CORE: Context-Aware Safety Reasoning and Grounding for Open-World Robots

The researchers propose a robot safety framework, CORE, which uses a vision-language model (VLM) to reason online about context-dependent safety rules from visual observations, without requiring environment priors such as maps or safety specifications. CORE grounds these rules in the physical environment and enforces the resulting spatial safety set via control barrier functions, while providing probabilistic safety guarantees that account for perception uncertainty. Simulation and real-world experiments show that CORE can enforce context-compliant safety behavior in unseen environments, significantly outperforming existing semantic safety methods that lack online context reasoning. Ablation experiments validate the theoretical guarantees and show that both VLM reasoning and spatial grounding are indispensable. Project resources are available at https://zacravichandran.github.io/CORE。