NVIDIA Drops Cosmos Policy for Robot Control
NVIDIA expands its Cosmos world foundation models with a new policy built specifically for robot manipulation tasks.
NVIDIA just gave its Cosmos platform a serious robotics upgrade. The company introduced Cosmos Policy, a new robot control policy layered on top of its existing Cosmos Predict-2 world foundation model.
The key idea: take a world model that already understands physics and spatial reasoning, then post-train it specifically for manipulation tasks. Instead of building robot control from scratch, Cosmos Policy leverages the foundation model's existing understanding of how the world works.
It's a move that blurs the line between generative AI and physical robotics. World foundation models were originally designed to simulate and predict environments. Now NVIDIA is turning that predictive power into actual robot behavior.
This positions Cosmos as more than a simulation tool — it's becoming a full-stack robotics AI pipeline, from world understanding to physical action.