Mifeng unveils million-hour robot-agnostic dataset

Operators no longer have to remotely control a real robot. Instead, they can wear first-person-view equipment or wrist-mounted devices, or directly hold a gripper to grasp, move and organize objects. On August 31, Mifeng Technology, a subsidiary of AgiBot, announced that its 20,000th MEgo-series device had rolled off the production line and released a 1-million-hour embodiment-free dataset. The data comes from people performing different tasks and is not tied to any particular robot embodiment, meaning that physical experience that previously went unrecorded is beginning to enter large-scale production.

The data's breadth is also being put on display: 22 broad scenario categories, more than 500 types of tasks, over 10,000 real-world scenarios and more than 50,000 object categories. Of this total, 500,000 hours came from bare-hand collection, 350,000 hours from wrist-mounted collection and 150,000 hours from gripper collection. Mifeng Technology was founded in February this year. Its chairman and CEO, Yao Maoqing, said that achieving artificial general intelligence and ultimately physical AI may require tens of millions or hundreds of millions of hours of data, but that today is only the starting point.

The first problem this approach addresses is “how to obtain data more cheaply.” He He, vice president of JD Technology, recalled that traditional teleoperation requires an operator to work for eight hours but may ultimately produce only one hour of usable data, while also incurring the costs of the robot, electricity, labor and facilities. Wearing equipment during normal work, by contrast, can generate as much as 5—6 hours of data in eight hours. Data-production efficiency has improved, expanding the supply of training material, but the amount of genuinely valuable action in one hour of data cannot be judged by duration alone.

More specifically, 1 million hours does not mean that robots have become equally more intelligent. Language models have ready-made digital material in the form of webpages, books, code and videos, but experience such as picking up a cup, folding clothes or placing objects made of different materials on a shelf does not naturally appear on the internet. However, the grippers, degrees of freedom, dimensions, torque and control frequencies of different robots vary, so actions performed by humans cannot be directly converted into control signals for every robot. Embodiment-free data is mainly used for pretraining and general representation learning; deploying a specific robot still requires real-robot data.

For factories, warehouses, shopping malls and future household settings, the most practical change for now is that data collection no longer has to rely solely on costly, single-robot teleoperation. In theory, a richer combination of environments, objects and tasks could help robots learn more general rules. But what Mifeng Technology demonstrated this time was mainly the scale of the data, its coverage, and its collection and governance capabilities. It has not yet disclosed how much capability gain the million hours of data can deliver for any specific model. Industrialization of data-production capacity has already taken place; whether intelligent capabilities will emerge alongside it remains for model training and real-world deployment to answer.