evaluating-cosmos-policy

To evaluate NVIDIA Cosmos Policy in MuJoCo robotic simulation, run the provided Python scripts targeting LIBERO or RoboCasa environments to execute tasks and collect performance data.

Yes, evaluating Cosmos Policy requires a GPU and supports headless rendering with EGL. This setup allows you to run robotic simulations and collect performance data on local clusters.

You need torch, mujoco, robosuite, robocasa, transformers, and the cosmos-policy package installed in your Python environment to run the evaluation scripts.

Yes, you can measure inference latency of a robotic policy during simulation. The Skill provides specific profiling tools to collect latency metrics for the policy on GPU systems.

The best way to validate robotic manipulation policies before physical deployment is running simulation evaluations on environments like LIBERO and RoboCasa to collect performance data.

You should use LIBERO and RoboCasa for robotic simulation evaluation because they provide standardized robotic manipulation tasks to measure and validate policy performance accurately.