NVIDIA Cosmos Policy is a new robot control and planning policy that fine-tunes the Cosmos Predictdirectly on robot demonstrations without any complex multi-stage pipelines. It injects actions, future states, and value estimates into the model’s latent sequence, so a single diffusion model can see, imagine, and decide.
With one unified model, you can:
- Run direct visuomotor control by sampling action chunks from images.
- Do model-based planning by rolling out candidate futures and picking the highest-value trajectories.
Cosmos Policy achieves state-of-the-art success on LIBERO and RoboCasa and transfers to real ALOHA bimanual manipulation, especially when combined with planning.
🍳 Want to get hands-on?
Explore Cosmos Policy alongside Cosmos Reason and Cookbook recipes in the ongoing Cosmos Cookoff 👉 Register: NVIDIA Cosmos Cookoff · Luma