MiraiBit Launches Full-Scale Physical AI R&D, Using OpenArm v2 and Unitree G1

The company has begun collecting robot-operation data in a teleoperation environment using OpenArm v2, with plans to use the data for AI training and evaluation and for testing on physical robots. It has also introduced the Unitree G1 humanoid robot into its development environment, broadening its scope from dual-arm robots to humanoids.

The plan to use OpenArm v2 and Unitree G1 for different purposes points to an effort to extend the application of real-robot data from dual-arm robots to humanoids. Including support for companies, universities and research institutions in building these systems could signal a shift in R&D focus from standalone models to data-collection infrastructure.

What is new in this announcement is that Physical AI is being treated not as a research concept but as a continuous process involving the collection of operation data with OpenArm v2, AI training and evaluation, and testing on physical robots. As an extension of an established approach, this follows the basic Physical AI model of moving robots to collect data, and the announcement sets out how the company plans to implement it. The introduction of Unitree G1 could be seen as a step toward extending data collection from dual-arm robots to humanoids, but for now the announcement describes only the R&D environment. We have no related coverage, so there is no continuity with our previous reporting. The focus here, however, is less on “what to build” than on “how to create training data.” Physical AI cannot be completed by a model alone; the ability to establish a cycle of data collection from physical robots, evaluation and retraining is important. MiraiBit is creating an entry point with OpenArm v2 and aims to broaden its scope with Unitree G1, shifting attention from the performance of the robots themselves to the reproducibility of data-collection procedures and testing environments. The industry implication is that an R&D framework built around collecting real-world data could become a differentiating factor. For robots to act on visual or language instructions, they need not only model training but also an environment that can accumulate operation logs and real-robot behavior. The announcement says the company plans to build such an environment in-house and is also considering support for companies, universities and research institutions. This suggests scope to extend its data-collection methods beyond its own research. Unresolved questions include which tasks will be collected and in what volumes, the format of the training data, evaluation metrics and the scope of proof-of-concept projects. How OpenArm v2 and Unitree G1 will be assigned different roles, and how far testing on physical robots will proceed, also remain to be seen.

The company said it has built a teleoperation environment using OpenArm, an open-source dual-arm robot, and started collecting operation data. The target is OpenArm v2, and the company has outlined an R&D process that repeatedly collects data, trains and evaluates AI models, and tests them on physical robots. It said it will continue collecting, training and evaluating real-robot data, and will also work on building real-robot data-collection environments for companies, universities and research institutions, as well as on Physical AI R&D and proof-of-concept projects.

In Japan, the extent to which real-robot data-collection environments can be established is likely to have a direct bearing on Physical AI R&D capabilities. The announcement describes a testing platform based on physical systems, including OpenArm v2 and Unitree G1, raising the question of whether research institutions and companies can build similar environments.