Japan Funds Five-Year Push to Build Domestic Multimodal 'Physical AI' for Robots and Industry

Japan’s Ministry of Economy, Trade and Industry, or METI, said Tuesday it has launched a five-year, government-backed program to develop a domestic multimodal foundation model for robots and other “physical AI” systems, with Noetra Inc. and the National Institute of Advanced Industrial Science and Technology, or AIST, selected to carry out the work.

The project, announced in coordination with the New Energy and Industrial Technology Development Organization, or NEDO, is aimed at AI for real-world machines rather than a general-purpose consumer chatbot. NEDO’s basic plan says the model is intended to combine language, images and video, audio, and physical or spatial information so machines can better operate in industrial and other real-world settings. METI listed the program period as fiscal 2026 through fiscal 2030; for international readers, that means the final milestone falls in March 2031, the end of Japan’s fiscal 2030.

NEDO said it chose Noetra and AIST as the planned implementers after a public solicitation that ran from March 24 to April 22 and drew 15 applications. Under NEDO’s materials, Noetra will take the main role in developing and providing the model, while AIST, one of Japan’s largest public research organizations, will lead exploratory research on theory, component technologies and system architecture. NEDO also noted that Noetra was formerly named Japan AI Foundation Model Development Co.

The government is framing the program as an industrial policy measure as much as a research effort. METI and NEDO say the work is meant to support manufacturing competitiveness and help address labor shortages. NEDO’s plan also lays out two specific policy goals: allowing Japanese companies to use and protect sensitive industrial and field data through domestic models, and prioritizing low-power, energy-efficient AI systems in line with Japan’s energy constraints and broader green transformation, or GX, goals.

The project’s oversight structure is also notable. NEDO said it will use an annual stage-gate review process with outside experts, and contracts are initially committed only for fiscal 2026 and fiscal 2027. Any continuation beyond that will be decided through yearly reviews. NEDO’s proposal summary also says the trained model weights developed in the project are to be published domestically during the project period, a provision that could shape how the work is shared inside Japan before the program ends.

The initiative builds on GENIAC, the Generative AI Accelerator Challenge, under which METI and NEDO have supported domestic foundation-model development since February 2024. But the new program marks a more specific shift toward multimodal AI for robots, factories and other physical systems, where the government says control over industrial data and lower-power computing matter as much as raw model capability.

That makes Tuesday’s announcement less a product debut than a statement of policy direction. Japan is using public R&D funding to try to build homegrown AI for real-world machines, with Noetra and AIST now tasked with turning that goal into a domestic foundation model over the next five years.