Physical AI
Fujitsu and Carnegie Mellon Launch Joint Physical AI Center in Pittsburgh
Fujitsu and Carnegie Mellon University launched the Fujitsu-Carnegie Mellon Physical AI Research Center on April 23, 2026, pairing Japanese enterprise infrastructure with one of the world's top robotics programs. The center, operating from CMU's 14,000-square-meter Robotics Innovation Center in Pittsburgh, will feed research into Fujitsu's Kozuchi Physical OS platform, designed as a vendor-neutral coordination layer for multi-robot deployments across manufacturing, logistics, and healthcare.
Fujitsu and Carnegie Mellon University announced on April 23, 2026, the launch of a joint Physical AI research center, pairing the Japanese tech giant's enterprise AI and networking infrastructure with one of the world's top robotics schools. The center, which will operate out of CMU's new 14,000-square-meter Robotics Innovation Center in Pittsburgh, is designed to move physical AI from controlled lab conditions into real-world deployments across manufacturing, logistics, construction, and healthcare.
The partnership involves 13 named CMU faculty members across robotics, machine learning, language technologies, human-computer interaction, electrical engineering, and philosophy. Fujitsu will contribute its Kozuchi Physical OS platform, which integrates cloud-to-edge infrastructure for coordinating multiple robots in live environments. Research outputs are slated to be incorporated into Kozuchi starting in fiscal year 2026.
AI-generated image
Conceptual render of physical AI simulation environment. Source: AI-generated.
Key Stats
13
CMU Faculty Members
14,000 m²
Robotics Innovation Center
FY2026
Kozuchi Integration Target
6
Research Focus Areas
Why This Partnership Happened Now
Physical AI, the discipline of building AI systems that operate in the real world rather than on screens or in data centers, has attracted serious institutional money in 2026. The basic problem it addresses is that robots need far more data than they currently get. Training a large language model draws on the entire text of the internet. Training a robot arm to handle a new object type may require months of controlled real-world trials.
Fujitsu has been building toward this for several years. The company's Kozuchi Physical OS platform, revealed before today's announcement, was designed to be exactly the kind of infrastructure layer that would let a manufacturer plug in physical AI capabilities the way they plug in cloud storage today. Coordinating multiple robots in a live production environment, reconciling sensor data from machines made by different vendors, and doing it with the latency guarantees real operations require, these are the kinds of engineering problems Fujitsu's telecom and enterprise background actually prepares it for.
Carnegie Mellon brings the counterweight: 13 faculty spanning robotics, machine learning, natural language processing, and philosophy. The School of Computer Science's robotics program is one of the few programs in the world where a researcher could plausibly move from training methodology to actuator design to ethics review within the same building. The Robotics Innovation Center, which opened in February 2026 in Pittsburgh's Hazelwood Green district, gives the collaboration 14,000 square meters of test space that bridges research and commercial deployment.
Why It Matters
Most physical AI efforts are either entirely internal to companies (Tesla, Figure AI) or academic without a clear commercial path. Fujitsu-CMU is explicitly structured to bridge both, with Fujitsu researchers working alongside faculty and technology results flowing directly into a commercial product platform.
Under the Hood: The Kozuchi Physical OS
Fujitsu's Kozuchi Physical OS is the concrete output target for the center's research. The platform is structured around two capability clusters that Fujitsu describes as brain intelligence and spatial intelligence.
🧠 Brain Intelligence
Enhances robots' adaptability to tasks based on prior experience and human imitation. This draws on the center's work in action generation and learning, pulling from both CMU's machine learning faculty and Fujitsu's enterprise AI work.
🗺 Spatial Intelligence
Provides real-time information about physical environments in which robots operate. Relevant to the center's spatial perception and environmental understanding research track, and to Fujitsu's unified cloud-to-edge infrastructure goal.
🤝 Multi-Robot Coordination
One of the six explicit research focus areas: coordinating multiple robots and optimizing their behavior as a system, not just as individual units. This is critical for real factory deployments where a single workflow may involve multiple machine types.
🌐 Sim-to-Real Bridge
The center will research integration of simulation and real-world environments, a core technical challenge for physical AI. Validating robot behavior in simulation before deploying it on hardware reduces iteration time and safety risk.
AI-generated image
Conceptual render of a robotics innovation testing facility. CMU's Robotics Innovation Center opened February 2026 in Pittsburgh's Hazelwood Green district.
Industry-Academia Physical AI Partnerships: A Snapshot
| Partnership | Announced | Research Focus | Commercial Output | Status |
|---|---|---|---|---|
| Fujitsu + CMU | April 2026 | Multi-robot coordination, sim-to-real, HRI | Kozuchi Physical OS | Active |
| NEURA + Dassault | April 2026 | Virtual twins, synthetic data, sim-to-real | 3DEXPERIENCE integration | Active |
| NVIDIA + Partners | GTC 2026 | Cosmos world model, Isaac robotics stack | Cosmos 3 platform | Deployed |
| Microsoft + OpenAI | 2024-2025 | Foundation models for robotics | Azure AI + OpenAI robotics API | In progress |
| Toyota Research Institute | Ongoing | Diffusion policy, large behavior models | Internal Toyota deployment | In progress |
Who's Involved: The Research Team
The 13 CMU faculty members named in the announcement span a wider range than most corporate AI partnerships. The roster goes beyond the expected robotics and machine learning names to include language technologies researchers and two philosophers, a signal that Fujitsu and CMU intend to take the human-trust and ethics dimensions seriously rather than treating them as an afterthought.
Participating Researchers
• Yonatan Bisk — Language Technologies
• Fernando De La Torre — Robotics (Research Prof.)
• Tim Dettmers — Machine Learning
• Laszlo Jeni — Robotics
• Kris Kitani — Robotics (Associate Research Prof.)
• David Lindlbauer — Human-Computer Interaction
• Yorie Nakahira — Electrical and Computer Engineering
• Graham Neubig — Language Technologies (Associate Prof.)
• Jean Oh — Robotics (Associate Research Prof.)
• Sean Qian — Civil and Environmental Engineering
• Sebastian Scherer — Robotics (Associate Research Prof.)
• Peter Spirtes — Philosophy (Dept. Head)
• Kun Zhang — Philosophy
On the Fujitsu side, Corporate Vice President and CTO Vivek Mahajan signed the announcement. CMU's School of Computer Science Dean Martial Hebert, himself a robotics researcher, framed the collaboration as building on CMU's longstanding model of combining fundamental AI research with practical deployment work.
What This Means for Physical AI's Commercial Path
The specific sectors Fujitsu named in the announcement, manufacturing, logistics, construction, infrastructure, and healthcare, are exactly the sectors where labor shortages are most acute and where current humanoid robots are being trialed. But the framing here is different from a company like Figure AI or Agility Robotics, which are robot hardware companies first. Fujitsu is positioning Kozuchi as platform infrastructure: the layer that sits between robot hardware and business applications.
This is a structural gap in the current physical AI market. Today, a manufacturer deploying humanoid robots typically has to build significant coordination software in-house. There's no equivalent to AWS or Azure for physical AI yet. If Kozuchi matures as intended, Fujitsu would occupy that infrastructure position, and the CMU research pipeline would be the technical differentiation sustaining it.
5
Target Industries Named
2026
First Kozuchi Tech Integration
Feb 2026
CMU RIC Opened
The academic side has a practical consequence too. A joint center produces publishable research, which creates public knowledge about techniques and failure modes. Several of the faculty named, including Tim Dettmers in machine learning and Graham Neubig in language technologies, have publication records that influence the entire field. Research out of this center will shape what tools and architectures other physical AI developers reach for, not just Fujitsu products.
The Data Gap Problem (and Why It Matters)
• Digital AI: Trained on billions of internet documents, images, and code. Vast, cheap, readily available.
• Physical AI: Requires real physical interactions. Each training scenario is slow, expensive, and often specific to one robot model.
• The fix: Simulation (synthetic data generation) + real-world feedback loops. This is exactly what the Fujitsu-CMU center is designed to accelerate.
• Scale needed: Current estimates suggest physical AI systems need 100x more interaction data than they presently get to match digital AI's generalization capability.
The 12-Month Outlook
Fujitsu committed to integrating center research into Kozuchi Physical OS starting in fiscal year 2026 (which for Fujitsu runs April 2026 to March 2027). That means the first commercial outputs of this partnership could appear within 12 months. The specifics will depend on which research tracks move fastest, but multi-robot coordination and sim-to-real transfer are the most likely to produce deployable results quickly, both because they have the clearest engineering benchmarks and because they're the most immediate bottleneck for existing customers.
On the CMU side, the Robotics Innovation Center in Hazelwood Green will provide real-world test environments that most university robotics programs don't have access to. The facility was designed explicitly to bridge fundamental research and commercial deployment, making it a better analog for the messy conditions of an actual factory floor than a conventional lab setup.
Watch for Fujitsu announcements of pilot deployments in manufacturing or logistics (likely Japan-based customers first, given Fujitsu's enterprise relationships there) and for research papers from named CMU faculty within the next two to three academic quarters.
Frequently Asked Questions
What is Physical AI, exactly?
Physical AI refers to AI systems designed to act in the physical world, not just process text or images. This includes robots of all kinds (humanoid, industrial arms, mobile platforms) as well as AI systems that control manufacturing equipment, autonomous vehicles, or infrastructure. The key distinction from conventional AI is that physical AI must interact with real environments where sensor noise, physics, and unpredictability are unavoidable constraints.
What is the Fujitsu Kozuchi Physical OS?
Kozuchi Physical OS is Fujitsu's platform for coordinating robots, sensors, and physical systems across a unified cloud-to-edge infrastructure. It combines what Fujitsu calls brain intelligence (task learning and adaptability) with spatial intelligence (real-time environmental awareness) to enable multiple robots to work together on operational tasks. Research from the Fujitsu-CMU center is slated to be incorporated into Kozuchi starting in fiscal year 2026.
Why does Carnegie Mellon matter for robotics specifically?
CMU's School of Computer Science houses one of the world's top robotics research programs, with faculty working across the full stack from perception algorithms to hardware design to ethics. The university's Robotics Innovation Center, which opened in February 2026, provides 14,000 square meters of real-world test space. Alumni from CMU's robotics and ML programs are spread across most major robot companies in the US and internationally.
How is this different from companies like Figure AI or Agility Robotics?
Figure AI and Agility Robotics are robot hardware companies: they design, manufacture, and deploy physical humanoid robots. Fujitsu is positioning itself as the infrastructure layer above the hardware, the platform that coordinates robots from multiple vendors in real facilities and connects them to business systems. If Kozuchi succeeds, a manufacturer could deploy robots from different vendors and manage them through a single Fujitsu platform, rather than needing a separate software stack for each robot type.
When will commercial products from this partnership be available?
Fujitsu has committed to integrating center research outputs into Kozuchi Physical OS starting in fiscal year 2026 (April 2026 to March 2027). This suggests early commercial integrations within 12 months of the center's launch. Full product offerings built on this research would likely follow in 2027 and beyond, depending on which research tracks mature fastest.
A Different Kind of Physical AI Bet
Most of the attention in physical AI goes to the robot companies: who raised the most money, which platform can carry more weight, whose demo looked most convincing. The Fujitsu-CMU center is a bet on a different layer of the stack. If physical AI is going to scale across industries the way cloud computing did, someone has to build the infrastructure that makes it interoperable, manageable, and reliable enough for enterprises to actually stake operations on it. That's what Kozuchi is designed to be, and that's what this research center is designed to make technically credible.
The research output, particularly from CMU faculty with strong publication track records, will matter beyond Fujitsu's own products. Techniques developed here for multi-robot coordination, sim-to-real transfer, and human-robot collaboration will likely appear in the broader field within a year or two. That's the nature of academic partnerships: the commercial application goes to the funder, the knowledge goes to everyone.
The Bottom Line: Fujitsu and Carnegie Mellon aren't building another humanoid robot. They're building the infrastructure layer that robot deployments at scale will need, and they're using one of the world's top robotics schools to make sure it's technically defensible.
The 12-month milestone to watch is whether Fujitsu announces Kozuchi-based deployments with real manufacturing customers. That will be the signal that the research center's output is translating into the commercial pipeline it was built to feed.