Mifeng, a subsidiary of ZHIYUAN, Launches One-Stop Physical AI Data Service Platform | Breaking News
In 2026, large language models and large video generation models are burning tokens like crazy, while the embodied robot industry is experiencing a situation of "having no tokens to burn". Large models can read books like humans, but embodied intelligence has to gain data through real - world experiences. The lack of data has become a bottleneck for the entire industry.
On April 16th, Mifeng Technology, a subsidiary of ZHIYUAN ROBOT, launched a one - stop physical AI data service platform, hoping to make data as readily available as water and electricity.
"GPT5 used 100 trillion tokens of training corpus. One token is approximately equal to 0.75 English words. If an average person can speak 150 words per minute, this corpus level is equivalent to a person speaking for 10 billion hours," said Yao Maoqing, the chairman and CEO of Mifeng Technology. "However, embodied intelligence is different. Today, the high - quality data from around the world, when combined, may only amount to about 500,000 hours."
The shortage of data resources and the slow expansion are due to the fact that the training data required for embodied intelligence is much more difficult to obtain than the training corpus needed for large prediction models. In the three - dimensional open world, companies in the industry have tried various methods such as real - machine remote operation or simulation data to accumulate data, but they still struggle to overcome the problems of high cost and slow growth.
Currently, the most advanced collection method is "object - free/body - less data collection".
Object - free/Body - less Data Collection refers to directly using human operations (hand - worn sensor grippers) or lightweight devices to record actions in embodied intelligence training, rather than relying on expensive physical robot bodies for remote control. It records high - quality, multi - modal action data through human hand grasping, moving, etc., with advantages such as low cost, high collection efficiency, and strong scene generalization ability.
At the press conference, Mifeng launched the MEgo series of object - free data collection hardware and the MEgo Engine data governance engine. The MEgo series includes collection grippers, head - mounted collection devices, etc. The devices have a panoramic perception of over 300° and sub - millisecond data synchronization capabilities, supporting lightweight collection anytime and anywhere in all scenarios such as factories, supermarkets, and homes.
This gripper named MEgo Gripper supports 1080P 60fps on all channels. The accuracy of trajectory reconstruction can reach one millimeter, ensuring that even the force of picking up a piece of paper can be restored. "It allows customers to directly implement the data once they get it," and it also has sub - millisecond wireless time synchronization. This device weighs only 480g, supports fast battery replacement and removal, gets rid of wires, and makes it convenient for people to "collect data wherever they go".
MEgo Gripper
Another head - mounted collection device, MEgo View, combines a field of view of over 300 degrees from the head and two local cameras attached to the wrists. It can not only cover a wide - area environment but also capture details of wrist and hand operations. It is equipped with 7 high - definition cameras and a vehicle - grade nine - axis IMU (Inertial Measurement Unit), and can output full - sensory data including RGB pictures, IMU, pose, and audio.
MEgo View
Like the gripper device, the head - mounted collection device also adopts a wireless design, supports fast battery replacement, and can achieve sub - millisecond wireless time synchronization.
The lightweight hardware has lowered the threshold for data collection. In the interview session after the press conference, Yao Maoqing told the media including 36Kr, "He believes that the ideal working mode of collectors in the future may be similar to that of 'Meituan riders' - 'People can do it part - time, but they also need to receive training at the station before taking up the job.'"
In terms of software, there is a MEgo Engine data governance engine behind the MEgo series of solutions, which is used to process all the data collected by MEgo devices, including data pre - processing, extraction, evaluation, etc. It can also evaluate the performance on robots and achieve various annotations of one - stop data.
Yao Maoqing said that Mifeng has achieved full - paradigm coverage of real - machine remote operation, object - free data collection, and simulation data, aiming to "make high - quality data as readily available as water and electricity." The company plans to achieve a data production capacity of tens of millions of hours in 2026 and reach a data production capacity of tens of billions of hours in 2030.
As a subsidiary of ZHIYUAN ROBOT, Mifeng positions itself as a To B data service platform for other robot companies. In the interview session after the event, a media asked Yao Maoqing, "How do you convince ZHIYUAN's competitors to buy your data?"
Yao Maoqing replied, "As an independent data service platform, Mifeng has strict agreements for all user data transactions. The data trading models are divided into 'right of use' and 'ownership'. In the past, most users chose to buy the right of use rather than the ownership. For the very few customers who choose to buy the ownership, we will conduct strict asset transfer and destroy the data locally."
"ZHIYUAN doesn't need all the data, and it can't obtain Mifeng's data either," said Yao Maoqing. "The only way for ZHIYUAN to obtain data from Mifeng now is to place orders in the market. There is no way for ZHIYUAN to obtain data from Mifeng for free."
At the press conference, Mifeng announced strategic signings with several enterprises such as JD Cloud, Baidu Cloud, Alibaba Cloud, Liepin, and Guizhou Big Data Group. All parties will carry out in - depth cooperation in fields such as data ecosystem, scenario collaboration, and computing power support.