China's top embodied intelligence companies are engaged in a battle of titans—
Investment community learned that today (July 6), StarDynamics completed a new round of 1 billion yuan financing. This round was led by the State-owned Assets Supervision and Administration Commission's (SASAC) state-owned capital operation company, Chengtong Fund. Other major state-owned capital participants included Jiangxi State-owned Capital Operation Holding Group, Guoyuan Equity, Yufu Zhongxin Fund, Hangzhou Capital, and others. Additionally, CICC Renault, Jiukun Venture Capital, Hony Capital, Juntai Capital, and Shenghe Capital followed on; existing shareholders Houxue Capital, Tsinghua Unigroup Tiancheng, and Qianshan Capital continued to invest.
With this, StarDynamics has raised 2.5 billion yuan in intensive financing over two months, constructing a triple capital matrix of "National-Level Capital Leadership + Top-Tier Financial Empowerment + Industrial Ecosystem Synergy," which assembles the industry's most extensive industrial capital injections, totaling over 20 entities.
In past communications with investors, StarDynamics left a deep impression on us: it is Tsinghua University's only equity-held embodied company, one of the earliest to propose world models in embodied AI, and despite being only three years old, it has already achieved scaled deployment in logistics scenarios. China's embodied intelligence landscape is filled with various schools, yet StarDynamics has quietly broken through.
Tsinghua's Direct Robotics Lineage
2.5 Billion in Financing in Two Months
The outside world may not know that StarDynamics is the proponent of the world model concept in embodied intelligence.
Rewind to 2024, VLA (Vision-Language-Action) was still the industry mainstream, but StarDynamics had already begun exploring research on world models. Not only did StarDynamics first propose the "world model" approach, it also released its world model achievement earlier than most globally—the PAD model released in September 2024 was the world's first World Action Model (WAM), integrating video prediction and action prediction, released nearly a year before similar solutions from NVIDIA (like DreamZero).
At the time, this was a pioneering, "minority" voice, somewhat esoteric.
It wasn't until the second half of 2025 that discussions about world models began frequently appearing in the embodied field. In October that year, StarDynamics collaborated with the Stanford University team led by Chelsea Finn to introduce Ctrl-World, using world models as data simulators to generate training data approximating real-world physics, improving performance by 45% over Pi0.5. Subsequently, we witnessed a global explosion of world models, with the entire industry converging towards this direction.
Being able to move slightly ahead of the industry is inseparable from the unique academic background of its helmsman, Chen Jianyu.
Born in 1992, he was admitted to Tsinghua University's Department of Precision Instruments in 2011—one of China's earliest institutions engaged in bipedal humanoid robotics research. During his undergraduate studies, Chen Jianyu began researching gait planning for bipedal robots. He later went to the University of California, Berkeley, for a direct Ph.D., studying under Professor Masayoshi Tomizuka, a member of the U.S. National Academy of Engineering and a pioneer in mechatronic control, delving deeply into robot reinforcement learning and motion control algorithms.
After earning his Ph.D. in 2020, upon invitation from Turing Award laureate and Chinese Academy of Sciences academician Yao Qizhi, Chen Jianyu returned to China and joined Tsinghua University's Institute for Interdisciplinary Information Sciences, serving as an assistant professor. At 28, he became one of the youngest doctoral supervisors at Tsinghua at the time.
These academic and research experiences made Chen Jianyu a rare talent in the industry with both hardware + brain capabilities, leading to very forward-thinking and keen choices in technical roadmaps and model iterations. In August 2023, leveraging Tsinghua's technology transfer policies, Beijing StarDynamics was officially established. It is the only embodied intelligence enterprise with direct equity held by Tsinghua University, with Chen Jianyu as its founder.
His core judgment is: to achieve true general intelligence, robots must simultaneously possess a smart "brain" and a dexterous "body"—a robot without a brain easily becomes scrap metal, and a brain without a body can hardly be called a robot.
Therefore, from day one, StarDynamics has been driven by AI Native principles, constructing the industry's only full-stack self-developed barrier encompassing "data-brain-motion control-dexterous hand-body." The core logic is "algorithm requirements first, data value first," starting from the actual needs of brain model training and deployment, and reverse-defining hardware design. To date, it possesses three major product lines: the full-size bipedal humanoid robot Star-L7, the wheeled humanoid service robot Star-Q5, and the full direct-drive five-finger dexterous hand Star-XHAND series.
In less than three years, a long queue of investors has lined up behind Chen Jianyu. From early backers like Tsinghua University and Alibaba, to top-tier financial institutions like CDH VGC, Sequoia Capital China, IDG Capital, Qingliu Capital, CICC Capital, to industry giants like SF Express, Samsung, Geely Capital, Haier, Lenovo, Singtel, BAIC Capital, Dongfeng Capital, CICC Porsche, CICC Renault, and funds under China Unicom, as well as local state-owned capital like the Beijing Artificial Intelligence Industry Investment Fund—StarDynamics' shareholder list almost covers the most complete capital spectrum in the embodied intelligence sector.
Entering 2026, the financing pace accelerated dramatically: In March, it completed a 1 billion yuan strategic financing round, with valuation breaking through the 10 billion yuan mark. South Korea's Samsung and Singapore's Singtel appeared among its shareholders as overseas industrial capital for the first time. Just a month later, $200 million in financing arrived, led by SF Express, with leading institutions like Sequoia Capital China and IDG Capital jointly following on, while the industrial investor lineup expanded simultaneously. And today, with national-level state-owned capital platforms like Chengtong Fund collectively joining, a new 1 billion yuan financing round was announced. Securing three consecutive large-scale financings within three months, it set one of the fastest financing speeds in the embodied sector this year.
Thus, StarDynamics has constructed a triple capital matrix of "National-Level Capital Leadership + Top-Tier Financial Empowerment + Industrial Ecosystem Synergy," with over 20 industrial capital entities assembled behind it—a lineup not commonly seen in the industry.
Strengthening the Brain Through the Hand, Forging a New Path in Embodiment
At this juncture, for embodied intelligence to deliver industrial value, the competition is no longer about individual strengths, but the systemic capability integrating software and hardware—the brain (model and data) determines the upper limit of value, while the body (whole machine and end-effectors) constrains the lower limit of capability.
But within this system, there is a core hub often overlooked: the dexterous hand.
StarDynamics adheres to the AI Native principle of full-stack self-development across the entire embodied chain of "data-brain-motion control-dexterous hand-body," meaning hardware design does not start from mechanical structures but is reverse-defined from "what kind of data is most valuable for the brain." The five-finger dexterous hand is precisely the richest and most precise data collection entry point for physical world interaction—whether a grasp is successful, how force feedback changes, if an object slips, these high-dimensional data points directly determine what the model can learn.
Over the past two years, StarDynamics' models have continuously evolved:
In the first half of 2024, the team first proposed the robot fast-slow system VLA architecture (frequency-separated VLA), achieving the unity of "acting in real-time" and "deep thinking"; in the second half, they combined world models with VLA, releasing the VLA algorithm framework PAD and VPP integrating world models, and consolidating them to launch the end-to-end native robot brain model ERA-42.
ERA-42 is one of StarDynamics' core models, integrating vision, understanding, prediction, and action into one, enabling control of whole-body dexterous operations through a single end-to-end VLA model.
In February this year, building upon Ctrl-World, StarDynamics launched VLAW, pioneering the world's first VLA reinforcement learning framework based on world models, achieving synergistic iteration of policy and simulator, making world models not only "look right" but also "be physically correct." To date, StarDynamics is one of the embodied companies with the most world model achievements.
Breakthroughs in brain capability depend on high-quality data. On this front, relying on its self-developed dexterous hand and deployment applications, StarDynamics possesses the world's largest dataset from real dexterous hand hardware.
Simply put, StarDynamics uses dexterous hand hardware to obtain high-quality dexterous manipulation data, then uses this data to train smarter "brain" models, and finally feeds the evolved models back to the dexterous hands, forming a virtuous cycle of "the more used, the smarter; the smarter, the more capable." Therefore, the "dexterous hand" business is not merely a single-point component on the robot body but a core hub connecting the entire full-stack capability, the central entry point for collecting physical world interaction data.
Having long anticipated that end-effectors would become a bottleneck for embodied intelligence deployment, StarDynamics' dexterous hand adopted a brain-friendly, pioneering full direct-drive technical route—joint module output directly acts on joints, with no transmission backlash, elasticity, or friction loss. The generated data inherently possesses high precision, low latency, and reproducibility, and can be directly used for model training, fundamentally solving the quality issue of physical interaction data.
Currently, StarDynamics employs a "dual-hand strategy" with two differently positioned dexterous hand products:
1. Star-XHAND 1 PRO (Brain Hand), emphasizing high performance, core positioning as a data collection and algorithm validation platform, anchoring the upper limit of model capability;
2. Star-XHAND 1 (Work Hand), achieving precise force control and flexible operation through a full direct-drive technical architecture, adaptable to various humanoid robot platforms, solidifying the lower limit for scaled deployment.
It is reported that Star-XHAND 1 is now widely adopted to meet diverse needs in scenarios like industrial sorting and routine operations, becoming a shared choice for global robot manufacturers—American embodied intelligence unicorn Skild AI, South Korea's Rainbow Robotics, UK's Extend Robotics and Discover Robotics, as well as the new-generation humanoid robot HMND 01 from UK's Humanoid AI, among others, have all adopted Star-XHAND 1 as a core end-effector.
When the valuation of a single dexterous hand within the industry can even exceed that of the entire robot body, StarDynamics' narrative of "strengthening the brain through the hand" gains more substance. But ultimately, "strengthening the brain through the hand" is not the goal, but a means—obtaining high-quality data through the hand, training a stronger brain through the data, driving more intelligent robots through the brain, ultimately running through the full-stack system capability of "data-brain-motion control-dexterous hand-body," becoming real productivity in industrial scenarios—that is StarDynamics' true moat.
Relying on its self-developed dexterous hand and deployment applications, StarDynamics has accumulated a leading-scale dataset from real dexterous hand hardware within the industry. Based on this, the company has built a triple-gradient data source system:
Core Value Layer: Long-duration real hardware interaction data from real scenarios like logistics and industry, with 100% physical authenticity, forming the core foundation for models to align with industrial needs;
Precision Training Layer: High-precision teleoperation data, providing standardized action reference paradigms; currently has over 12 million clips of real hardware teleoperation data, including over 1.5 million clips of dexterous hand real hardware teleoperation data, one of the largest datasets of its kind in the industry;
Breadth Expansion Layer: First-person human behavior data and large-scale internet video data, enabling million-hour scaling, cost-effectively covering vast daily behaviors and scenarios.
Among these, long-duration real hardware interaction data from real scenarios and large-scale human video data simultaneously satisfy authenticity and diversity, constituting the dual core engines of the data system.
Currently, StarDynamics' overall dataset covers 100+ real scenarios and 1000+ dexterous manipulation tasks, sufficiently ensuring scene richness and behavioral diversity, laying a high-quality data foundation for the continuous iteration of the general embodied brain.
On this foundation, StarDynamics begins to answer the next question: Can this data support the brain in achieving commercial closed loops in real industrial scenarios?
Robots Begin Stable Deployment
The Watershed Moment
The answer lies on the industrial front lines where robots can stably deploy and work continuously.
StarDynamics adheres to a "B2B first, then B2C" commercialization path, gradually deploying in logistics, high-end manufacturing, and commercial service scenarios.
As we have seen, StarDynamics first achieved the industry's first PMF (Product-Market Fit) in logistics scenarios, deeply cooperating with leading clients like SF Express and China Post, batch-deploying to over 10 logistics centers across 5 provinces/municipalities in North, East, and South China, with some achieving normal 7×24 hour operations.
Here, embodied robots can flexibly handle various parcel sorting tasks involving packages of different shapes, materials, and sizes—grasping, turning parcel labels face-up, placing, etc. It's reported that efficiency in some scenarios even surpasses human levels, with processing speeds reaching over 1200 pieces per hour.
There's an interesting anecdote here—when the overseas embodied leader Figure conducted a live stream from its testing room, it was called out by foreign media: "StarDynamics' robots in China are already working at SF Express and China Post. When will Figure step out of the testing room and truly deploy?"
It is revealed that in the future, StarDynamics will continue to expand logistics scenarios, building a full-process embodied logistics service system covering inbound logistics, in-plant logistics, sales logistics, after-sales logistics, and extended express logistics links.
Meanwhile, high-end manufacturing and commercial service scenarios are also being rolled out. In high-end manufacturing for 3C electronics and automotive, StarDynamics has partnered with industry leaders like Samsung, Lenovo, Haier, and Geely. The real hardware data flowing back from these industrial sites continuously provides iterative nourishment for the embodied brain. Additionally, the Star-Q5 super-realistic service robot has been deployed in scenarios like Haier, Lenovo, and Century Golden Resources, providing services like store attraction, guided tours, explanation, and product delivery, exploring B2C entry points.
The underlying body capabilities supporting these deployments are also being validated by global top-tier institutions. StarDynamics' dexterous hands, universal body base, and development kits are not only used internally but have also been shipped globally, serving 9 out of the world's top 10 tech companies by market capitalization, becoming the shared choice of top tech firms and research institutions including OpenAI, Boston Dynamics, NVIDIA, Apple, Google, Amazon, ByteDance, as well as MIT, UC Berkeley, Stanford University, Tsinghua University, and Shanghai Qizhi Research Institute. Feedback from top-tier research clients further drives continuous upgrades to StarDynamics' hardware base.
As robots begin stable deployment, the prelude to the era of embodied intelligence leaders opens. As more robots enter the real world to perform tasks, those who can first establish data moats and evolution barriers for the next-generation robot brain that others find difficult to replicate will occupy key positions in the new round of competition.
As widely believed, 2026 is a watershed year for embodied models: in the first half, model capabilities diverge; in the second half, commercialization closed loops diverge. Capital markets no longer pay for stories, only recognizing one thing—whether technology can deploy, whether deployment can scale, and whether scale can feed back into technology.
The watershed moment has emerged—not under the spotlight of the laboratory, but on the logistics sorting lines at 3 a.m.
This article is from the WeChat public account "Investment Community" (ID: pedaily2012), author: Yang Jiyun
