AIRoA Publicly Releases About 5,000 Hours of Robot Motion Data and a Foundation Model

The AI Robot Association (AIRoA), a general incorporated association, on July 31, 2026, publicly released about 5,000 hours of cleaned data covering 68 task types and a continued pretraining model for the open-source VLA model π0.5. It also released AiroPi, source code for training. The materials are outcomes of research and development and verification using the mobile manipulator HSR (Human Support Robot). The materials were compiled as outcomes of a project selected for GENIAC, a program run by the Ministry of Economy, Trade and Industry and NEDO: “Research on building a robot motion dataset for training and verifying its potential application to robot foundation models to accelerate the development of general-purpose robots using AI.” AIRoA said the release will support research and development and verification of robot foundation models by research institutions and development companies.

By releasing about 5,000 hours of robot motion data along with training code, AIRoA makes training resources close to real-world robot data more accessible to researchers and development companies working on robot foundation models. The release as an outcome of a project selected for GENIAC points to the formation of a data-sharing framework backed by government support.

The release stands out because it makes available not just a demonstration or individual robot trial, but about 5,000 hours of motion data, a foundation model and training source code together. It includes a dataset covering 68 task types and a continued pretraining model for the open-source VLA model π0.5, suggesting an effort to support a workflow from input data through training and model development to reuse, rather than simply providing data. The work was conducted using the HSR, with the development of a training foundation based on data from real robots as its central aim. In a related article, AWS Japan holds final results presentation for physical AI support program, the program’s final results were made visible through presentations by multiple companies. The AIRoA release differs in that it goes beyond announcing individual projects and makes the data and model themselves available for external use. In physical AI implementation, not only model performance but also the volume and quality of real-world training data and the availability of reproducible code are likely to shape barriers to entry. From an industry-structure perspective, the release raises the question of how Japan can build the data preparation and distribution layer, in addition to developing higher-level models for robot control. AIRoA says the data release will support research and development and verification, a direction that contrasts with companies keeping their data within their own organizations. However, given that the materials released are AIRoA MoMa 5k, pi05_hsr_base and AiroPi, the next questions are how widely they can be adapted to other robot types and tasks, and under what conditions implementers can use them. Training performance, generalization range, licensing terms and plans to expand the data have yet to be verified. The release brings the discussion of robot foundation models from abstract concepts to concrete implementation resources, while also showing that several questions remain before practical deployment. Areas to watch include results from retraining with the released data and model, the potential to expand the tasks covered, and terms of use beyond research applications.

The released dataset is called “AIRoA MoMa 5k,” and a public page on Hugging Face was provided as its source. The model weights were released as “pi05_hsr_base,” and the training source code was made available on GitHub as “AiroPi.” The press release says each item is provided under the license and terms of use specified on its respective release page. Users are asked to check the technical documentation, terms of use and disclaimers on each page.

In Japan, how data from real robots such as the HSR is accumulated and released in a reusable form could become a constraint on the development of robot foundation models. The release suggests that the competitive focus could shift from closed, company-by-company data collection toward the development of shared datasets and training code.