# ARMOR preprint reports gains in detecting and explaining robot failures

_Published Friday, October 16, 2026 at 9:42 AM EDT · Science, Robotics · Latest · Tier 2 — Notable_

Researchers report that ARMOR, a vision-language model for detecting and explaining robotic failures, improved failure detection rates by up to 30% over previous approaches in experiments across diverse environments. The preprint also reports improvements of up to 100% in reasoning measured by an LLM fuzzy match score.

ARMOR repeatedly revises its detection decisions and explanations using its earlier outputs. It trains on large collections of binary labels alongside smaller collections of detailed explanations, then selects among multiple prediction trajectories using a confidence measure. The reasoning improvement measures agreement under an LLM-based score, rather than a demonstrated reduction in real-world robot accidents.

## Sources

- [cs.LG updates on arXiv.org](https://arxiv.org/abs/2602.12405)

---
Canonical: https://techandbusiness.org/newswire/VsQGwUv1428GIzFFxHLWvt
Published: 2026-10-16T13:42:34.405Z
Story chronology: 2026-10-16T04:00:00.000Z
Retrieved: 2026-10-16T17:03:30.634Z
Publisher: Tech & Business (techandbusiness.org)