evaluating-cosmos-policy

You can benchmark NVIDIA Cosmos Policy in LIBERO and RoboCasa by running the cosmos_policy module with explicit evaluation entry points and configuration files to generate reproducible performance metrics.

Yes, headless GPU evaluation is supported using EGL rendering, allowing scalable benchmarking across multiple LIBERO task suites and RoboCasa tasks without requiring a display server.

Benchmarking requires the Cosmos Policy repository, LIBERO and RoboCasa simulation environments, and explicit evaluation entry points via the cosmos_policy module with appropriate configuration files.

Latency profiling is handled through evaluation workflows applying reproducible configurations and command surfaces to ensure consistent latency measurements across LIBERO and RoboCasa runs.

Yes, the workflow supports reproducible benchmarks across multiple LIBERO task suites and RoboCasa tasks by applying reproducible configurations and command surfaces to ensure consistent results across runs.