company, which only formally became operational in March 2025, has completed five successive rounds of financing.
According to the AI Science and Technology Review, the market for finance for the recently established industry is not very promising. Over the past year, the narratives of financing this track have changed several rounds: the world model was the first to get the money; then the monograph technology was not enough, it had to go down, it had to go into the factory's ROI; and later, it had to have the imagination of "RSI."
Behind this story of inflation is the fact that most teams cannot afford the escalation of narratives, and every time the wind turns, a new narrative is needed. Capital then became picky, and the same BP, which was available last year, did not even meet this year. (Ray Peak Network (WeChat Official Account:Ray Peak Network))
But the rhythm of RoboScientific machine science is completely different, and the reverse is constantly favoured by capital and thus becomes a researchable sample company in the field.
Looking back at the time of the last five rounds of machine science financing, almost every round, there are verifiable product technology landings: the ICRA best thesis, the mirage platform, RoboMirage, the generic body LLM Visics, the cloud end-up with Tencent clouds with smart services, EAAS, the first wheeled generic robot REX G1 and the real project landing.
In a sense, this means that financing in a market should not follow the story of financing, which in turn should follow the technology.
According to AI, this is the most interesting signal of the current round of finance: smart capital screening criteria, moving from "showing" to "realising."
The financing tempo of a company, if matched by the technological tempo, suggests that it holds something that can be taken on a sustainable basis. According to AI Science and Technology Review, this is inextricably linked to the founding team.
01
Stanford Twin Star,
How to raise a Chinese body head company
In the Chinese industry, the academic history of the founders of the head company has shown several clear parties: the Qinghua, Berkeley, Stanford, etc.
The most profound of these is Stanford Circle.
Galactic General cranes, Dome Smart Lutzius, Hillbot Sow, RoboScientific Machine Science () have different paths with their founders, but most of them go up to the same lab, where a fellow from the U.S. Third College, Geometric Calculator Leonidas J. Guibas, is part of Stanford. In the same vein, Open AI researchers have been very effective as an assistant professor at Qinghua University.
It's not a coincidence.
Stanford is almost one of the sources of smart research around the world: Li Fei Fei, the initiator of ImageNet, the founder of the visual and data-driven paradigm, Real Lab of Chelsea Finn, and geometric calculator Leonidas J. Guibas, as well as the founding members of the foreign headmaster company Pi, came from Stanford.
Around 2016, the geometrical group of Prof. Stanford Leonidas J. Guibas (United States Member of the Third College) gathered a group of Chinese students, including Wang crane, Shaolin and their senior brother Soo-hoon.
Professor Guibas has a fairly consistent research line: 3D visual origin, focusing on generalization of the operation and trusting in abstract representation and simulation. The personal trajectory of the sorghum is almost the complete sample of the main line.
In 2014, Shaolin entered Stanford, where he started with a 3D visualization in Geometric Taekwondo, a member of the United States Third College, Leonidas J. Guibas, and then turned to robotics, with a teacher from Jeannette Bohg, who is currently Assistant Professor at the National University of Singapore.
On the academic level, a paper by Shaolin UniGrasp is a benchmark for data-driven smart-hand grabs; in 2025, the team, through the D(R, O) Grasp, won from over 4,000 contributions the award of the best thesis on the operational and movement direction of the ICRA robot, becoming the first-ever team in Asia in the first unit in nearly five years, breaking the long-standing European-American monopoly; and in 2026, Bi-Adapt returned to the nomination of the award, making it the only team in Asia that has been in that position for two consecutive years. The contestants were Berkeley, MIT, Stanford, Qinghua.
In addition, the Stanford Department has two other sub-lines, the Li Fly and U Nda. The Li Fei Fly line went out of Lutzego, and the U Nda line went out of the field.
In 2014, Fields entered Stanford AI Lab as the first major in the major physics department, and the division entered the same year as Shaolin from Stanford AI. He then spent seven years in Apple, making Apple AI Platform one of the youngest chief engineers at Apple Headquarters, leading the construction of the AI platform known as Apple and CUDA, and the first cloud around the world, in collaboration with the LLM reasoning system, which serves not laboratory test machines, but more than 1 billion users, 2 billion devices, including iPhone, Vision Pro and Apple Industries.
The core of the experience left to him was that the barrier to apples was never a single-point technology, but a soft and hard system capability.
According to AI, this experience has shaped a fundamental judgement in the field: it is not a single-point technological breakthrough, but a system that allows technology to work across equipment and across scenes.
The trees and fields are old acquaintances.
At the end of the year, one came out of the apple tube project, one came out from the academic world and two feeders synthesized the RoboScience machine science. In the Stanford system, this may be one of the most complete start-up companies with “academic depths plus engineering depths”: the route from Unigrasp to D(R,O) Grasps to Bi-Adapt corresponds to the model's cross-introvert capability; the field's experience of working with apples as a cloud ender, serving a billion-class equipment, corresponds to a system that can be deployed and iterative. The combination of the two is another formulation of the RoboScience technology road map.
02
Three major nodes with closed rings.
From publish RoboMirage Simulation Engine in September 2025 to publish Visions Universalized Body LLM in June 2026 and to publish REX G1 Wheeled Humanized Robot in August of August 2026, RoboScientity has spent about one year studying and understanding all three stages of "Data, Models, Body". This technology mix is not unique to RoboScience, but in all the companies that do the whole house, AI Science and Technology reviews that RoboScience's technology choices are the clearest.
Data layer, first solve the problem of affordability. The most intelligent bottlenecks are not algorithms, but data and the environment, where real telemetry is expensive and slow, and the industry is generally trapped in a cycle of "data sizes that cannot afford Scaling Law".
RoboMirage, starting with the bottom solver, has completed the coverage, efficiency and co-activity of the entire object space in the world, which has also reached an advanced level of industry. Hard soft material coupling, the contact power of very thin objects, and even distant exercises like "wash" pulling rubber bands are difficult to complete and can be simulated with high precision. With a fully automated video apostille line, this "simulation+video" system puts the cost of a single piece of data at 0.05 yuan, about 1/20 to 1/200 for the traditional scheme, which expands at hundreds of thousands of hours a week. From early 2026 to October, video training data increased from 1 million to 4 million. There is now an industry consensus that when a task is sufficiently efficient and comprehensive to validate it, it is often solved by AI. RoboScience ' s data engine provides sufficient kinetic energy for models to evolve.
Model level, the core is learning what. Visics uses the VLOA (Vision-Language-Object-Action) structure, with an internal double engine: the upper world model uses massive Internet video pre-training to understand missions, 3D motion tracks of rehearsing objects; the lower operating model is responsible for translating the tracks into control signals for specific robots. The middle interface is the object 3D point cloud trajectory.
The essence of this design can be summed up by a question: teaching robots how to grab a cup, how to move a cup, or how to turn a joint of this robot? The former is the common language of the physical world, while the latter is the dialect of specific hardware. Learning from the former model, with its natural cross-intellectual capability, is precisely the industrialization of sorghum forests from UniGrasp to D (R, O) Grasp: a model driven by a variety of dominant forms, such as Leaphand and Shadow Hand.
The ability to stand the test depends on real task.
In May 2025, RoboScientism, based on Visics' universal body LLM, completed a highly complex, precise and step-by-step furniture-sharing mission, with a full lens. On site WAIC 2026 a few months ago, AI, in a technical review, observed that RoboScientity demonstrated that multiple different brands run with the same smart hands, 30 seconds to identify new hardware, without fine-tuning, to actually change hands, to randomly hand over objects, to model self-capturing, and to be truly generalized.
These demonstrations have the ingredients to show, but the "random objects" and "replacement hand" settings are indeed open tests of the ability to generalize.
So, at this congress, RoboScientific Machine Science and Sordo is two of the most impressive companies we've ever had.
The inner layer, the last ring of the ring. In August this year, REX G1 of RoboScientific Machine Science publish integrated mobile, sensory and arms operations.
The meaning of self-study to RoboScientity is not to sell an extra piece of hardware, but to give a carrier for a "mutual training model, model driven body, real data flow back" to create a real data wheel.
Another external variable that deserves attention is the rapid iterativeization of General LLM. The penetration of LLM into the physical world is faster than most would have expected. Models such as GPT Astra already understand spatial relations and plan mission steps, and there are teams in the industry that have started to use General LLM directly as the robotic brain.
But if you've seen some live broadcasts or demonstrations about LLM to mobilize the arm of the machine, you'll find that success rate is very low, and this is the real scene, and the short plate is soon exposed. Let a robot move the cup on the table to the left, and LLM can tear it down pretty well: find the cup, plan the path, move the arm. But if the glass is glass or paper, how much energy should it be? Exposure, friction, elasticity, physical properties that are not in the text and video data, and the model has not been seen or known.
And that's the hardest part of being smart: not trying to understand, but being successful. The broad capability of the operation, the speed of implementation, the task success rate, determines whether the robot can get out of the demo video. This requires AI to learn about the physical attributes of the real world, the corresponding data, and the ability of companies to collect and build.
RoboScience's choice of machine science is to make this layer of power itself.
Its Visics LLM and RoboMirage Simulation Engines, with a layered decoupling structure, plus high precision, high diversity physical data, have trained a unified operational model across introverts, cross objects and missions. The robot "understands" how the object moves, then drives different bodies through the same interface, whatever the body may be.
03
Talk about the idea of commercializing RoboScience.
In a sense, we think that the commercialization of RoboScientific machine science is both soft and hard.
REX G1 wheel-style human robots are oriented towards specific scenes for project delivery, while the unique design of the decoupling of the VLOA technology used by Visics General with LLM has made Visics the industry ' s first cloud-like LLM, RoboScience Machine Science, in collaboration with Tencent Clouds, has launched the first cloud-based body-smart service, EAAS (Embodied-AI-as-a-Service), with users acquiring operational capabilities like API.
The two paths are one and the other, and one and the other, to verify whether model capabilities can be reused rather than being developed from scratch.
At the same time, in the business model design of RoboScientific Machine Science, a diversified business landscape was designed, from cloud-top API to whole machine delivery.
The nearest layer to the model is cloudy API.
Visics is sealed into a very simple interface for its own manufacturers, integrators and developers, and is available on demand, similar to a cloud service provider, at API call or License charge. For a company that makes smart hands or human robots, the least is never hardware, but the brain that makes hardware work. Now it doesn't need to train its own model, to move an API so that it can operate across objects. They recently worked with the Tencent cloud in the Eaas doing what is the lightest and easiest to start.
On the next floor, the third-party hardware is equipped with a "brain" — that is, a "general operating brain" for which third-party hardware is available immediately.
The RoboScientific Machine Science provides a controller for the integration of models, algorithms and control interfaces, which is immediately plugged in. It is aimed at hard bones that are traditionally automatically gnawed: unscrewd, mixed, cargo containers piled together, with different sizes and materials. Traditional solutions need to be modeled one by one, and Visics, based on the trajectory of objects, naturally doesn't matter what is caught. This layer, which supports the privatization deployment by adding a charge to the controller, is in fact extending the capacity boundary of the automated equipment manufacturer: the non-platform that was previously unavailable is now available.
The heavyest layer is its own going down and is oriented towards such terminal scenarios as instant retail forward warehousing, commercial supers and storage sorting.
RoboScientific Machine Science uses REX G1 as an integrator delivery, which detects, renews, moves, and operates in a way that is designed to meet the needs of the scene, in conjunction with the general end-of-pipe implementer interface, in which the whole machine is sold and deployed.
Why leave the heavyest layer to yourself?
In our view, the answer lies in the data value of the real scene. Retail and logistics scenes SKU are complex and environment-variable, both as a test site for modelability and as an entry point for the accumulation of real physical interactive data.
REX G1 with a common Visics body with LLM, has the role of breaking through this loop. Visics uses the 3D point cloud trajectory of the object as an intermediate representation, reducing the operational capacity to rely on specific goods and robotic hardware, and allowing models to have cross-physical, cross-object operational experience in the face of unknown SKU. RoboScientific Machine Science is known to be advancing multiple POCs with customers, including large domestic retail, supply chain enterprises. For them, these projects are both commercial validation and a continuous iterative data entry for models: Successes and failures in real operations can be channelled back to training and evaluation systems to promote model capabilities.
From this perspective, it is hoped that the self-study inputs will be translated into the compound benefits of the entire technology system. The more the scene is validated, the more the data accumulates; the stronger the modelling capacity, the higher the value of API and the controller delivery. The logic is clear when three layers are spelled together: the whole machine is delivered to grind models and sediment data in the most complex scenes; the controller takes proven capabilities into a thousand businesses; and the API turns it into a shared capability for all, which sounds a little bit like a body-style Open AI.
04
A year and a half, industrial capital continues to increase.
Finally, attention is drawn to the composition of the investors of the current round of RoboScience, through which the financial architecture reveals the emergence, in addition to financial institutions, of amber capital, pariah capital, 360 war capital funds and a number of publicly listed corporate platforms, with a remarkable presence of industrial capital.
Industrial capital is closer to the application site, and the question is more specific, as is how long the equipment will run, whether it will be guarded or not, and the new category of commodities will be tested for a few days. Their entry could be understood as the industry's bet on the line of "capability can be reused", but the investment does not amount to an order, and ultimately they have to speak on the spot. As a result, RoboScientific machine science has completed the perfect adjustment of industrial capital.
Back to finance itself: the story of smart industry inflation is essentially due to the lack of sustainable leverage for most teams. The continuous financing of RoboScience machine science indicates that it has a complete chain of research, engineering and products. After the A2 wheel, the criteria will be much simpler: whether robots can stabilize their work when they enter the field, and how much more it costs to change one item or equipment, and whether the more the project is delivered, the less it is.
In the second half of 2026, all the new stories on financing that could be told were finished, followed by a new phase requiring the use of hard nuclear technology and real landing scenes. (Ray Peak Network)
It's an original Ray Peak article, which is forbidden to reproduce without authorization. For further information, see the information below.
For your account security, please verify the mailbox.
Check your mailbox.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.