Reproduce “CAPSafe: A Benchmark for Contextual Action and Plan Safety Reasoning in Embodied Agents” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Youmna Farag, Svetlana Stoyanchev, Simon Keizer, Mohan Li, Rama Doddipatla

CAPSafe tests safety in structured household scenes at both single-action and multi-step planning levels. It defines hazards through actions, object categories, and properties, then uses a symbolic simulator to check hazards during plans and at their end. Qwen3-32B reaches 91.8% unsafe-action recall but 51.65% plan-level safety, showing a gap under conditions that omit perception and low-level control.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.