Mifeng delivers 20,000th MEgo device to JD.com

On August 31, Mifeng Technology Chairman and CEO Yao Maoqing handed JD.com its 20,000th MEgo body-free data-collection device. The device will be used in JD.com’s real-world business scenarios. Mifeng Technology also announced that it had produced more than 1 million hours of body-free data and had begun providing data services to customers including Tencent’s Robotics X laboratory and Ant Lingbo.

MEgo does not record a single image stream. It combines a panoramic field of view of more than 300° with wrist close-ups, while simultaneously capturing RGB, depth, inertial-measurement, tactile and audio information. Three-dimensional tactile sensing works with millimeter-level trajectory reconstruction, while hardware-level synchronization aligns data from each channel to the sub-millisecond level. A person’s hands, body and head, along with the robot’s end effector and the surrounding environment, are thereby converted into unified spatial data.

The data comes from real-world collection in open environments and covers 22 major scenario categories, more than 10,000 real-world environments, more than 50,000 types of objects and more than 500 specific tasks. These range from residential, office, retail, production, warehousing, dining, entertainment and leisure, and transportation scenarios to long-tail applications in healthcare, education, elderly care and agricultural production. Mifeng said that eight leading scenario categories account for approximately 82% of the data, including around 500,000 hours of bare-hand Ego data, 350,000 hours of Ego+Wrist data and 150,000 hours of Ego+UMI gripper data.

To process this data, Mifeng built MEgo Engine, which covers processing, perception, annotation and quality assessment, and released its high-precision HandPose hand-reconstruction technology. The accompanying ManiEval system grades data across dimensions including data type, task type, sensor quality, semantic quality and action quality. Its approach is to “not filter, only grade,” matching data of different quality levels to different model-training needs.

The specific change is that robot-model training is beginning to receive a more continuous supply of data closer to real work: JD.com has received a set of data-collection devices about to enter business operations, while Tencent’s Robotics X laboratory and Ant Lingbo have become data-service customers. Mifeng aims to link collection, reconstruction, quality control, training, deployment and data feedback into a single pipeline. However, the disclosed production figures, data hours and coverage remain company-reported figures, and the devices’ performance in business settings—as well as the extent to which this data improves model capabilities—will require validation through subsequent applications.