⌘K
Change language Switch ThemeSign In
Narrow Mode
20 亿砸向 00 后创业机器人公司！估值一年暴涨 7 倍，国家级资本重仓 =====================================
量 量子位 @思邈
One Sentence Summary
Lingchu Intelligence secures 2 billion RMB in funding, building a "human-centric" embodied AI data collection paradigm and a general-purpose manipulation brain through its self-developed Psi-SynEngine solution.
Summary
The article provides a detailed report on embodied AI startup Lingchu Intelligence completing a funding round totaling approximately 2 billion RMB, heavily backed by state-level and local government capital. Lingchu Intelligence posits that the deciding factor in embodied AI lies in the "data paradigm" rather than the robot hardware itself, and has released Psi-SynEngine, the world's first embodied native human data collection solution. This solution collects native human operational data via portable exoskeleton gloves at only 10% of the cost of teleoperation, featuring cross-embodiment migration capabilities. The company adheres to a "model-driven data" strategy, focusing on high-complexity, highly flexible task delivery in logistics scenarios, aiming to feed models with high-density problems from real-world environments to create an "embodied brain" with general manipulation capabilities.
Main Points
* 1. The core competition in embodied AI lies in the choice of data paradigm rather than robot hardware.Lingchu Intelligence believes traditional teleoperation and UMI (Universal Manipulation Interface) solutions suffer from "robot-centric" limitations, leading to tight coupling between data and hardware and restricted potential. By shifting to "human-centric" data collection, the system can learn the essence of tasks and achieve cross-platform portability. * 2. The self-developed Psi-SynEngine solution achieves low-cost, high-fidelity native human data collection.This solution utilizes tactile gloves to capture 21 degrees of freedom in human hands, with a total cost of only 10% of teleoperation methods. This hardware-agnostic data source allows models to learn the essence of human movement and migrate it to various dexterous hand configurations. * 3. Adhering to a "model-driven data" closed-loop logic to transform data into structured assets.The company does not act merely as a data provider; instead, it uses models to verify capability boundaries, which in turn defines the data architecture. This closed loop ensures that collected data aligns closely with model objectives, evolving into the core fuel that drives a "general-purpose manipulation brain." * 4. Deeply engaging in real-world logistics scenarios to build technical compounding by solving high-difficulty long-tail problems.Lingchu Intelligence has moved away from showy demos toward delivering high-complexity scenarios like "garment induction." High-density problems in real-world settings serve as hardcore nourishment for model evolution, creating a competitive advantage with strong exclusivity and compounding effects.
Metadata
AI Score
82
Website qbitai.com
Published At Today
Length 3323 words (about 14 min)
Sign in to use highlight and note-taking features for a better reading experience. Sign in now
> 允中 发自 凹非寺 > > > 量子位 | 公众号 QbitAI
具身智能的胜负手，可能并不在机器人本身。
这种认知差，正在被最敏锐的资本迅速兑现为筹码。
近日，灵初智能宣布完成总额约20亿元的天使轮及Pre-A轮融资。
* 天使轮由国开金融、国中资本、央视融媒体产业投资基金等国家级“国家队”资本，某数千亿上市公司旗下战投、长飞光纤旗下基金，两大核心产业龙头资本，沃德尔等知名产业资本，及元生创投、珠海科技产业集团、钧山投资、燕缘创投、大米资本、沃赋资本、彬复资本、泰合资本等多家知名基金共同投资； * Pre-A轮由上海国资徐汇资本等基金领投，梁溪科创产业二期母基金（博华资本管理）、锡创投等地方国资，及普丰资本、钛铭资本等市场化基金跟投，多家老股东实现超额跟投。华兴资本担任长期财务顾问。
这笔20亿元的资金，将加速灵初智能在物流场景的规模化落地与数据采集体系建设。
值得注意的是，这是灵初智能首次系统披露公司的融资进展。
过去很长一段时间，这家公司几乎没有在资本层面发声，而是把大部分精力放在技术路线与数据体系的打磨上。
而之所以能吸引国家级资本与地方国资的大规模重仓，也与团队极其互补的底色分不开：
创始人兼CEO王启斌在手机、智能音箱及机器人领域拥有20年产品操盘经验，曾担任黑莓、Sonos及云迹科技高管；
联合创始人陈源培是00后，在北大人工智能研究院读研时师从强化学习（RL）代表人物杨耀东，在斯坦福曾与李飞飞有过深入交流，曾拒绝了华为“天才少年”的高薪offer。
在具身智能公司纷纷卷Demo、卷参数的当下，灵初智能的突围逻辑清晰且冷峻：放弃昂贵且低效的机器人遥操，All in人类原生数据。
长期以来，具身智能被三座数据大山死死压住：
仿真环境数采存在无法逾越的Sim-to-Real差距，尤其在处理布料等柔性物体时捉襟见肘；
机器人遥操数采则像是一场昂贵的人力外包，碎片化的试点导致成本居高不下，且无法穷尽物理世界的复杂分布；
而硬件本体与数据的深度耦合则成为了“房间里的大象”。数据与硬件强绑定，采集于哪种本体，便服务于哪家体系。这种封闭结构使数据难以跨平台流通，整个生态将逐渐演变为彼此割裂的孤岛。
风头正盛的UMI设备数采，在灵初智能看来更像是一个“美丽的陷阱”，同样存在结构性问题。
在灵初智能看来，这并非一场工具之争，而是一场数据范式之争。
UMI本质上是让人拿着机器人的器官去模拟机器，是一种“Robot-Centric”的逻辑。短期内可以降低门槛，长期却可能锁死上限。
如果数据从一开始就围绕机器人本体采集，那么模型的能力边界也会被提前锁定；而如果数据源头来自人类本身，那么模型学习的将是“任务本质”而非“机器结构”。两种路径，决定的是具身智能未来的天花板。
“UMI采集的是机器人夹爪的数据。它无法泛化，今天用UMI采的数据，根本无法直接用到五指灵巧手上。”陈源培指出，这种方案强行将人类拥有20多个自由度的五指灵巧手，降维成了一个只能“开合”的简易夹爪。
为了破解这一冷启动难题，灵初智能全栈自研并发布了全球首个具身原生人类数据采集方案——Psi-SynEngine。
其核心逻辑只有四个字：以人为本。
便携式外骨骼触觉手套可精准捕捉人手21个关节自由度及全手触觉信息，并不影响工人正常作业；系统同步记录头戴与手部视角的视觉、触觉、动作及语言数据，为预训练阶段的多模态对齐提供真值支撑。
更关键的是成本结构。
据王启斌透露，通过手套采集数据的综合成本，仅为真机遥操方案的10%。
真正的护城河，则来自跨本体迁移能力。
“机器人会迭代，夹爪会更换，但人手是不变的。”陈源培表示。
通过基于世界模型与强化学习的迁移算法，灵初能够将人类动作高质量映射到不同构型的灵巧手上，弥合Embodiment Gap。
当数据源头脱离硬件本体，模型的能力上限也会随之被重新打开。
数据基建解决的是“矿从哪来”，即“有没有数据”的问题，而真正拉开公司之间差距的，是“矿炼成什么”，也就是把这些数据转化为模型能力的效率。
灵初发布数采体系，外人看可能觉得就是个“卖铲子”的生意。但在创始团队的逻辑里，这只是飞轮的起点。
陈源培直言：
> 我们不会停留在数据供应商这个角色。
数据是用来训练具身大脑的燃料，而非终点。真正具备长期价值的，是由数据喂养出来的、可迁移的通用操作能力。
灵初智能不卖矿石，也不卖铲子，卖的是“会干活的脑”。
在灵初的逻辑里，单纯靠人力做数采没什么门槛，本质上赚的是劳动力的钱。而模型对数据的消化、泛化与迁移水平，才直接划定了这些数据价值的天花板。
因此，灵初走出了一条在行业中相对少见的路径：
* 先通过模型验证能力边界，再反过来定义数据体系； * 先通过模型训练与任务实验，判断哪些数据真正有价值，再围绕这些关键数据构建规模化采集能力。
这种“模型驱动数据”的闭环，让灵初在持续推进模型落地的过程中，能不断修正数据结构、标注体系与采集方式。
这就把原本死板堆积的“原材料”给带活了，让它们变成了一种紧贴模型目标、不断进化的结构化资产。
相比行业里还在兜售“大物流”“全场景泛化”这类宏大叙事，灵初表现得有些反直觉的克制。
王启斌透露，2025年下半年灵初内部曾有过一次关键掉头：停止资源投入纯展示型Demo，全面转向真实数据采集与细分场景交付。
为什么？
这种转变背后的逻辑其实很务实，因为模型进化需要养料，而最硬核的养料，只有在真实交付中碰撞出的“高密度问题”里才能淘出来。
灵初切入的口子极细，甚至有些“挑剔”，比如专门盯着“衣服供包”或“入箱检”这种高复杂度、强柔性的活儿。
以衣服供包为例，目前灵初已实现对上千件衣物的泛化抓取，节拍提升至800 UPH（即Unit Per Hour，指每小时产出数量），形成从场景部署、数据采集到模型优化的闭环。
这套方法在团队内部被沉淀为一种“能力飞轮”：每一个新场景的落地都在喂养模型，而变强了的模型，又成了他们敲开下一个复杂场景的敲门砖。
飞轮由此启动。
在具身智能这个“软硬耦合”的赛道上，全栈几乎是必选题。
算法离不开硬件，硬件又反过来塑造数据分布，两者缺了谁都跑不通系统闭环。
但灵初的“全栈”被赋予了某种分寸感，他们通过战略筛选，将精力高度克制地集中在核心链路，走的是一条精准布局、有的放矢的路子。
在王启斌看来，市场上能买到且够用的，灵初绝不碰；但凡是卡住核心能力的环节，必须紧紧攥在自己手里。
之所以费力气自研数据手套和灵巧手，是因为市面上现有的方案在规模化数采和底层电流环控制算法上，根本达不到高精度操作的要求。
假如将这些环节外包出去，无异于把数据质量和模型演进节奏的“命门”交给了别人。
而对于像轮式底盘这样已经高度成熟的赛道，他们则选择通过定制合作来解决，因为那已经不构成技术瓶颈，强行投入只会分散自身精力。
如此取舍背后，其实是在重新厘清灵初这家公司的能力边界：自研是为了守住核心能力，整合是为了调用通用资源。
这也让灵初的定位变得愈发清晰——他们本质上是一家通用灵巧操作能力的“大脑驱动公司”，核心算法与数据链路自控，硬件形态保持开放，能够根据不同场景灵活适配。
对于具身智能公司来说，眼下跟时间赛跑很重要。
因为真实场景的数据反馈是有复利效应的，那些进场早、碰到的复杂任务多的人，才能抢先一步触达那些决定胜负的长尾数据。
这种由规模和密度堆叠出来的领先优势，是后来者很难用资本直接抹平的。
所以灵初的方法论重心在于模型能力的沉淀。随着数据飞轮的启动，数据成本在降，模型能力在升，这种此消彼长会带他们步入一个更高阶的战场—— 去解决那些更复杂、更广阔的应用场景。
值得注意的是，灵初智能此次披露的融资信息也释放出一个信号：具身智能正在进入资本与产业共振阶段。
据行业人士估算，在过去一年中，灵初估值已提升约6–7倍，正在向具身智能领域的独角兽迈进。
从国家级产业基金、地方国资平台，再到通信与光通信产业链龙头，这种资本结构背后，其实是产业界对“具身数据基础设施”的提前押注。
在这个赛道里，大家争夺的筹码早已超出了资金本身，时间才是最稀缺的通货。
随着数据飞轮转速加快，比拼的本质也从聚焦于谁账面上钱多，转变成了谁跑在了时间的最前面。
而这种领先优势极具排他性，一旦时间差形成，差距就会以一种极快的速度被放大，留给后来者的空间只会越来越窄。
量 量子位 @思邈
One Sentence Summary
Lingchu Intelligence secures 2 billion RMB in funding, building a "human-centric" embodied AI data collection paradigm and a general-purpose manipulation brain through its self-developed Psi-SynEngine solution.
Summary
The article provides a detailed report on embodied AI startup Lingchu Intelligence completing a funding round totaling approximately 2 billion RMB, heavily backed by state-level and local government capital. Lingchu Intelligence posits that the deciding factor in embodied AI lies in the "data paradigm" rather than the robot hardware itself, and has released Psi-SynEngine, the world's first embodied native human data collection solution. This solution collects native human operational data via portable exoskeleton gloves at only 10% of the cost of teleoperation, featuring cross-embodiment migration capabilities. The company adheres to a "model-driven data" strategy, focusing on high-complexity, highly flexible task delivery in logistics scenarios, aiming to feed models with high-density problems from real-world environments to create an "embodied brain" with general manipulation capabilities.
Main Points
* 1. The core competition in embodied AI lies in the choice of data paradigm rather than robot hardware.
Lingchu Intelligence believes traditional teleoperation and UMI (Universal Manipulation Interface) solutions suffer from "robot-centric" limitations, leading to tight coupling between data and hardware and restricted potential. By shifting to "human-centric" data collection, the system can learn the essence of tasks and achieve cross-platform portability.
* 2. The self-developed Psi-SynEngine solution achieves low-cost, high-fidelity native human data collection.
This solution utilizes tactile gloves to capture 21 degrees of freedom in human hands, with a total cost of only 10% of teleoperation methods. This hardware-agnostic data source allows models to learn the essence of human movement and migrate it to various dexterous hand configurations.
* 3. Adhering to a "model-driven data" closed-loop logic to transform data into structured assets.
The company does not act merely as a data provider; instead, it uses models to verify capability boundaries, which in turn defines the data architecture. This closed loop ensures that collected data aligns closely with model objectives, evolving into the core fuel that drives a "general-purpose manipulation brain."
* 4. Deeply engaging in real-world logistics scenarios to build technical compounding by solving high-difficulty long-tail problems.
Lingchu Intelligence has moved away from showy demos toward delivering high-complexity scenarios like "garment induction." High-density problems in real-world settings serve as hardcore nourishment for model evolution, creating a competitive advantage with strong exclusivity and compounding effects.
Key Quotes
* The deciding factor in embodied AI may not lie in the robot itself. * UMI is essentially having humans use robot 'organs' to simulate machines... it may lower the barrier in the short term, but it could lock the ceiling in the long run. * Robots will iterate and grippers will be replaced, but the human hand remains constant. * Lingchu Intelligence sells neither the ore nor the shovels; it sells the 'brain that knows how to work'. * Lingchu never touches what is readily available and sufficient on the market; however, any link that bottlenecks core capabilities must be held firmly in our own hands.
AI Score
82
Website qbitai.com
Published At Today
Length 3323 words (about 14 min)
Tags
Embodied AI
Robotics
Data Collection
Reinforcement Learning
Funding News
Related Articles
* Faith and Breakthroughs: 2026 Artificial Intelligence Trend Outlook * After Topping Open-Source Rankings with its Programming LLM, the Zhipu GLM Team Faced a 3-Hour Questioning Session * GPT-5.4 Released: OpenAI's First Unified Model, Truly Native * MiniMax Hailuo Video Team's First Open-Source Release: Tokenizer Exhibits a Clear Scaling Law * AI Starts to "Take Action", Alibaba's Qwen Leads the World * MiniMax M2.5 Released: $1/Hour, the King of Real-World Work through a native Agent RL framework.") * Yao Shunyu Lectures Face-to-Face with Tang Jie, Yang Zhilin, and Lin Junyang! Four Schema Heroes Debate Heroes at Zhongguancun * Year-end AI Reflections by Salvatore Sanfilippo, the Creator of Redis * Deconstructing Agentic Coding from First Principles: From Theory to Practice * After Reading 200 Papers: Analyzing DeepMind, Meta, DeepSeek - What AGI Narratives Are Chinese and American Giants Describing? | 2025 AI Annual Review HomeArticlesPodcastsVideosTweets
2 billion RMB Invested in Gen-Z Robotics Startup! Valuati... ===============