Nvidia has unveiled Cosmos 3 Edge, a new AI model designed to help robots and intelligent cameras understand their surroundings and make decisions locally.
The company also expanded its physical AI ecosystem in Japan, where major robotics, manufacturing and technology groups plan to build on Nvidia’s models, computing platforms and simulation tools. The move shows how Nvidia wants to become more than the chip supplier behind generative AI.
According to Nvidia’s official Cosmos 3 Edge announcement, the model will support on-device vision reasoning and robot action generation on its Jetson platforms.
Cosmos 3 Edge puts more intelligence inside machines
Cosmos 3 Edge is a 4-billion-parameter model built on Nvidia Nemotron. It combines visual understanding with real-time reasoning, helping robots and vision AI systems decide what action to take after analysing their surroundings.
That could help a warehouse robot locate and collect a package, an inspection camera identify faulty equipment or an agricultural machine respond to changing field conditions. The model focuses on physical tasks rather than generating text for a chatbot.
The “Edge” name matters. Nvidia designed the model to run on computers close to the robot or camera, including its Jetson hardware, rather than relying entirely on a remote data centre.
Local processing can reduce delays and keep a machine responsive when connectivity becomes limited. It can also suit factories, farms and infrastructure sites where every movement may require an immediate decision.
| Feature | What it means |
| Model size | Four billion parameters |
| Primary role | Visual reasoning and robot action generation |
| Supported systems | Jetson, RTX GPUs and DGX systems |
| Customisation | Adaptable to specific robots, sensors and environments |
Japan’s industrial giants join the Cosmos ecosystem
Nvidia says AIRoA, FANUC, Fujitsu, Hitachi, Kawasaki Heavy Industries, Kubota, NEC, SoftBank, Sony and Yaskawa Electric are among the companies planning to join its expanded Cosmos Coalition in Japan.
The coalition gives members access to open models, datasets, data-curation libraries and development frameworks. Members can also contribute industry knowledge and training data to improve models for real workplaces.
This matters because a general-purpose robot model won’t automatically understand every production line, hospital, farm or construction site. Each environment has different equipment, safety risks and operating rules.
Fujitsu is separately exploring a collaborative control platform with FANUC, Yaskawa Electric and Kawasaki Heavy Industries. The companies plan to combine Cosmos with Nvidia Isaac, Omniverse NuRec libraries and the Newton physics engine.
The proposed platform could help developers create digital twins, train robots in simulation and test their behaviour before putting machines near workers. That may shorten development cycles while reducing the cost and risk of physical trials.
The companies haven’t announced a firm timeline for bringing these robots into daily life. However, they expect the first phase of the collaboration to begin later in 2026, according to AP’s report on Fujitsu and Japan’s physical AI initiative.
Nvidia is building a full stack for physical AI
We think the real story isn’t only the launch of another AI model. Nvidia is connecting computing hardware, AI models, robotics software and simulation into one physical AI stack.
Jetson provides computing inside machines. Cosmos gives those machines a way to understand and predict the physical world. Isaac supports robotics development, while Omniverse tools help companies simulate factories and other environments.
The strategy could make Nvidia deeply embedded in how companies build, train and operate intelligent machines. Once a business adopts several parts of the stack, moving to another provider may become difficult and expensive.
Japan gives Nvidia an especially strong testing ground. The country has world-leading robotics and manufacturing companies, but it also faces a serious labour shortage linked to its ageing population.
Reuters reported that government-backed Noetra plans to buy 27,500 Nvidia Rubin chips for physical AI development. Construction is scheduled to begin in April 2027, with operations planned for June 2028.
That infrastructure project connects directly with Memeburn’s breakdown of Japan’s plan to deploy 10 million AI robots. Japan isn’t treating physical AI as a collection of small experiments. It’s building models, computing infrastructure and industrial partnerships at national scale.
What this could mean for South Africa
For South African readers, physical AI may sound distant, but the potential applications are close to home. Mining, ports, farming, manufacturing, logistics and infrastructure inspection all involve repetitive, dangerous or time-sensitive physical work.
A machine that analyses its surroundings locally could inspect mining equipment, identify damaged rail infrastructure or monitor crops without sending every video frame to the cloud. Edge processing could prove especially useful in areas with unreliable connectivity.
But hardware alone won’t create a working ecosystem. South African businesses would need skilled engineers, local training data, reliable maintenance, clear safety standards and a strong business case.
Large mining, automotive and logistics groups may adopt these tools first because they can absorb the early costs. Smaller businesses may need cheaper robots, shared infrastructure or robotics-as-a-service models before physical AI becomes practical.
FAQs
What is Nvidia Cosmos 3 Edge?
Cosmos 3 Edge is a 4-billion-parameter AI model for robots and vision systems. It helps machines interpret visual information, reason in real time and generate actions locally. Nvidia designed it to run on edge hardware, including Jetson platforms.
What does physical AI mean?
Physical AI refers to artificial intelligence that controls or guides machines in the real world. Examples include robots, autonomous vehicles, intelligent cameras and smart industrial equipment. These systems must understand physical environments, not only text or images on a screen.
Which Japanese companies are working with Nvidia?
Nvidia named companies including FANUC, Fujitsu, Hitachi, Kawasaki Heavy Industries, Kubota, NEC, SoftBank, Sony and Yaskawa Electric. Many plan to join the Cosmos Coalition, while others are building systems with Nvidia’s wider physical AI stack. Their work covers robotics, manufacturing, mobility, infrastructure and smart spaces.
The post Nvidia Unveils Cosmos 3 Edge as Japan Expands Physical AI in 2026 appeared first on Memeburn.