Japan’s Physical-AI Race Moves From Robot Demos to the Hard Work of Keeping Machines on the Job

The most important part of a working robot may not be the moment it walks onto a factory floor. It may be what happens three months later, when a joint fails, a software update changes behavior, a customer needs a new task programmed, or an engineer has to determine whether the machine can safely reconnect to a corporate network.

That less glamorous layer of robotics is at the center of a capital and business alliance announced September 14 by GMO AI & Robotics Corporation, or GMO AIR, and Tokyo-based TechShare. GMO AIR plans to acquire newly issued TechShare shares and the companies will cooperate on sales, maintenance, after-sales support and solution development for robots made by China’s Unitree Robotics. The investment is scheduled to be executed on September 30. The companies did not disclose the investment amount or GMO AIR’s resulting ownership percentage.^[1]

The deal connects two very different histories in Japan’s Unitree market. TechShare has sold Unitree machines domestically since 2020 and built experience around research customers, repairs, secondary development and demonstrations. GMO AIR became an authorized Unitree distributor only in June 2026, but arrives with the broader GMO Internet Group’s customer relationships and capabilities in cloud infrastructure, communications and cybersecurity.^[1][2][3]

A capital alliance, not an announced takeover

The September 14 release is precise about what has and has not happened. The companies signed a basic agreement for a capital and business alliance. GMO AIR is to purchase TechShare shares, with execution planned for September 30. They plan to cooperate on Unitree distribution, maintenance, support and joint solutions. No purchase price, share count or ownership ratio was announced.^[1]

TechShare also says it will retain its role in areas where it holds exclusive Unitree sales rights for educational institutions. That detail matters: this is not simply the disappearance of one distributor into another. The alliance is designed to connect technical support and market reach while preserving specialized channels.^[1]

TechShare’s Unitree history began with a robot dog

TechShare started selling Unitree’s A1 quadruped in Japan in May 2020 as an authorized distributor. The roughly 12-kilogram machine was marketed as a relatively accessible development platform for work involving SLAM, machine learning and mobile robotics. In 2023, TechShare announced an exclusive domestic agreement covering the R&D version of Unitree’s Go1, with an emphasis on customers doing secondary development in corporate labs, universities and public research organizations.^[2][4]

That history explains what TechShare contributes to the new alliance. Development robots are not appliances. Customers may add sensors, rewrite control software, connect the machine to ROS, create custom autonomy, repair mechanical damage and tune the system to an experimental environment. The value is partly in the machine and partly in the engineering organization that can make the machine useful.

Unitree’s path from quadrupeds to humanoids

Unitree Robotics, based in Hangzhou, traces its company formation to founder and CEO Wang Xingxing’s quadruped work in 2016. Its official history lists Laikago in 2017, Aliengo in 2019, A1 in 2020 and Go1 in 2021. The company moved decisively into humanoids with the full-size H1 in 2023 and the smaller G1 in 2024, applying capabilities in motors, reducers, control systems, sensing and dynamic locomotion that had been developed through quadrupeds.^[5]

Unitree and its distributors describe the company as the world leader in 2025 shipments of bipedal humanoids. That ranking is based on Unitree’s own data, not an independently verified market census, so Japan.co.jp treats it as a company claim rather than an established industry statistic.^[1]

What is less disputed is the widening product range. Unitree now offers research and industrial quadrupeds alongside multiple humanoid platforms. GMO AIR began accepting Japanese orders for newer H2 Plus and H2-D humanoids and the wheeled As2-W quadruped in late August, adding them to a lineup that already included G1, H2, Go2 and B2.^[6]

GMO entered robotics only two years ago

GMO Internet Group created GMO AIR in June 2024 to build a business around AI, robots and drones. The company positioned itself not as a robot manufacturer but as a commercial integrator: identify a customer problem, provide hardware, combine it with software and infrastructure, and support deployment.^[7]

That strategy accelerated in 2025 and 2026. GMO used Unitree’s G1 in robot staffing and conversational-AI demonstrations, opened the GMO Humanoid Lab in Shibuya in April 2026, and signed its authorized-distributor agreement with Unitree on June 19. The Shibuya lab was designed as a physical-AI research, engineering and business-development base rather than merely a showroom.^[3][8]

In September, GMO AIR added two services that reveal where it believes the market is going. “GMO LOOP for Physical AI” cycles field data back into model improvement and then back onto deployed robots. A dedicated “humanoid ambulance” is intended to carry engineers, diagnostic equipment, parts and replacement machines to failures in the field. The branding is playful; the operational problem is not. Downtime can destroy the economics of automation.^[9][10]

Physical AI changes what counts as an AI system

Generative AI can fail inside a document or a browser window. Physical AI acts through motors, sensors and tools in environments occupied by people and valuable equipment. A mistaken answer is no longer only an information problem; it can become a collision, dropped load, damaged machine or unsafe movement.

That is why physical AI brings together disciplines that are often separated in software: perception, control, real-time computing, networking, cybersecurity, functional safety, mechanical reliability and data governance. A successful model is necessary but insufficient. The full system has to sense the world, decide in time and execute reliably.

Japan’s government is now treating that stack as industrial policy. In June, METI and NEDO launched a program to develop a domestic multimodal foundation model for AI robots and physical AI. The program explicitly targets models that can integrate language, images, video, audio and information about physical properties and real-world conditions, while making use of industrial data that Japan wants to keep secure.^[11]

METI’s revised AI Robotics Strategy goes further. The minister said the government is targeting roughly 10 million deployed robots by 2040 across 18 fields, and intends to support user adoption, research, deployment hubs and human-capital development. The policy emphasis is shifting from proving that robots can perform tasks to building an ecosystem capable of deploying them at scale.^[12]

Japan has been pursuing a “robot revolution” for more than a decade

There is a useful historical contrast. Japan’s 2015 New Robot Strategy described the country as a “robot powerhouse” and sought to make it the world’s leading robot showcase. The policy emphasized manufacturing, services, nursing and medical care, agriculture, infrastructure and disaster response, while also calling for more systems integrators and lower deployment costs.^[13]

The industrial robots behind that reputation were exceptionally good at a particular model of automation: operate behind safeguards, repeat defined motions, interact with standardized parts and maximize speed and precision.

Humanoids promise something else. Their economic case rests on the possibility that a machine can enter an environment built for people — doors, stairs, carts, tools, shelves, workbenches — and perform multiple tasks without rebuilding the entire site. That promise remains partly unproven at commercial scale, but it explains why companies and governments are willing to experiment.

Japan’s demographic pressure raises the stakes

The labor argument is not abstract. Japan’s Statistics Bureau estimated the population aged 15 to 64 at 73.29 million in August 2026, down 265,000 from a year earlier. The population aged 75 and over reached 21.488 million, up 479,000 year on year.^[14]

Those numbers do not automatically make a humanoid economical. A robot must still beat alternatives: a wheeled autonomous mobile robot, a fixed arm, a conveyor, redesigned software, better scheduling or simply a different workflow. The correct automation depends on the task.

But demographic pressure changes the question companies ask. A deployment is valuable not because a robot can walk across a stage, but because it can reliably cover a shift, reduce dangerous manual work, or allow scarce workers to concentrate on jobs that still require human judgment.

The maintenance problem may matter more than the headline hardware

This is where the GMO-TechShare alliance becomes more interesting. TechShare brings six years of Unitree support history in Japan. GMO brings a business model built around infrastructure and recurring operations. Put together, the target is not a one-time robot sale but a service chain: procurement, setup, connectivity, customization, field learning, security, repair and replacement.

Japan already has an enormous ecosystem of companies that keep industrial equipment running — machine-tool service organizations, factory integrators, telecom field engineers and maintenance contractors. Physical AI will need an equivalent layer. The firms that own that layer may capture substantial value even when the robot hardware itself comes from overseas.

Humanoid or specialized machine?

The case for humanoids should not be exaggerated. Many jobs are better served by simpler machines. Wheels are more efficient than legs on flat floors. Fixed automation can be faster and more reliable than a general-purpose robot when the task never changes. Human-shaped machines add joints, software complexity, balance problems and potential failure points.

The strongest argument for the humanoid form is compatibility with human infrastructure. GMO AIR explicitly cites buildings, doors, stairs, desks, appliances and vehicles as environments designed around the human body. If a humanoid can use those spaces without expensive reconstruction, its flexibility could offset some of its complexity.^[3]

An airport test announced by JAL Ground Service and GMO AIR illustrates the hypothesis. Ground-handling work takes place in constrained areas around aircraft and involves equipment of many shapes. The companies began a 2026 demonstration to test whether humanoids can help automate tasks that are awkward for fixed systems.^[15]

The experiment also illustrates the gap between possibility and production. Real airports bring weather, irregular objects, strict safety rules, mixed traffic, time pressure and expensive consequences for failure. Those are exactly the conditions that turn robotics from a demonstration problem into an operations problem.

Field data becomes a strategic asset

Physical AI also changes the value of data. A robot operating in a warehouse or factory can generate video, depth maps, motion traces, force readings, failure events and examples of successful task execution. Used well, those records can train better policies and improve the next deployment.

GMO LOOP is explicitly based on that cycle: collect operating data, strengthen models, return improvements to the robot. METI’s domestic multimodal-model program similarly treats Japan’s industrial field data as a potential competitive advantage.^[9][11]

That creates hard governance questions. Who owns the data generated on a customer site? Can it be transferred abroad? Can examples collected in one factory improve a model used by another company? How long is raw video retained? What happens when a model update changes behavior? How can a customer audit what the robot learned?

These issues will increasingly sit inside purchasing contracts, cybersecurity reviews and labor-management discussions. The physical-AI market will not scale on model accuracy alone.

Foreign hardware, domestic control

There is also an industrial-policy tension in Japan’s reliance on Unitree hardware. The Japanese government is funding domestic foundation models and data infrastructure at the same time Japanese companies are accelerating deployment of Chinese-made humanoids and quadrupeds.

The two trends can coexist. Using mature, purchasable hardware can speed research and give Japanese engineers real field data now. But dependence on overseas platforms creates questions about software control, components, communications, data flows, supply continuity and economic security.

Japan.co.jp’s analysis is that the more useful question is not simply “Japanese or foreign?” It is which layers of the physical-AI stack Japan controls: motors and sensors, robot bodies, operating software, foundation models, cloud infrastructure, communications, cybersecurity, field data, integration and maintenance.

A country can import a robot and still build a valuable domestic ecosystem around it. It can also lose strategic value if all of the learning, data and software intelligence flow elsewhere.

The real metric is not units sold

If the GMO-TechShare alliance works, the easiest number to publicize will be sales. But unit volume may be the least revealing metric.

Better measures would include uptime, mean time to repair, the percentage of tasks completed without human rescue, the number of jobs one platform can perform, safety incidents, the cost per productive hour and whether workers actually experience less dangerous or repetitive labor.

TechShare’s first A1 sales in 2020 belonged to an era when legged robots in Japan were still strongly associated with research labs and experiments. Six years later, the conversation includes maintenance vehicles, continuous field-data learning, cybersecurity and industrial deployment.

The robot industry’s next milestone may therefore look surprisingly ordinary: a machine that works on Monday, gets repaired on Tuesday and does Wednesday’s job a little better because of what it learned. The September 14 alliance is a bet that the companies able to make that ordinary routine possible will matter as much as the companies building the robots themselves.