NVIDIA keeps pushing the boundaries of robotics by tying world models (WFMs) directly to physical control. Now they introduce Cosmos Policy, a control policy that post-trains Cosmos Predict-2 so the same model that understands how a scene evolves also generates actions, predicts future states, and estimates reward returns. What does that mean for you if you're building robots or researching physical AI? Fewer separate modules to stitch together, smoother temporal coherence, and better generalization — so a home robot, for example, can not only see a cup start to tip but plan the sequence of moves to catch it and judge whether that effort is worth the reward. Less engineering overhead, more unified behavior.