skills.rest
box755
box755COMMUNITY
Explore all skills by this author →
GitHub

evaluating-cosmos-policy

Evaluate NVIDIA Cosmos Policy on LIBERO and RoboCasa simulations.

Updated Aug 23, 2026
Search Tags:#cosmos policy#egl rendering#robot manipulation#simulation benchmark#vla evaluation#latency profiling#libero robocasa
One-click install
npx skills add https://github.com/box755/simlens-research --skill evaluating-cosmos-policy-box755
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: evaluating-cosmos-policy
Source: https://github.com/box755/simlens-research/tree/main/skill-packs/AI-Research-SKILLs/18-multimodal/cosmos-policy
Command: npx skills add https://github.com/box755/simlens-research --skill evaluating-cosmos-policy-box755

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires torch>=2.1.0, mujoco>=3.0.0, robosuite>=1.4.0, robocasa @ git+https://github.com/moojink/robocasa-cosmos-policy.git, transformers>=4.40.0, cosmos-policy @ git+https://github.com/NVlabs/cosmos-policy.git, and includes references (resource) components.

What problem does it solve?

This Skill streamlines evaluating NVIDIA Cosmos Policy on the LIBERO and RoboCasa simulation benchmarks so you can reliably measure success rates and inference performance without manual setup chaos.

Core Features & Use Cases

  • End-to-end evaluation workflows: Runs official public eval entrypoints for both LIBERO (cosmos_policy.experiments.robot.libero.run_libero_eval) and RoboCasa (cosmos_policy.experiments.robot.robocasa.run_robocasa_eval) using the correct checkpoint/config flags.
  • Headless GPU rendering setup (EGL): Provides a reproducible checklist for aligning CUDA_VISIBLE_DEVICES, MUJOCO_EGL_DEVICE_ID, MUJOCO_GL=egl, and PYOPENGL_PLATFORM=egl on GPU clusters.
  • Smoke-to-benchmark scaling: Guides a fast smoke run (few trials) followed by multi-trial suite runs for consistent comparisons (seed/task/trials).
  • In-depth command variants: Includes reference matrices for common interactive, local GPU, and batch workflows.

Quick Start

Run a minimal LIBERO smoke evaluation by executing the official module cosmos_policy.experiments.robot.libero.run_libero_eval with the provided Cosmos Policy checkpoint and dataset/statistics arguments.

Frequently Asked Questions about evaluating-cosmos-policy

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I evaluate a Cosmos Policy checkpoint on the LIBERO simulation benchmark?▼

Configure headless EGL rendering by aligning `MUJOCO_GL=egl`, `PYOPENGL_PLATFORM=egl`, `MUJOCO_EGL_DEVICE_ID`, and `CUDA_VISIBLE_DEVICES` on your GPU cluster. The skill provides a reproducible checklist for aligning these environment variables for deterministic MuJoCo rendering.

What dependencies do I need to install before benchmarking VLA robot manipulation policies?▼

You need `torch>=2.1.0`, `mujoco>=3.0.0`, `robosuite>=1.4.0`, `transformers>=4.40.0`, plus the `robocasa` and `cosmos-policy` repositories. Proper GPU and EGL environment alignment is required for deterministic reproducibility during VLA robot manipulation benchmarking.

Can I run RoboCasa single-task and multi-task evaluations using this Cosmos Policy setup?▼

Yes, you can run RoboCasa single-task and multi-task evaluations by executing the `cosmos_policy.experiments.robot.robocasa.run_robocasa_eval` entrypoint. The skill supports smoke-to-full scaling and provides reference matrices for batch workflows across interactive, local GPU, and cluster execution scenarios.

How do I profile inference latency for VLA models in simulation benchmarks?▼

Execute the provided evaluation workflows on LIBERO and RoboCasa to measure both success rates and inference latency for VLA robot manipulation. The skill targets latency-oriented profiling by running repeatable benchmark suites with correct checkpoint and config flags.

Why does my MuJoCo simulation fail to render on a headless GPU cluster?▼

Headless MuJoCo rendering fails when EGL environment variables are misaligned. Ensure `CUDA_VISIBLE_DEVICES`, `MUJOCO_EGL_DEVICE_ID`, `MUJOCO_GL=egl`, and `PYOPENGL_PLATFORM=egl` are correctly configured to match your GPU cluster setup for deterministic reproducibility.

Related Skills

Explore More Education & Research Skills→
handsomelong922handsomelong922

evaluating-cosmos-policy

Evaluate NVIDIA Cosmos Policy in LIBERO and RoboCasa simulation environments.

Community
Advanced
KappTech88KappTech88

evaluating-cosmos-policy

Evaluate NVIDIA Cosmos Policy on LIBERO and RoboCasa simulation environments.

Community
Advanced
hhhi21ghhhi21g

evaluating-cosmos-policy

Evaluate NVIDIA Cosmos Policy in LIBERO and RoboCasa simulation environments.

Community
Advanced
qcmuuqcmuu

evaluating-cosmos-policy

Evaluate NVIDIA Cosmos Policy on LIBERO and RoboCasa simulation environments.

Community
Advanced