J-HRTI Kanto Data Factory in Narashino, Chiba Prefecture, Japan, has begun operating what its operator describes as Japan’s first privately operated data factory for physical AI and robotics run by domestic companies. The site currently has 35 humanoid robots, and the data collected is shared among the participating companies.
The facility comes at a time of growing expectations for using physical AI to automate real-world tasks. The article says the Japanese government has designated this field as strategic, with a total investment of 10.5 trillion yen by the public and private sectors through fiscal 2040. Estimates by the National Institute of Population and Social Security Research also indicate that Japan’s working-age population will decline from 75.09 million in 2020 to 62.13 million in 2040.
How Is Training Data Produced?
The site is divided into an area for operating the robots, an area for adding annotations to recorded footage, and an area for testing the participating companies’ environments, although the last area was not open to visitors. Operators control the robots remotely using virtual reality headsets and controllers, while the robots perform repetitive tasks such as moving dolls or food samples, removing products from boxes, and pouring resin pellets with a scoop.
Three cameras mounted on the robot’s head and wrists capture movement details. The footage is then divided into time periods and descriptions of the movements are added, such as using the left arm to pull an object. The process is not limited to recording successful attempts; failures from which the robot can recover are also labeled, such as dropping an object inside the box and picking it up again, while cases in which the object cannot be recovered or which result from operator error or a camera malfunction are excluded.
From Data to the Model and Then to the Field
The refined data is used to develop models and test them on physical robots, after which data generated during field operation is returned to the facility to update the model. Under the announced plan, the project aims to apply simple tasks such as transportation during fiscal 2026, expand to ten sites across Japan during fiscal 2027, and then provide industry-specific models in fiscal 2028.
The consortium comprises five companies, including Tsumura, DMS, Fuyo General Lease, and Yamazen, as well as one company whose name has not been disclosed. INSOL-HIGH serves as the secretariat, while the motivations for the participation model include sharing data and equipment costs and designing tasks based on the needs of actual workplaces.
What Limits Expansion?
Operations began on August 26, 2026, after the plan announced in March had referred to a site capable of accommodating up to 50 robots. However, building constraints reduced the suitable number to approximately 35 or 40 robots. The site covers approximately 1,400 square meters, while its detailed address has not been disclosed for security reasons. The facility primarily uses G1 robots from AgiBot, with the five-fingered G2 model being introduced gradually.
Editorial perspective: The project’s significance lies not only in the number of robots, but also in its attempt to turn the collection of motor-skill data into shared infrastructure among different industrial companies. The success of this model will depend on the amount of data returned from actual use and on the ability of shared data to reduce retraining for each new task. Whether the service will reach small and medium-sized companies remains uncertain; the idea of lending first-person data-collection tools or tools for controlling the robot’s end effector is still under consideration, and final usage fees have not been determined.