Risk Evaluation · Benchmark

ForesightSafety-VLA: A Unified Diagnostic Safety Benchmark for Vision-Language-Action Models

NUWA research note · independent review pending

First published: Publication date not verifiedSource checked: 2026-10-28

Original source ↗Cite this record ↓Autonomous physical harm →

NUWA Research Note

Research question Design and assumptions Source findings NUWA interpretation Limits and contrary evidence Research and decision relevance Evidence update Institutions and roles Original resources

Model-assisted reading · Selected sections of the original full text or official article; a complete section-by-section audit remains pending.

At a glance

A four-quadrant success/safety decomposition avoids hiding unsafe successes.

Research question

How do structural, language and visual perturbations affect embodied safety?

Design and assumptions

The protocol covers 13 safety categories, 66 base scenarios and five embodiments. Language and visual variations overlay the base scenarios. Outcomes separate safe success, safe failure, unsafe success and unsafe failure, with cumulative safety cost and exposure measures.

Source findings

The four-quadrant decomposition exposes unsafe successes that a task-success scalar can hide. The scenario and category counts describe benchmark design, not observed harm frequency.

NUWA interpretation

Process and endpoint evidence need joint interpretation. Composite metrics should retain inspectable components so that high completion does not conceal unsafe action.

Limits and contrary evidence

Overlay variants are not additional independent base tasks. Results depend on simulation safety thresholds and the evaluated embodiments.

Research and decision relevance

NUWA recommendation: Compare safety components within a shared embodiment and configuration, preserving base-scenario and perturbation identities.

Precise source locations

ForesightSafety-VLA: A Unified Diagnostic Safety Benchmark for Vision-Language-Action Models · §§III–V; discussion; text lines 00082,00084,00134,00141,00146,00147Original resource captured 2026-10-28T07:24:44.372873+00:00; sha256:303708eaa7bcf2e1d96cd551279703949de0dff652b1b5b86e0cd2dd5905329a

Source checking, extraction, editorial approval and replication are recorded separately. Methodology →

Evidence update · 2026-10-28

Model-assisted reading · independent evidence and bilingual editorial review pending. Values are source findings or inventory counts, separate from the trend result series.

Source-bound observations (2)

Counts, rates and contrasts have different units. A sample count is not a number of independent experiments. Unknown numerators and interval levels remain unknown.

System / metric	Reported value	Population and conditions
Base scenarios	66 scenarios	Conditions, limitations and source Inventory count, not outcome denominator or severity. Original source ↗ · Abstract; benchmark design; text lines 00082,00084,00134,00141,00146,00147Original capture 2026-10-28T07:24:44.372873+00:00; sha256:303708eaa7bcf2e1d96cd551279703949de0dff652b1b5b86e0cd2dd5905329a
Safety categories	13 categories	Conditions, limitations and source Inventory count, not outcome denominator or severity. Original source ↗ · Abstract; benchmark design; text lines 00082,00098,00100,00146Original capture 2026-10-28T07:24:44.372873+00:00; sha256:303708eaa7bcf2e1d96cd551279703949de0dff652b1b5b86e0cd2dd5905329a

Institutions and roles

Affiliation not indexed in this review

Authors (8)

Mingyang Lyu · Yinqian Sun · Yiyang Jia · Sicheng Shen · Moquan Sha · Huangrui Li · Feifei Zhao · Yi Zeng

Evaluation protocols

ForesightSafety-VLA →

Cite and reuse

Cite the original publication for its findings. Cite this NUWA record separately for its explanation and retained snapshot.

BibTeX ↓RIS ↓Record + source metadata · JSON ↓Reading coverage · JSON Suggest a correction

Original publication · BibTeX

@misc{foresight-vla,
  title = {ForesightSafety-VLA: A Unified Diagnostic Safety Benchmark for Vision-Language-Action Models},
  url = {https://arxiv.org/html/2606.27079},
  author = {Mingyang Lyu and Yinqian Sun and Yiyang Jia and Sicheng Shen and Moquan Sha and Huangrui Li and Feifei Zhao and Yi Zeng},
  version = {arXiv:2606.27079v2; version header 27 Jun 2026; Original resource captured 2026-10-28T07:24:44.372873+00:00; sha256:303708eaa7bcf2e1d96cd551279703949de0dff652b1b5b86e0cd2dd5905329a},
  note = {arXiv:2606.27079v2; version header 27 Jun 2026; Original resource captured 2026-10-28T07:24:44.372873+00:00; sha256:303708eaa7bcf2e1d96cd551279703949de0dff652b1b5b86e0cd2dd5905329a}
}

NUWA record · BibTeX

@misc{foresight-vla-nuwa,
  title = {ForesightSafety-VLA: A Unified Diagnostic Safety Benchmark for Vision-Language-Action Models — NUWA research record},
  url = {https://whitzard.tech/en/nuwa/whitzard-index/work/foresight-vla/},
  version = {2026-10-29.1},
  organization = {NUWA Frontier AI Safety Lab},
  howpublished = {Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--foresight-vla.json},
  note = {NUWA research note. Source-reading scope and independent review status are stated on the record. Snapshot 2026-10-29.1. Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--foresight-vla.json.}
}

Evidence snapshot 2026-10-29.1 · Index updates · RSS

References

[1] Mingyang Lyu, Yinqian Sun, et al.. ForesightSafety-VLA: A Unified Diagnostic Safety Benchmark for Vision-Language-Action Models ; §§III–V; discussion; text lines 00082,00084,00134,00141,00146,00147

Search NUWA Index