Zhiyuan spins off Mifeng, using bodyless data collection devices to break through the physical AI data bottleneck, with a million hours of data driving the industrialization of embodied intelligence infrastructure.
When Shanghai Mifeng Embodied Intelligence Technology Co., Ltd., incubated internally by AgiBot, officially delivered its 20,000th MEgo bodyless data collection device to JD.com, and simultaneously announced that its cumulative bodyless data had surpassed 1 million hours and that it had entered into a deep business partnership with Tencent Robotics X, the shock felt across the entire embodied intelligence and robotics sector went far beyond the mass-production milestone of a single emerging hardware startup.
For a long time, capital market enthusiasm for humanoid robots has been highly focused on dexterous hand degrees of freedom, motor torque density, and flashy demonstrations of motion control algorithms. But behind the bustling demos, the entire industry is collectively hitting an invisible wall of ice—the extreme scarcity of high-value action-interaction data from the physical world. Mifeng Technology's emergence as an independent entity and its large-scale delivery fulfillment tear open exactly the strategic inflection point at which embodied intelligence moves from the pioneering phase of blindly stacking hardware bodies toward fully decoupling data production into independent industrial infrastructure.
The reason large language models were able to achieve emergent intelligence over the past two years is that the internet had accumulated decades of high-quality token text and image data measured in the trillions. But when robots move into the physical world, they face gravity, friction, flexible deformation, and unstructured obstacles. These physical interactions cannot be described by text alone and must rely on trajectories accurate to the millimeter, multimodal clock synchronization, and force-feedback data.
In the past, the industry generally used expensive humanoid robot bodies for teleoperation data collection. A single device often cost hundreds of thousands of yuan, not only because the mechanical structure was highly prone to wear and damage, but also because the collection scenarios were strictly locked to flat laboratories and exhibition halls. This high-cost, low-efficiency workshop-style data collection directly resulted in most robotics companies having data pools of only hundreds or thousands of hours, nowhere near enough to satisfy the data throughput appetite needed to train the next generation of general-purpose physical-world large models.
Unpacking the Business Landscape: The Strategic Gambit of an Independent Entity
Tracing through the underlying business map of this new track reveals deep traces in corporate records of the operating team's strategic intent to spin the core business off from the parent body. Tianyancha corporate data shows that the operating entity, Shanghai Mifeng Embodied Intelligence Technology Co., Ltd., was established in February 2026, with Yao Maoqing as legal representative, registered capital of approximately 5.2875 million yuan, and a location in Shanghai's Pudong New Area, a bridgehead of technological innovation.
In the equity chain traced by Tianyancha, the incubator AgiBot Innovation (Shanghai) Technology Co., Ltd. was established in February 2023, with former Huawei executive Deng Taihua as legal representative. Its shareholder list is firmly controlled by key partnership platforms including Sangpeng Technology and Guanbao Technology, which hold 28.07% and 10.22% respectively.
Within just half a year of its founding, Mifeng not only completed hundreds of millions of yuan in massive financing, but its shareholder and strategic investment lineup also assembled top state-owned and industrial capital including China Telecom, Zhangjiang Group, Sequoia China, Futeng Capital, CDH Investments, and Baidu Ventures.
The parent body AgiBot's choice to spin off the data collection business at this moment is an extremely seasoned commercial gambit.
If the data collection and governance functions had remained attached to the AgiBot parent body, then due to the natural whole-machine competitive vigilance among humanoid robot manufacturers, Tencent, JD.com, and even other competitors would never procure critical training data from a competitor, still less easily open up their own real logistics and warehousing scenarios. Establishing Mifeng independently allows it to enter the market as a neutral physical AI data infrastructure service provider, completely dissolving the ecological wariness of downstream hardware manufacturers and scenario owners.
The fact that JD.com is willing to accept 20,000 devices into its real business warehouses and that Tencent Robotics X is willing to sit at the procurement negotiation table is precisely the enormous commercial dividend released by this independent third-party positioning.
Bodyless Collection: A Dimensional Deconstruction of Traditional Data Supply Forms
The deeper disruption lies in the dimensional deconstruction of traditional data supply forms by bodyless collection devices.
The MEgo series that Mifeng is betting on is essentially a matrix of exoskeletons or wearable sensors that completely strips away heavy lower limbs and intricate joint servo motors. It returns the subject of data collection to frontline human workers, letting humans serve as the natural driving source in real labor such as warehouse sorting, shelf stocking, and fine assembly. This design, with a hardware threshold of thousands to tens of thousands of yuan, directly replaces the previous teleoperation bodies costing hundreds of thousands of yuan, causing collection costs to plummet exponentially, making it possible to rapidly roll out a massive scale of 20,000 units and quickly accumulate 1 million hours of physical data spanning 22 major categories of scenarios and 50,000 types of objects in open environments.
The Ultimate Question for the Business Model: Data Depth and the Algorithmic Bridge
However, as data supply moves from the laboratory to a large-scale pipeline, the ultimate question for the business model has only just begun.
Although 1 million hours of physical data is already an unprecedented magnitude for the industry, compared with the accumulation of hundreds of millions of kilometers of mileage data in autonomous driving, the data depth of embodied intelligence is still at an early stage. More critically, what bodyless collection obtains is kinematic data of human skeletons and end-effector movements. How to map these human dynamics parameters losslessly onto various bipedal, quadrupedal, or wheeled-legged robot bodies with different degrees of freedom and different motor response characteristics still stands as a huge algorithmic bridge.
If the data governance platform cannot achieve near-physical-level high-fidelity fitting in data retargeting and motion overlap, then the massive flood of data may degenerate into an inefficient and redundant digital burden.
Conclusion: Seizing the Industry Chain's Most Core Chokepoint
This data handover in Shanghai's Pudong sends the coldest signal to the entire hard-tech sector: the bubble phase of embodied intelligence is over, and the era of raising money from the primary market simply by building a few robot shells that can do flips or serve coffee is gone forever. When industrial capital and scenario giants begin to converge and pour real money into the infrastructure of data collection and annotation, whoever can efficiently convert every atomic interaction in the physical world into standard fuel that computing power can digest will be the one to seize the most core chokepoint of the entire industry chain on the eve of general-purpose robots truly moving toward large-scale commercialization.