On July 16, 2026, Fujitsu began exploring business initiatives toward the social implementation of physical AI together with three companies: FANUC, Yaskawa Electric, and Kawasaki Heavy Industries. Incorporating NVIDIA's AI models and simulation technology, the initiative aims to standardize and open up a coordinated control platform usable across manufacturing, retail/logistics, and healthcare. The star of the announcement is not a new robot model. Rather, it is "Fujitsu Kozuchi Physical OS," which connects business systems—such as production management and electronic medical records—to robots from multiple manufacturers, and allocates work across an entire site.
This business exploration represents a stage in which the full-stack AI platform that Fujitsu and NVIDIA unveiled in October 2025 is extended to physical machines and on-site operations. However, this is not the launch of a finished product or a joint venture. The specifications and rollout timing for the common interface have not yet been disclosed. Pricing and deployment targets have also not been announced, and each company will now begin building roadmaps for technology development and business expansion.
A Four-Layer Structure Spanning from Business Instructions to Robots
Fujitsu's announcement materials depict the concept in four layers. At the top sit systems for production management, logistics management, and medical/nursing care information; below that, Physical OS converts business instructions into robot tasks. The design separates control for individual machines from the computing infrastructure used for AI training and validation.
| Layer | Main Components | Role |
|---|---|---|
| Business | Production management, logistics management, medical/nursing care information systems | Instructs what to execute, when, and under what conditions |
| Coordinated Control | Fujitsu Kozuchi Physical OS, Takane, spatial prediction | Task planning, allocation, coordination, and monitoring for multiple robots |
| Physical Machines | Robots and equipment from FANUC, Yaskawa Electric, and Kawasaki Heavy Industries | Executes movement, transport, grasping, processing, and other actions |
| AI/Computing | NVIDIA Cosmos, Omniverse, Isaac, Newton, FUJITSU-MONAKA | Supports on-site prediction, simulation, training, and validation |
NVIDIA Cosmos is incorporated as a world foundation model that predicts upcoming changes from on-site video and status data. Omniverse, Isaac, and Newton support Sim2Real, in which robots are trained and validated in virtual environments before the results are transferred to physical machines. Fujitsu's plan places Takane, multi-robot task planning, site-wide prediction, and cyberattack countermeasures on top of Physical OS. The robot manufacturers, for their part, bring control technology that moves motors and sensors accurately and safely, along with operational knowledge accumulated on-site.
The "sovereignty" referred to here does not mean a purely domestic stack that excludes NVIDIA. Under the announced division of roles, NVIDIA handles models and simulation, Fujitsu handles the business, integration, and computing infrastructure, and the three companies handle machine control and on-site data. Whether user companies can manage where data is stored and who has access rights—down to model update authority and audit mechanisms—will determine the substance of this sovereignty.
The Three Companies Are Already Using Separate AI Systems
FANUC has maintained a long-running collaboration with NVIDIA. In May 2026, it integrated its own robot simulator "ROBOGUIDE" with NVIDIA Isaac Sim and demonstrated a dual-arm robot folding T-shirts using GR00T N. In the exhibited system, which upgraded from Jetson AGX Orin to Jetson T5000, AI computing performance reportedly improved by more than 7.5 times. This figure does not represent the performance of today's common platform, but rather shows that FANUC's own on-machine AI is already advancing.
Yaskawa Electric has also been selling the autonomous MOTOMAN NEXT since 2023. The day before this announcement, the company unveiled a system combining MOTOMAN NEXT with Google DeepMind's Gemini Robotics ER 1.6. In this design, generative AI breaks down "what to do" into procedures, while the machine's own vision, path planning, and force sensing handle "how to move." The system was shown to recognize situations and redo work if it drops a part, and to share order information with internal systems upon detecting a parts shortage—though the fields and timing for deployment have not yet been disclosed.
Kawasaki Heavy Industries opened a physical AI development center in San Jose in May 2026, dividing collaboration themes among NVIDIA, Analog Devices, Microsoft, and Fujitsu. The role expected of Fujitsu is integration across business systems, robots, and AI. Looking across the moves of all three companies, what Physical OS requires is not lock-in to a specific machine or a single model, but the ability to standardize business operations at a higher layer while preserving existing AI pathways. If it becomes confined to an NVIDIA-only interface, the value of the open platform that Fujitsu champions will narrow.
Business-Driven Autonomy Connecting Factories, Logistics, and Hospitals
Conventional industrial robots have excelled at repeating fixed processes—such as welding or transport—with high precision. What Physical OS aims for is to use business information from outside the process to redistribute work across multiple machines. In factories, it reorganizes overall plans while monitoring production fluctuations, including orders and equipment downtime; in logistics, it reflects changes in sales and inventory into transport plans. Rather than making individual robots smarter, it changes site-wide operations at a layer above them.
For hospitals, the concept presented involves transporting medications and specimens, receiving outpatients, and providing guidance, all triggered by instructions issued from in-hospital systems such as electronic medical records. Kawasaki Heavy Industries envisions a future "hospital one-stop solution" connecting everything from hospital arrival through treatment to post-surgical care, but the concrete autonomous tasks presented this time are transport, reception, and guidance. This is not an announcement that surgical robots will perform autonomous surgery based on AI judgment.
Connecting to business systems expands the scope of potential impact from failures even as it improves efficiency. Fujitsu itself listed cyberattacks, system-wide outages or malfunctions, and leaks of confidential information as anticipated risks. An error in a production plan could halt a line; misdelivery or authority overreach in a hospital could affect patient safety. Physical OS will need functions that restrict authority, halt operations during anomalies, and track operation histories—at the same level of sophistication as its task-generation functions.
A Domestic Policy Target of 10 Million Units, Against an 18.3% Drop in Domestic Shipments
Japan's robot industry faces both a recovery in exports and weakness in domestic adoption simultaneously. According to the Japan Robot Association, orders for 2025 (excluding service robots) rose 25.7% year-on-year to ¥1,045.6 billion, while production value rose 21.0% to ¥945.3 billion. Export unit volume grew 28.2% to 173,323 units, while domestic shipments fell 18.3% to 37,816 units. There is a gap between the ability to manufacture and actual adoption at domestic sites.
Under its 2026 AI robotics strategy, the government has set a target of deploying approximately 10 million units domestically by 2040, capturing over 30% of the global market share, and securing a ¥20 trillion market. The scope covers 18 fields, including manufacturing and logistics as well as medical and nursing care. In the near term, indoor and outdoor inspection and transport will be prioritized. The policy further calls for advancing adoption across eight common task categories, including cleaning, loading/unloading and palletizing, handling, and welding/painting/processing.
What the policy requires is a cycle that rapidly rotates through deployment at sites, collection of real-machine data, evaluation, and model improvement. Fujitsu's concept aims to claim the layer within that cycle that connects business systems with heterogeneous robots. If standardization progresses, task plans and safety measures gained at one site could be more easily transferred to other machines and industries. Conversely, if data formats and operations remain siloed by manufacturer, increasing the number of deployed units is unlikely to accelerate the pace of improvement.
What Do "Open" and "Sovereign" Actually Guarantee?
At the time of announcement, no common API or list of supported models has been disclosed. Data formats and the scope of real-time control also remain unclear. Whether "open" means source code disclosure, or rather publication of specifications and an expanding roster of participating companies, is also uncertain. Who will operate Physical OS, whether a standardization body will be established, and the procedure for swapping in models other than NVIDIA's all remain issues for the future.
On the safety front, the government's strategy also lists proof and certification of safety, allocation of liability in the event of accidents, and privacy and security as shared challenges. If it is impossible to record which component—business system, Physical OS, generative AI, or robot controller—made an erroneous judgment, operators will be unable to determine root causes. Conditions for human intervention, safe shutdown during communication failures, re-validation after software updates, and maintenance contracts spanning multiple manufacturers all remain to be decided.
Fujitsu plans to progressively incorporate the results of joint research with Carnegie Mellon University into Physical OS starting in fiscal 2026, but for this particular business initiative, the roadmap itself is still to be developed. Whether the first demonstration can show machines from different manufacturers being safely invoked from the same business system—complete with clearly defined boundaries for shutdown and authority—will be the condition for advancing this concept into an industrial platform.