evaluating-cosmos-policy

To run Cosmos Policy evaluation on LIBERO, you need to configure benchmark flags, repeatable seeds, and task-suite settings. This Skill prepares the environment and executes the official Cosmos Policy LIBERO entrypoint for both smoke tests and full benchmark sweeps.

Headless GPU execution during robot evaluation requires EGL rendering environment variables, dataset statistics paths, and checkpoint paths. You must set these alongside explicit benchmark flags to properly validate manipulation benchmark results on local or cluster machines.

Yes, you can run RoboCasa single-task or multi-task evaluations using this Skill. It configures the required object split, dataset statistics, and EGL rendering setup to launch RoboCasa runs with the correct environment variables and reproducible seeds.

Yes, Cosmos Policy evaluation supports inference profiling and latency checks. This Skill helps you check inference latency, verify results, and troubleshoot common issues like GPU mismatches, asset setup problems, or host-runtime incompatibilities during evaluation.

Robot manipulation benchmark results may fail to validate due to missing reproducible seeds, incorrect dataset statistics paths, or GPU mismatches. This Skill ensures explicit benchmark flags and EGL rendering settings are correctly applied to validate evaluation results.