Nvidia partners with Japan's robotics industry for its open world models.

      Nvidia has brought together a majority of Japan's industrial robotics sector into the Cosmos Coalition, its initiative for open-world models aimed at advancing its physical AI stack, coinciding with Jensen Huang's visit to Tokyo.

      The coalition includes twenty-two companies: AIRoA, classmethod, Enactic, FANUC, Fujitsu, GROOVE X, Hitachi, Honda R&D, Kawasaki Heavy Industries, Kubota, Mitsui & Co., Mitsubishi Corp., Mujin, NEC, Preferred Networks, SoftBank Corp., Sony Group, Telexistence, TIER IV, TRON K.K., Turing, and Yaskawa Electric. Nvidia states that all of these companies “intend to join,” but there are no disclosed binding commitments or financial specifics involved.

      This announcement is significant, particularly because FANUC and Yaskawa are the largest industrial robot manufacturers globally based on installed base, and they have historically utilized proprietary control stacks. Their intention to join an initiative centered on someone else’s open models is noteworthy. This strategy mirrors Nvidia's previous collaboration with Hyundai regarding the Atlas humanoid and its competition with Tesla's vertically integrated Optimus program: provide the technology stack while allowing the industrial partner to manage deployment.

      At the technical core is Cosmos 3 Edge, a model comprising four billion parameters that operates on Nvidia’s Nemotron series and runs on Jetson edge hardware instead of data centers. Nvidia claims that developers can customize it for a particular robot, vehicle, or sensor setup in roughly a day, and it will function across RTX GPUs, DGX systems, and the newly launched Jetson T2000 and T3000 modules.

      Fujitsu is spearheading a significant aspect of the program, which involves a collaborative control platform being developed in conjunction with FANUC, Yaskawa, and Kawasaki, intended to connect digital and physical operations across various industrial sectors. This platform is based on Cosmos world foundation models, the Isaac robotics platform, Omniverse NuRec libraries, and the Newton physics engine, aimed at managing digital twins, robot learning, and simulation-to-real validation before anything enters the factory.

      Additionally, the applications being discussed become quite specific. Kubota is exploring the use of Cosmos for autonomous agriculture, while Enactic is refining Nvidia’s Isaac GR00T model for semi-humanoid robots in elder care. Shimizu Corporation is employing Metropolis for construction-site safety, GROOVE X is developing Jetson-powered companion robots, and Kawasaki is implementing the technology in healthcare, shipbuilding, transportation, aerospace, and energy sectors.

      “Japan invented modern manufacturing,” Huang remarked in the announcement. “Now, it has the opportunity to reinvent it for the era of intelligent industries.” He described physical AI as “a once-in-a-generation opportunity” for the country.

      A second announcement on the same day addressed the language aspect. The Institute of Science Tokyo created its Swallow open models using Nemotron datasets, while SB Intuitions, SoftBank’s generative AI subsidiary, trained its Sarashina series with Nemotron libraries. The Sarashina3 mini has been adopted by Japan’s Digital Agency. SoftBank Corp. has implemented a Large Telecom Model for autonomous network management, and NTT DATA has utilized Nvidia’s Japanese personas dataset to enhance training for its tsuzumi 2 model.

      Throughout the messaging, the emphasis is on sovereignty. “Every nation and every company should own and control its intelligence infrastructure,” Huang emphasized. “Open models make that achievable.” This argument has varied implications depending on one's perspective, as the open models in question function optimally on hardware sold by Nvidia. The same assertion has been made by Nvidia in Europe, where it recently launched 35 new AI supercomputers.

      Japan’s interest in these developments is more pragmatic than the rhetoric implies. Nvidia’s announcement cites the aging population and workforce transition as a motivating factor, which is a subtle way of addressing the labor shortage issue Japanese manufacturers have faced for a decade.

      Meanwhile, Sakana AI is integrating Nemotron into its Fugu model-routing platform, which selects the best model for each task rather than relying on a single option.

      What remains unresolved is whether these world models actually accelerate the process from demonstration to deployment. This question has lingered unanswered amid China's bustling robot industry, and a coalition of intent does not provide clarity on the matter.