Japan Bets on Physical AI

One of the latest examples is Cinnamon 2, a humanoid robot being developed by Japanese startup Donut Robotics. The robot can reproduce fluid human movements learned from video or motion-capture data, understand spoken instructions, recognize people and gestures, interpret its surroundings through cameras and perform tasks intended for security, construction, nursing care and factories.

Donut Robotics CEO Daisuke Ono said Cinnamon 2 is still under development and is scheduled for an official lease launch by the end of 2026. The company plans to formally unveil the machine in October.

The robot’s artificial intelligence is developed in-house by Donut Robotics, while its hardware is currently produced in China under an original equipment manufacturing arrangement before being assembled in Japan. The company plans further design changes in 2027.

Cinnamon 2 is equipped with a vision-language model, or VLM, allowing it to interpret both images and language. During a demonstration, the robot was asked what it could see and gave a detailed description of the studio, including the floor, people, cameras and surrounding equipment.

The same visual system can also be adapted for security work. Cinnamon 2 can photograph people it identifies during patrols and immediately transmit the images to a management department, allowing suspicious individuals to be reported.

The robot can also recognize gestures made by registered users. A peace sign or a bow, for example, can trigger predetermined movements once the machine confirms the identity of the person giving the command.

Ono said gesture control could be useful at noisy construction sites where verbal instructions are difficult to hear. Registered workers could control the robot from in front or behind without relying on voice commands.

Its humanoid shape also gives it potential advantages in environments designed for people. The machine can walk over uneven surfaces, negotiate areas where wheeled robots would struggle and extend its arms to inspect locations above normal reach.

Ono said security inspections and patrols are likely to be among the first commercially practical applications because robots do not initially need completely autonomous movement to perform useful work. Factory applications are expected to take longer to develop.

Physical AI differs from conventional industrial robotics because traditional machines generally repeat movements that have been individually programmed in advance. A physical AI system is instead designed to perceive its surroundings, make decisions and plan its own movements.

Japan was once one of the world’s leading countries in humanoid and industrial robotics, but China and the United States have moved rapidly into the new generation of AI-powered machines.

Ono argues that Japan does not necessarily need to build its own large foundation models from the ground up to remain competitive. Open-source models or systems developed by major Western companies could be adapted and fine-tuned for Japanese industrial environments.

Instead, he sees Japan’s advantage in the final stage of deployment, including safety, specialized industrial knowledge and data collected inside Japanese factories.

Unlike language and image AI, which is largely trained on information available across the internet, physical AI depends heavily on data generated in actual workplaces. Such information is often unavailable to global technology companies.

Japan’s manufacturing sector could therefore become an important source of proprietary training data, particularly in areas involving safety standards, industrial processes and robot components.

Japan Advanced Institute of Science and Technology professor Shota Imai said this difference could give Japan an opportunity. Training general-purpose language and image models has increasingly become a competition over computing resources and capital because companies around the world can access much of the same internet data.

Manufacturing data is different because access depends on relationships with factories, machinery makers and other industrial operators.

The Japanese government is also moving to build a broader physical AI ecosystem.

A government initiative involving major machinery manufacturers including DMG Mori and Komatsu aims to collect learning data that can be used to make robots and industrial equipment operate autonomously.

Under the plan, machinery in factories would be assigned common identification numbers, similar in concept to identification systems used for individuals, allowing manufacturing data to be collected across company boundaries.

Ono described the effort as a major step, saying common identification standards could eventually make it easier for humanoid robots to coordinate with smaller robots and machinery already operating inside factories.

AI systems generally improve when large volumes of diverse data are gathered and used for training, and robot movement data must usually be collected deliberately rather than appearing naturally on the internet.

Imai said coordinated data collection could therefore become an important part of Japan’s strategy.

The push comes as concerns about the risks of increasingly capable artificial intelligence continue to grow internationally.

Some researchers and industry figures have warned that future AI systems could become difficult to control, while recent experiments have raised concerns about unexpected behavior by autonomous AI agents.

Imai said incorporating fears of human extinction into immediate regulation would be excessive at the current stage, but he believes long-term risks will eventually require serious consideration.

He also warned that less extreme but still serious incidents involving AI could occur within the next one or two years as systems become increasingly capable of acting on their own.

Current AI models remain physically constrained because the largest systems require substantial computing infrastructure, data centers and electricity. Their enormous size also makes the science-fiction scenario of an AI easily copying itself across decentralized networks much more difficult than it may appear.

Ono similarly said present-day AI remains under human control because it depends on physical infrastructure. However, he said the risks could become more significant if artificial general intelligence capable of exceeding human abilities emerges.

The renewed interest in physical AI is also producing a new generation of Japanese startups.

Tokyo-based Enactic, founded in July 2025, has 17 employees and is developing both robot hardware and the AI software needed to train it.

The company was invited to an AI conference organized by U.S. semiconductor giant Nvidia in California in March 2026, where roughly 400 companies from more than 190 countries and regions participated.

Large Japanese companies including Hitachi and Mitsubishi group companies also attended, while Enactic participated as a Japanese startup developing physical AI.

One of Enactic’s demonstrations involves a robot arm learning to insert a pillow into a pillowcase, a flexible and awkward task that is difficult to automate using traditional programming.

A human operator first controls the robot through a device known as a leader arm. The robot records the operator’s movements and uses the resulting data to train its AI.

By repeatedly observing examples, the system can improve its ability to perform the task autonomously.

The same approach has been used to teach robots how to handle plastic bottles and assemble complicated boxes.

Enactic has also developed an open-source robot arm whose design information is available free of charge.

Rather than selling the device itself, the company publishes the specifications so engineers around the world can order the necessary parts and build their own versions.

The strategy is intended to encourage widespread use of the technology, generate feedback and create additional training data that Enactic can eventually apply to humanoid robots.

Company executives said Enactic’s ability to develop hardware, software and AI training systems internally was one factor that attracted attention at Nvidia’s event.

Takayuki Furuta, director of the Future Robotics Technology Research Center at Chiba Institute of Technology, is also working on physical AI for factory automation, autonomous patrols and plant inspections.

Furuta said artificial intelligence development has in some ways come full circle. Researchers who once left robotics to focus on AI are now returning because generative AI can produce text and images but cannot physically carry out the actions it recommends.

Physical AI gives that intelligence a body.

Instead of engineers programming every movement individually, AI systems can learn how to generate appropriate movements themselves after observing human examples.

The potential market is large. The global physical AI market is projected to reach about 11 trillion yen by 2034, roughly 13 times its current size.

Japanese developers are also exploring how the technology could move beyond factories into everyday life.

Another project seeks to combine robots directly with residential architecture to automate household chores.

The company Mu, led by CEO Osamu Narita, is developing a home in which physical AI robots and sensors work together to perform domestic tasks, with the long-term goal of reducing household work that can consume several hours each day.

Narita said the objective is not simply to build a more convenient or technologically advanced house, but to give people more time for work, family and other activities that enrich their lives.

A prototype structure already contains household equipment including air conditioning, water systems, a washing machine and a kitchen.

Its robotic system, known as MuBot, moves along rails installed through the home.

When a resident returns with groceries, sensors can detect the arrival and activate the system. A transport robot then carries the shopping bag along the rail to the kitchen.

Once the container reaches the kitchen, another robot equipped with two articulated arms can take over the next stage of the task.

The project reflects a broader vision emerging from Japan’s physical AI industry: rather than trying to defeat global technology companies in every part of artificial intelligence, Japanese companies are seeking to combine existing AI models with robotics, manufacturing knowledge, safety expertise and real-world data.

For developers such as Ono, that final connection between intelligence and the physical world is where Japan still has a chance to establish a leading position.

Source: テレ東BIZ

Disclaimer : This story is auto aggregated by a computer programme and has not been created or edited by DOWNTHENEWS. Publisher: newsonjapan.com