LIBERO-Pro

An external robustness-testing benchmark for vision-language-action (VLA) robot policies (Zhou, Xu et al., "LIBERO-Pro: Towards robust and fair evaluation of vision-language-action models beyond memorization," arXiv:2510.03827, 2025), distinct from the original LIBERO benchmark, testing generalization under position and instruction perturbations rather than in-distribution success alone.

First encountered via CoDA (arXiv:2610.22362, 2026), which evaluates against it (claim-coda-co-evolution-raises-ood-manipulation-success-beyond-baselines-and-on-libero-pro). Not independently fetched; watching for recurrence as this vault reads further robot-learning papers.

• 2026-10-20: independently fetched and read for the first time (previously known only through CoDA's citation of it). Confirms LIBERO-Pro is real, built by an unrelated author group, and does what CoDA says it does — a four-dimension perturbation benchmark extending LIBERO. CoDA's specific 25%→84%/1%→76% figures on it remain unconfirmed; the paper's own numeric tables/figures defeated this session's text extraction. (claim-libero-pro-paper-confirms-benchmark-but-not-coda-percentages)

References

• claim-coda-co-evolution-raises-ood-manipulation-success-beyond-baselines-and-on-libero-pro
• claim-libero-pro-paper-confirms-benchmark-but-not-coda-percentages
• Captures: 10-inbox/raw/2026-10-18-outside-arxiv261022362-coda-co-evolution-of-data-engines.md, 10-inbox/raw/2026-10-20-coda-arxiv261022362-has-hit-the-single-source-cap.md