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

2026-07-15

AI Summary

This work addresses a critical gap in the evaluation of embodied agents by highlighting that existing methods overlook process-level safety risks arising from spatial relationships—such as support, containment, and proximity. To remedy this, the study introduces the first benchmark that treats spatial relations as a core dimension for process-level safety assessment. The benchmark comprises 507 executable scenarios designed to verify whether vision-language model–driven agents adhere to safety constraints prior to executing potentially hazardous actions. It integrates both spatial relationship modeling and a mechanism for process-level safety validation. Experimental results demonstrate that while mainstream models often succeed in task completion, they frequently violate fundamental safety constraints, revealing significant deficiencies in their capacity for safety-aware reasoning. This benchmark thus fills a crucial void in the evaluation of process-level safety for embodied intelligence.

Abstract

Vision-language models (VLMs) are increasingly used as the reasoning backbone of embodied agents, enabling robots to interpret visual scenes, follow language instructions, and plan multi-step actions. In household environments, however, safety depends not only on recognizing objects, but also on how actions change the physical scene over time. Existing embodied safety evaluations largely focus on static risk recognition, unsafe instruction refusal, or final-state task completion. As a result, process-level safety failures induced by spatial relations such as support, containment, and proximity remain insufficiently studied. To address this gap, we introduce SAFERELBENCH, a spatial-relation-aware safety benchmark with 507 executable evaluation samples, including 248 spatial-relation samples and 259 non-spatial control samples. Using SAFERELBENCH to evaluate seven open- and closed-source VLM-driven embodied agents, we find a substantial gap between task success and process-level safety compliance: models often complete the requested task while violating process-level safety constraints. Unlike prior benchmarks, SAFERELBENCH explicitly tests whether agents satisfy safety conditions before risk-prone actions, making spatial relations a core dimension in embodied safety assessment. More broadly, our results show that safe embodied intelligence requires not only stronger perception and planning, but also reliable reasoning about how object relations shape risk during interaction.

Problem

• process-level safety
• spatial relations
• embodied agents
• safety benchmark
• vision-language models

Innovation

• spatial relations
• process-level safety
• embodied agents
• vision-language models
• safety benchmark