evaluating-cosmos-policy

To evaluate NVIDIA Cosmos Policy in LIBERO and RoboCasa, you run headless GPU benchmarks using this skill to quantify deployment performance, measure latency, and ensure reproducible results across simulated robotic environments.

Running headless Cosmos Policy benchmarks requires PyTorch, MuJoCo, Robosuite, the Robocasa fork, transformers, and the Cosmos Policy repositories to execute end-to-end evaluations and generate latency profiles.

Yes, this skill supports latency profiling for robotics policies by utilizing headless EGL rendering to quantify performance metrics on GPU clusters or local hardware during LIBERO and RoboCasa evaluation workflows.

Yes, the skill provides reproducibility guidance and deployment notes for RoboCasa evaluation workflows, allowing researchers to reliably benchmark Cosmos Policy behavior and validate results across multiple simulation experiments.

The best way to benchmark a Cosmos Policy checkpoint is to run the LIBERO evaluation workflow in a headless environment, which quickly validates policy behavior and quantifies performance metrics in simulated robotics tasks.