NVIDIA on August 27, 2026 announced a major refresh of its Cosmos world foundation models (WFMs) and physical-AI data tools — including an open, fully customizable reasoning model, a new Cosmos Reason 2 vision-language model, and two Omniverse-Cosmos blueprints designed to churn out post-training data for robots and autonomous vehicles at scale. Agility Robotics, Figure AI, Foretellix, Skild AI and Uber are among the first named adopters.
Cosmos Reason 2 And Controllable World Generation
The headline release gives developers unprecedented control over what Cosmos generates — steering scenes, physics parameters, agent behaviors and edge cases rather than relying on stochastic sampling. Cosmos Reason 2 is a new reasoning vision-language model tuned for embodied agents that need to see, understand, and act in the physical world with higher accuracy. NVIDIA CEO Jensen Huang framed the launch as a step-change: "Just as large language models revolutionized generative and agentic AI, Cosmos world foundation models are a breakthrough for physical AI."
Synthetic Data For Robots And Robotaxis
The two new blueprints — built on the NVIDIA Omniverse and Cosmos platforms — act as controllable synthetic-data generation engines aimed squarely at post-training robots and autonomous-vehicle stacks. That plugs directly into how Japan's Cosmos coalition and Boston Dynamics' Atlas program are already using WFM-driven simulation to compress data-collection cycles.
Launch Adopters
Agility Robotics will use Cosmos to expand training data for its Digit humanoid; Figure AI will feed the models into its Helix policy training pipeline; Skild AI plans to accelerate its omni-bodied brain roadmap; Foretellix is folding Cosmos scenarios into its AV verification suite; and Uber is exploring the new blueprints for its robotaxi partner networks. The launch cements NVIDIA's foothold in the physical-AI training stack even as merchant GPU competition intensifies — and gives every robotics lab a new default choice for world-model pretraining.
Reporting based on coverage from NVIDIA Newsroom, NVIDIA Blogs and Seeking Alpha.