SafeRelBench: A Spatial-Relation-Aware Benchmark for Process-Level Safety in VLM-Driven Embodied Agents

arXiv cs.AI · 6 hours ago

Researchers introduced SafeRelBench, a benchmark with 507 samples to evaluate whether vision-language model-driven robots maintain safety during physical interactions by understanding spatial relationships like support and containment. Testing seven VLM-based embodied agents revealed that models frequently complete tasks while violating process-level safety constraints, with the benchmark explicitly measuring whether agents satisfy safety conditions before executing risk-prone actions. The findings indicate that safe embodied AI requires improved reasoning about how spatial relationships between objects affect risk during physical interactions.