英伟达Cosmos 3：统一理解、模拟与行动的物理AI模型

Nvidia's Cosmos 3: 1 model that can understand, simulate, and act across many physical AI tasks.

It treats action as a first-class language of the world.

Most AI models look at reality from the outside: images become captions, videos become descriptions, and motion becomes something to label after the fact.

Cosmos 3 tries to collapse that distance by putting language, image, video, audio, and action into one shared system, so a robot can connect what it sees with what might happen next and what it should do.

A home robot cannot simply recognize a plate, a table, and a human instruction, because the useful question is what changes when it moves, grasps, slips, bumps, or waits.

That is why the paper’s action-token design matters: it turns movement into something the model can condition on, infer from video, or generate alongside a future scene.

Link – arxiv. org/abs/2606.02800

Title: "Cosmos 3: Omnimodal World Models for Physical AI"