Japan wants to turn its factory floors into the proving ground for the next frontier of artificial intelligence. The question is whether it still has time.
At her New Year’s news conference, Prime Minister Sanae Takaichi said Japan would harness decades of high-quality operational data from its manufacturing and service sectors to advance “physical AI” — systems that allow robots to autonomously support human workers with unprecedented precision. Around the same time, Nvidia CEO Jensen Huang told the CES technology show in Las Vegas that physical AI would define the next wave of innovation.
The concept is straightforward, even if the technology is not. It refers to systems that understand the real world’s spatial structure and physical laws and act accordingly. Techniques include imitation learning, reinforcement learning and vision-language models, along with video-action models that generate behavior from video. Embedded in robots and linked to actuators, these systems enable autonomous real-world action.
In Japan, interest in this field has accelerated. The AI Robot Association was established in 2025 to develop foundation models for AI-powered robotics, while the New Energy and Industrial Technology Development Organization has launched research calls related to AI robotics and physical AI. With media coverage increasingly highlighting AI-enabled humanoid robots, Japanese companies — long competitive in manufacturing and industrial robotics — have intensified their focus on this emerging area.
Because physical AI represents a fusion of software and hardware, some observers argue that Japan — despite lagging in the global AI race — could leverage its manufacturing strengths to regain competitiveness. But value creation is increasingly shifting away from hardware and toward AI foundation models that serve as the “intelligence” of robotic systems. To understand the implications for Japan, it is necessary to examine changes in technological architecture, application domains and value chains.
Physical AI is not simply an extension of traditional industrial robotics. Conventional industrial robots excel in controlled environments, performing predefined tasks with high speed, precision and reproducibility. Physical AI, by contrast, is designed to function flexibly in unfamiliar environments or unexpected situations. This capability, often described as “zero-shot adaptation,” allows robots to adapt without explicit programming of every movement.
Technologically, physical AI places its center of gravity on AI models that integrate multimodal perception, diffusion models for high-fidelity data transformation and large language models, or LLMs. Hardware, in this architecture, becomes subordinate to the AI “brain.” In demonstrations by companies such as the U.S.-based startup Physical Intelligence, robots equipped with such models can retrieve clothes from a dryer, place them in a basket, carry the basket to a workspace and fold the clothes — tasks requiring general adaptability rather than rigid programming.
This shift underscores a key difference: While traditional industrial...