On August 26, 2026, INSOL-HIGH began operating the J-HRTI facility in Japan, an abbreviation for Japan Humanoid Robot Training & Implementation, to collect training data for humanoid robots and develop “physical AI” applications that connect software decisions with movement and interaction with the real world.
The facility is the first project under the J-HRTI initiative and covers approximately 1,400 square meters. It can accommodate up to 40 humanoid robots, with operations initially beginning with 35 robots. According to the article, the facility will collect data by having human operators control the robots, with the goal of building a dataset that can be used to train AI models.
How Is Training Data Collected?
The facility is divided into three main areas. The first is dedicated to training humanoid robots and collecting movement data; the second converts the collected data into usable training data through labeling and annotation processes; and the third is used to test what the models have learned and verify their ability to perform tasks under conditions close to practical operation.
The examples presented include the use of AgiBot G1 and AgiBot G2 robots, with operators wearing virtual reality headsets to perform different movements and tasks. Scenarios include moving containers, using the hand to pick up or move objects, and sorting various items. Additional data is also collected through cameras positioned around the work area to monitor the robot’s movements and its interaction with the environment.
The process is not limited to recording ideal movement; it also includes situations such as deliberately dropping items and then retrieving them, so that models learn how to handle errors and unexpected conditions. After the data is recorded, it is reviewed and labeled to identify movement details and context, then converted into training data that can be used during the learning and testing stages.
What Changes in Practice?
The importance of the initiative lies in moving humanoid robot training from the laboratory to environments closer to actual work lines. The required tasks and necessary data vary depending on the operating site and the nature of the work. Therefore, INSOL-HIGH plans to begin collecting data according to the targeted operations, then change the training subjects and areas as needed.
This approach indicates that developing a robot capable of working depends not only on the software model, but also on movement and interaction data drawn from varied real-world situations. The presence of a separate testing area also makes it possible to compare the robot’s performance after training with its behavior during data collection, rather than simply displaying a preprogrammed movement.
Limitations and Open Questions
The article does not provide details such as the software models used, the expected size of the dataset, or the timeline for making the data available to other parties. Therefore, based solely on the published information, it is not possible to determine the extent to which the results can be generalized to different factories or robots. However, operating a facility capable of accommodating 40 robots places the collection of organized data, including error and recovery data, at the center of the development path for humanoid robot applications in Japan.