Into the Omniverse: How Open World Models Push the Frontier of Physical AI

NVIDIA just shipped Cosmos 3, and it's the clearest sign yet that the company wants to own the plumbing under every robot, autonomous vehicle and smart camera on the planet — without necessarily owning the models that run on top of it. The pitch is simple: physical AI needs to understand cause and effect in the real world, not just recognize objects in a photo. World models are how you teach a system what happens next, whether that's a forklift swerving around a spilled pallet or a delivery robot navigating rain-slicked pavement it's never seen in training data.

Cosmos 3 comes in three sizes. Super, at 64 billion parameters, handles high-fidelity world modeling for teams that need maximum accuracy. Nano, at 16 billion, is built for efficient reasoning and post-training. Edge, at just 4 billion parameters, is small enough to run directly on NVIDIA Jetson Thor hardware or RTX GPUs, meaning a warehouse robot could do vision reasoning and policy decisions on-device rather than phoning home to a data center. That range matters more than the benchmark scores, frankly, because physical AI deployments live and die on latency and power budgets, not leaderboard bragging rights.

Still, the numbers are hard to ignore. Cosmos 3 tops Artificial Analysis for open-weights text-to-image and image-to-video generation, leads PAI-Bench for world generation, and takes the No. 1 spot on RoboLab for robot policy work. It also ranks first among open models on VANTAGE-Bench for vision understanding. NVIDIA is releasing the whole family under the Linux Foundation's OpenMDW 1.1 license, which is the part that actually matters for adoption — teams can pull the weights, retrain on their own robot's sensor data, and deploy without licensing headaches.

The adoption list already reads like a who's-who of hardware makers: Doosan Robotics, LG, Samsung and Skild AI are building on it for robotics, Li Auto and Xiaomi for autonomous vehicles, and companies like Milestone Systems for vision AI in industrial settings. NVIDIA also just expanded its Cosmos Coalition into Japan, roping in manufacturing and robotics firms to build out world models for factories and logistics. This is NVIDIA doing what it does best — not necessarily inventing the smartest model, but making sure every serious physical AI project ends up running on its stack, whether that's Jetson chips, Omniverse simulation tools, or now Cosmos itself.

The Omniverse and OpenUSD tooling that ships alongside Cosmos is arguably the less flashy but more durable part of this announcement. Building simulation environments to test a robot before it touches the real world used to mean duplicating work every time you changed a sensor or a lighting condition. NVIDIA's framing here is that Cosmos plus Omniverse removes that duplication, letting teams generate synthetic training data and validate behavior in simulation before anything ships. Whether that promise holds up at scale is the thing to watch over the next year, not the benchmark charts.