Japan is accelerating development of physical AI to blend sensing, decision-making, and real-world action, leveraging its manufacturing strengths to compete with rivals in the US and China.

Industry leaders like Hitachi and Mitsubishi group entities are deploying physical AI, signaling a national strategy that fuses hardware, software, and industrial data to train AI in real-world environments.

Donut Robotics’ Cinnamon 2 humanoid robot showcases advanced capabilities such as video-based movement learning, vision-language understanding, face recognition, gesture-triggered actions, and security patrol functions, with an official lease launch planned by the end of 2026 and a formal unveiling in October.

Security and patrol applications are viewed as among the first commercially viable uses, while factory automation may take longer due to development needs for autonomous movement.

New startups like Enactic are introducing robot arms that learn tasks from leader-arm data collection, contributing to a broader ecosystem showcased at Nvidia events that attract global participation.

Cinnamon 2’s visual-language motor (VLM) enables it to describe surroundings and classify visual data, while gesture control can trigger predefined actions, addressing noisier work sites where voice commands are impractical.

Enactic also offers an open-source robot arm to encourage widespread adoption and data generation, reinforcing a Japan-centric approach to accumulating proprietary training data through factory and industrial collaborations.

Experts warn of AI risks, but proponents emphasize that current physical AI remains under human control due to reliance on physical infrastructure, though long-term governance will require ongoing attention.

Cinnamon 2 uses in-house AI with Chinese-made hardware under OEM arrangements and is expected to undergo further design changes in 2027, signaling ongoing refinement for deployment in security, construction, nursing care, and factories.

The government is pushing a broader physical AI ecosystem, including a plan with DMG Mori and Komatsu to collect factory data via common identification numbers across companies to enable cross-boundary data use and improved robot coordination.

Japan argues against building large general-purpose models from scratch, promoting adaptation of open-source or Western models to industrial contexts with safety, specialized knowledge, and factory-specific data as competitive advantages.

Physical AI differs from traditional robotics by enabling perception, decision-making, and self-planning of movements rather than pre-programmed repetitive motions, allowing Japan to leverage deployment timing alongside foundational model-building.