NVIDIA has launched Cosmos Policy, a cutting-edge robot control system that builds on the Cosmos Predict-2 model. This policy enhances robot manipulation by directly encoding actions and future states, demonstrating superior performance on LIBERO and RoboCasa benchmarks. It works by fine-tuning a model that predicts future frames, enabling the robot to make decisions based on visual inputs without the need for separate perception and control networks. Instead, it utilizes a unified approach that treats robot actions and observations similarly to video frames, streamlining the learning process. Cosmos Policy shows significant advantages in robotic tasks, outperforming traditional methods. In testing on the LIBERO benchmark, it achieved an average success rate of 98.5%, surpassing various other policies, including diffusion-based and vision-language-action models. On RoboCasa, it also demonstrated improved generalization in household manipulation tasks, achieving a 67.1% success rate with just 50 training demonstrations per task. This performance underscores the efficiency gained from initializing with Cosmos Predict, which models temporal dynamics and physical interactions. The policy can function as a direct action generator or as part of a planning system, where it evaluates multiple potential actions before execution. In real-world tests, Cosmos Policy has been effective in managing complex bimanual tasks with the ALOHA robot platform. To further engage the robotics community, NVIDIA is hosting the Cosmos Cookoff, a hackathon aimed at fostering innovation and collaboration around these new models. The event runs from January 29 to February 26, offering cash prizes and valuable hardware incentives for participants.