vla-robustness-kit 0.1.0

Simulator-light diagnostics for VLA policy robustness.

vla-robustness-kit

vla-robustness-kit generates lightweight robustness diagnostics for vision-language-action policies. It perturbs task instructions, camera and lighting metadata, embodiment metadata, and task phase order, then reports failure clusters and follow-up data collection recommendations.

The default runner uses a deterministic mock policy. No simulator, GPU, or real robot is required for local tests.

These reports are diagnostics, not robot safety claims.

Quick start

python -m pip install -e .
vla-robustness-kit run examples/mock_episode_set --policy mock --out report.md

Output

The report includes:

• Perturbation family pass rates.
• Failure clusters.
• Recommendations for additional data collection.
• A clear diagnostic-only warning.