How a Robot Tests an Action Before It Moves

How can a robot test an action without moving? NVIDIA's Ming-Yu Liu explains how a world model predicts the outcome first. Cosmos 3 learns three abilities together: forward dynamics (what happens after an action), inverse dynamics (which action caused a change) and policy (what the robot should do next). The results in the paper show that each one helps the others. The payoff is a closed-loop simulator. Instead of setting up a thousand kitchens, developers can check a robot policy in simulation