Abstract
Laboratory safety is procedure-dependent: required personal protective equipment (PPE) and permissible actions vary across experiments and across experimental steps, yet most vision-based monitoring remains appearance-driven and often produces generic warnings without reliable procedural context. We propose a multimodal framework for step-aware safety monitoring in laboratory videos. The framework first localizes procedural context through clip-level step prediction and protocol alignment to identify the experiment and current step. Given this context, it retrieves step-specific safety constraints, extracts evidence of step-relevant equipment and interactions using an equipment database, and prompts a video-capable vision-language model (VLM) to generate structured (JSON) monitoring reports supported by retrieved constraints and visual evidence. Experiments on protocol-annotated molecular biology lab videos show that our approach improves the mean score from 0.4352 to 0.6430 and reduces the missing rate from 65.00% to 33.75% relative to a video-only baseline, demonstrating more faithful and step-specific safety judgments.