Mifeng Launches Crowdsourced Robot Training Data Platform With 20,000 Collection Devices

Users download the Mifeng Pai app, apply for a collection device, claim tasks and then upload the data they collect. After processing and grading, the data is sold for robot training. Mifeng says it has produced 20,000 MEgo devices, which it describes as the industry's first large-scale mass production of such equipment. The MEgo line includes MEgo View, a head-mounted panoramic camera with a wrist close-up camera, and MEgo Gripper, a lightweight UMI gripper that weighs less than one jin and includes 3D tactile sensing.

Mifeng chairman and CEO Yao Maoqing compared embodied data collection to electricity infrastructure, saying that power truly changed the world only after the grid was built. He said MEgo production capacity has reached tens of thousands of units per month. The 20,000 devices have entered real-world settings and collected one million hours of effective data, according to the company. Mifeng says the data comes from open environments, covers 22 major scene categories and includes bare-hand, wrist-worn and gripper collection forms, with modalities such as color images, depth, inertial measurement, touch and audio. It claims to be the first data service provider with million-hour-level non-embodied data available for sale.

Mifeng says its MEgo Engine handles preprocessing, spatial perception, annotation and quality evaluation. Its HandPose technology combines five-view panoramic perception, hand tracking and three-view depth fusion with parametric hand reconstruction and trajectory recovery. The company says seven core metrics reach state-of-the-art levels in real work scenes involving occlusion, interaction and fast movement. In milk tea preparation, it says, sub-centimeter 3D reconstruction is possible even when hand visibility is below 30 percent; in nail salons it can distinguish the collector's hands from a customer's; and in pet washing it can stably reconstruct hands despite motion blur and foam. Mifeng also built the MPR unified algorithm base, which it says achieves trajectory reconstruction error below 1 centimeter for head, hand and gripper poses, and sub-centimeter dense depth reconstruction of environments.

For quality control, Mifeng uses ManiEval, which evaluates data by type, task, sensor quality, semantic quality and action quality, then assigns S, A, B and C grades for different training needs. Precise action trajectories can be used for policy imitation learning; scene-rich data with clear task meaning suits general pretraining and representation learning; operational errors can train error recognition; rare scenes and difficult samples help test and expand generalization. Mifeng describes the approach as not filtering, only grading, while saying compliance and basic quality requirements remain prerequisites. The company is exploring uses around embodied brains, vision-language-action models and world action models. Large-scale non-embodied data mainly serves general pretraining, while robot-specific real-machine data is still needed for post-training and optimization. Mifeng's REMORA task progress model tracks what a robot is doing and where it fails, and a Value Model selects valuable failure and difficult samples from deployment for return to the data system, forming a loop from collection and reconstruction to quality control, training, deployment and data return.

On the crowdsourcing side, Mifeng says a one-month beta before the official launch drew 20,000 registered users who submitted 13,000 collection tasks. Participants claim tasks, wear MEgo devices, complete specified operations and receive pay based on valid duration after review; one person earned more than 5,000 yuan in a month, according to the company. Mifeng Pai has a scene network covering hotels, catering, supermarkets, logistics, factories, medical care, homes, elderly care and intangible cultural heritage, and a people network spanning different ages, industries and regions. Mifeng calls participants robot trainers.

Figure's Index, which emerged from stealth in August, is pursuing a similar crowdsourcing route. QuantumBit reported that Index accumulated 16 million uploaded videos in four months, covered 108 countries, passed 264,000 app downloads, had more than 43,000 weekly active contributors and paid $15 million to participants. Figure plans to invest more than $1 billion in data and computing over the next year. Figure's Helix 2.5 first used large-scale human behavior data collected by Index for pretraining, then learned specific tasks; in 30 unfamiliar homes, task success rose from 9 percent to 56 percent after adding human data pretraining. Both companies conclude that the embodied AI data bottleneck cannot be solved by relying only on full-time collectors.

Mifeng and Index differ in a key way. Index is used for Figure's own model training, while Mifeng Pai serves various data buyers, organizes production according to customer demand and delivers processed data. Mifeng says China's large industrial scene base, including the world's largest manufacturing system, dense logistics and warehousing networks, and rich retail and service businesses, provides real-world classrooms for robots. It has launched a Scene Data Alliance that brings together scene owners, collection networks, robot and model companies; more than 50 enterprises and institutions joined the first batch. Mifeng says its million-hour-level non-embodied data has begun serving customers including Tencent Robotics X lab and Ant Lingbo.