JAPAN’S INDUSTRY. YOUR PERSPECTIVE.
JAPAN MANUFACTURINGINSIGHTINDIA EDITION
OUR PARTNERS · SPONSORED PROFILE
JIMTOF 2026 — JAPAN INTERNATIONAL MACHINE TOOL FAIR
PARTNER

JIMTOF 2026

READ MORE ↗
Technology & Products

From Hardware to Factory Data: Why FANUC, Yaskawa and Kawasaki Are Joining Forces on Physical AI

Three rival robot makers are aligning how they collect factory-floor data, under a shared format. Kawasaki Heavy Industries, FANUC and Yaskawa Electric have teamed up to build a physical-AI foundation for the factory floor. In the first year the project plans to gather 5,000…

From Hardware to Factory Data: Why FANUC, Yaskawa and Kawasaki Are Joining Forces on Physical AI
NVIDIA

Three rival robot makers are aligning how they collect factory-floor data, under a shared format. Kawasaki Heavy Industries, FANUC and Yaskawa Electric have teamed up to build a physical-AI foundation for the factory floor. In the first year the project plans to gather 5,000 hours of on-site data, backed by up to ¥2 billion (about US$13 million) in state support. The aim is not a new product but the common data specifications and collection method that will underpin the AI itself. The move signals that the basis of competition in industrial robotics is shifting — from the performance of the machine to the volume and quality of data.

Key points

  • Kawasaki Heavy Industries, FANUC and Yaskawa Electric are collaborating to build a physical-AI dataset for the factory floor. The project has been selected for "GENIAC", a generative-AI development programme run by Japan's Ministry of Economy, Trade and Industry (METI) and NEDO.
  • They will build the foundations of a "VTLA model" that handles vision, touch, language and action together, collecting 5,000 hours of video, tactile and motion data from factories in the first year.
  • The programme runs from August 2026 to July 2027, with up to ¥2 billion (about US$13 million) in support from NEDO. Osaka University, tactile-sensor start-up FingerVision and AI firm ABEJA are also taking part.
  • What the three standardise is the data specification and collection method. Each keeps its own shop-floor data, control technology and hardware. Where cooperation ends and competition begins is the real question.

Top robot makers team up — on data

In July 2026, Kawasaki Heavy Industries, FANUC and Yaskawa Electric announced a partnership to develop a foundation model for physical AI on the factory floor. The project has been selected for "GENIAC", a generative-AI development programme run by Japan's Ministry of Economy, Trade and Industry (METI) and NEDO (the New Energy and Industrial Technology Development Organization). It runs from August 2026 to July 2027, with up to ¥2 billion (about US$13 million) in support from NEDO.

Physical AI refers to AI that acts in the physical world. Rather than generating text or images on a screen, it moves a robot's arms and hands to grip, assemble and carry — performing real work.

All three rank among the world's leading makers of industrial robots, and normally compete for the same customers. What makes this line-up unusual is that the object of their cooperation is neither a finished product nor a component technology, but data.

The roles are divided. The three makers handle the standardisation of the data specification and collection infrastructure; Osaka University is the project's lead organisation, and ABEJA, commissioned by the university, advances the underlying technology. FingerVision, a start-up, brings the tactile-sensing know-how — competitors, a university and a start-up, brought together around one shared body of data.

Why now — the deciding factor has moved from hardware to data

For decades, competition in industrial robots turned on hardware: positioning accuracy, payload, speed and durability. Japan's strength rests on the depth of that mechanical craftsmanship.

Physical AI asks for something else. Told to "pick that up", a robot must judge the situation and move its hand, even for a part it has never seen. That kind of general behaviour cannot be achieved by precise hardware alone; it needs an AI trained on vast amounts of task data. The same rule of thumb that governs large language models applies here: the volume and diversity of training data largely determine how capable the model becomes.

The problem is that this data has been fragmented across makers. Recording formats, and what is measured and how, differ from one company to the next, so data collected by one company cannot readily be used to train another's models. Align the format, and data gathered across different companies and processes can be reused together. The range available for training widens, and development speeds up. Standardisation is the groundwork that makes this mutual use possible.

That is why competitors have come together. Diverse shop-floor data cannot be gathered by one company alone. Collecting compatible data from different sectors and processes under a common format grows a usable dataset faster than any single firm could on its own. Teaming up became the rational move once the contest had shifted to data.

VTLA aims to turn "teaching" into "learning"

At the core of the effort is the "VTLA (Vision-Tactile-Language-Action) model" — an AI model that handles vision, touch, language and action as one. It aims to understand a spoken instruction, see the object through a camera, and move while adjusting its force by the feel at its fingertips, all within a single model.

The distinctive feature is the inclusion of touch. Conventional industrial robots have worked mainly from vision and position data. But for delicate handwork — pinching a soft part, inserting a component into a slightly misaligned hole — the ability to sense slippage and contact force is essential. FingerVision's technology fills that gap.

What VTLA seeks to change is how work is taught to a robot in the first place. Today's shop floors rely on "teaching": a setup task in which each motion is programmed by hand, and a person must re-enter coordinates and steps every time the object changes. That labour has been a barrier to automation in high-mix, low-volume plants and in factories short of workers. If VTLA reaches practical use, the relationship inverts: instead of being taught in detail, the AI learns behaviour from accumulated data and can respond, to a degree, to work it has never seen. The premise of automation shifts from "teaching" to "learning".

Where cooperation ends and competition begins

The essence of this three-way tie-up lies in the boundary between what the companies share and what they continue to compete over.

What they standardise is the data specification and the way it is collected — the rules for how data is recorded and stored. What stays with each company is the shop-floor data actually gathered in its own plants, its robot control technology and hardware, and its final products and services. The base rules are shared; the contest happens on top of them.

The real question in this arrangement is who holds the levers. Who owns the data collected under the common format, and who may use it for training? Who decides what goes into the standard? Whoever holds that ground secures a position close to the foundations of the physical-AI era. Setting a standard is also a contest over the right to decide its details.

Viewed in a global context, the move runs alongside a race for data. US firm NVIDIA is extending its footing in physical AI by building a strong position in the computing platforms and simulation environments used to develop robots. The three Japanese makers are, separately, in talks with Fujitsu and NVIDIA on social implementation, and will rely partly on overseas technology for computing. Yet on the single point of a data standard for the shop floor, they are moving to hold a home-grown foundation themselves. Borrow part of the platform, but do not hand over control of the data — that is how this VTLA effort is best read.

Industry ripple — the work of integrators and parts makers shifts

If a general-purpose dataset and foundation model take hold, the effects spread beyond the robot makers.

System integrators (SIers) are likely to see the biggest change. Installing robots and setting them up through teaching has been a core source of revenue for integrators. As learning takes over more of how motion is acquired, the work of teaching each motion by hand shrinks. In its place, the centre of gravity moves toward data-centred support — how to collect data suited to a site, and how to adapt a model to a plant's own processes. From setters of equipment to enablers of data use: the role is being rewritten.

For parts and sensor makers, an opening appears. VTLA's emphasis on touch could lift demand for force and tactile sensors and the mechanical components behind them. The accuracy of physical AI depends not only on the AI but on the quality of the sensors that read reality — an area where Japanese parts makers are strong. Even so, the locus of value may move from selling a sensor as a standalone unit to how the data it captures is fed into the platform.

What to watch — who decides the right to use the data

The central question is who controls the data and the standard. Who owns the data collected under the common format, and how far may each company use it? How is the right to decide the standard's contents divided among the three? If this stays vague, the rhetoric of standardisation may continue without ever growing into a real standard. The operating rules the project publishes will be the first thing to judge it by.

Scale matters too. Whether the first year's 5,000 hours is enough to train a general-purpose model is hard to assess from outside. Set against the vastly greater volumes of text that large language models have learned from, 5,000 hours looks like a starting point. Shop-floor data is dense per hour, so a simple comparison by time does not hold — but how much can be added from the second year on will decide the model's real capability.

And there is the question of whether a domestic standard can travel. Japan's makers took the world on hardware; can they hold the initiative on the new ground of data? This one-year project is the first test of that.

OUR PARTNERS

JAPANESE INDUSTRY, CONNECTED.
VIEW ALL PARTNERS ↗