Panasonic Launches KMC to Turn Factory Know-How Into Shared AI Models

By Nathan Brooks

Panasonic Holdings announced on November 11, 2026 a Manufacturing Commons initiative, called KMC, to build AI and robotics reference models in Japan through industry, government and academic collaboration — and deploy them across real workplaces.

The project is the company's vehicle for what it calls field-driven Physical AI: artificial intelligence applied to physical work rather than confined to digital environments. Panasonic wants the initiative to improve research, design and production processes as AI spreads beyond screens and servers.

How the learning loop works

The model Panasonic describes is a continuous two-way exchange. AI first learns from workers' experience, skills and judgments. Workers then use the analysis and proposals the AI produces, and the new knowledge and data generated on the job flows back into the system for further training. The loop is meant to run across research, design and manufacturing settings simultaneously.

Under KMC, organizations would share manufacturing technologies, knowledge and human connections across corporate and organizational boundaries. Panasonic's stated ambition is blunt: pool Japanese manufacturing capability into a globally competitive production model. For practitioners, that means proprietary process know-how becomes a shared asset — with the competitive question shifting from who holds the technique to who deploys it fastest.

The announcement ties KMC to the Technology Division's 2040 technology vision, which pairs technology development with partner co-creation. That vision names AI and cyber-physical systems — connecting digital analysis to physical operations — as common technologies spanning robotics, sensing and model-based research.

What does this change on the factory floor?

The strategy draws on more than a century of Panasonic manufacturing experience in materials, devices, sensing, control and production processes. Crucially, it also leans on skilled workers' methods for turning physical phenomena into usable data — grounding the AI effort in factory operations, not just software.

Two concrete programs show where the money and engineering are going:

• Glass-type perovskite solar cells. Panasonic's materials technology plus inkjet coating, laser processing and sealing techniques borrowed from display manufacturing enable power-generating glass for windows and walls. The company will apply informatics and AI to materials development and process-condition optimization. Sales of pre-mass-production units began in September 2026; full mass production is targeted for April 2029 at the Nagahama development site in Shiga Prefecture.
• Advanced semiconductor back-end equipment. The work targets three-dimensional chip stacking and other next-generation manufacturing uses, combining plasma surface modification, precision bonding, printing, inspection and transport technologies. A newly launched AI Semiconductor Packaging Project unites research and business development, with AI analysis of process data and inspection images intended to support quality prediction, process-condition searches and equipment specifications.

T-Rex shows the robotics payoff

A separate research release dated November 5, 2026 illustrates the kind of work behind the strategy. Panasonic Holdings and Panasonic R&D Company of America developed T-Rex with researchers from UC Berkeley, NVIDIA, Stanford University, Sapienza University of Rome and ItalAI.

T-Rex integrates fingertip tactile information into a vision-language-action model, which converts images and natural-language instructions into robot actions and corrects movement as contact conditions change. The engineering trick is speed layering: a vision-and-language module generates broad actions at roughly 5 Hz — five updates per second — while a tactile module corrects motion at about 20 Hz. That lets the robot respond to slipping or over-pushing without rerunning slower image recognition for every adjustment. The T-Rex paper documents a 100-hour tactile-rich dataset and the variable-rate Mixture-of-Transformers architecture that handles the different processing rates.

The results are hard to ignore. Panasonic reports a 65% average success rate across 12 real-world contact tasks, versus 35% for its strongest comparison method and 42% when tactile information was stripped out. The tasks — page turning, transferring eggs, applying toothpaste, unlocking locks, removing cards and screwing in lightbulbs — all demand delicate force control and handling of deformable objects.

Panasonic sees future uses in inserting, removing and tightening components in manufacturing and handling flexible or fragile items in logistics, aiming to reduce worker burden and ease labor shortages. KMC's central test is whether shared manufacturing knowledge can convert research results like T-Rex into reusable reference models across production and logistics workplaces — the difference between an impressive demo and a repeatable industrial asset.