Mengfengpai Launches Data Crowdsourcing to Teach Robots via Head-Mounted Devices
On the afternoon of September 23, at the global launch event of Mengfengpai, a giant screen displayed dense figures: 22 major scenario categories, over 5,000 tasks, more than 50,000 real environments, 20,000 sets of production measurement equipment, and 1 million cumulative valid data hours.
At the center of the stage, a Suzhou-style mooncake intangible cultural heritage inheritor wearing a head-mounted device demonstrated how to teach a robot using the hands that make mooncakes.
This was not a traditional tech product launch, but more like the inauguration of a new profession and a new industry.
The core challenge is not a lack of data, but a lack of tools to mine it. Yao Maoqing offered an analogy: large language models advanced rapidly because they consumed all the knowledge accumulated by humanity over millennia. However, the actions robots need to learn—grasping a cup, unscrewing a bottle cap, folding clothes, stocking shelves—are repeated countless times daily but have never formed large-scale training data.
Yao said they do not lack data for teaching robots, because every second, people across all industries perform massive numbers of actions that potentially generate vast data for robot learning. What is missing is the shovel to mine this data treasure. The industry's previous solution was real-machine teleoperation, but collection speed is constrained by the number of robots, operators, and sites; hardware, labor, and operational costs grow with data scale. Industry estimates suggest that achieving embodied AGI requires scaling data from millions to billions of hours.
Mengfengpai's solution is an infrastructure of three parts: hardware, an app, and a data engine. The first is the MEgo collection device, available in two products: MEgo View, a head-mounted multi-view device with cameras covering over 300 degrees, plus wrist-mounted close-up cameras to capture hand details; and MEgo Gripper, a wireless UMI-type gripper with millimeter-level spatial trajectory reconstruction and optional 3D tactile sensing. Both can be used independently or wirelessly networked for sub-millisecond time synchronization. By the end of August, 20,000 MEgo units had entered various real scenarios, accumulating 1 million hours of valid data.
The second part is the Mengfengpai app, which transforms professional data collection into standardized tasks for ordinary users, simplified into four steps: accept order, collect, verify, settle. The third is the MEgo Engine data governance engine, handling cleaning, segmentation, annotation, and quality assessment to turn raw data into trainable finished data.
Li Guoying, 42, a ride-hailing operations staffer at Zhejiang Meitu Travel, learned about Mengfengpai a month ago. Initially thinking it required technical background, she found it simply involved recording daily chores like washing, cooking, mopping, organizing, and building blocks with her child. In one month of beta testing, she earned over 3,100 yuan.
Short-video blogger Chu Siyu, who has experienced a hundred professions, found being a robot trainer completely different: she said that previously she was the one learning, but this time the robot is learning from her, and she needs to teach it the skills she knows.
Lu Xiaoxing, the sixth-generation inheritor of the Renchangshun Suzhou-style pastry-making technique, demonstrated how the MEgo device recorded the making of rose bean paste mooncakes. In one month of beta testing, Mengfengpai reached 20,000 registered users who submitted 13,000 collection tasks. On the earnings leaderboard, the top earner, Ms. Zhu, made 5,111 yuan, and the tenth place earned 3,595 yuan. The Ministry of Human Resources and Social Security recently recognized embodied AI robot application technicians as a new profession.
At the launch, Mengfengpai also announced two initiatives. The first is a 100-million-yuan subsidy plan: 50 million for task subsidies, 20 million for equipment, 30 million to build a global offline service network with over 50 service points in 40 domestic cities, and 2 million for insurance.
The second is a scenario data alliance whose first members cover retail supermarkets, hotels, catering, elderly care, healthcare, manufacturing and real estate, including over 50 enterprises such as Dossen Hotels, CTS Group, Huayi Hotel, Chenxianggui, Pushang Fresh, Pushan Nursing Home, Shanghai Deji Hospital, Zhenro Group, Longqi Technology, Haitian Ruixin, and Zhangjiang Group.
Mengfeng Technology also partnered with the China Academy of Information and Communications Technology, the National Robot Testing and Evaluation Center, Guizhou Big Data Group, China Telecom, and others to compile the 'Embodied Intelligence Crowdsourced Data Development Report.' Yao Maoqing concluded his speech by saying that the experience of billions of hands will ultimately converge into robots' capabilities.