With the rapid development of Large Language Models (LLMs), Embodied Artificial Intelligence (EAI) has entered a new phase of advancement. However, in real-world human–robot interactions, the ambiguity of user instructions continues to pose significant challenges to the feasibility and safety of task execution. Existing approaches have notable limitations: rule-based static safety constraints struggle to adapt to diverse and dynamic scenarios; multimodal models often lack fine-grained risk identification; and reinforcement learning-based approaches rely heavily on trial-and-error, resulting in delayed responses to potential risks. To address these challenges, we propose RoboClarify, a scenario-guided instruction clarification approach. Based on objects in the environment, this approach evaluates both the inherent basic risk and the distance-based risk between objects. It then integrates a constructed Scenario Risk Graph with the semantic reasoning capabilities of LLMs to identify potential risk relations and guide the transformation of ambiguous instructions into semantically clear and executable ones. This process enhances task understanding and execution in complex environments. We conducted empirical evaluations using the public SafeAgentBench dataset and our newly constructed HomeAgent dataset. The results show that three mainstream LLMs (GPT-4o mini, DeepSeek, and Qwen) achieved clarification rates exceeding 94% on both datasets, significantly improving their ability to interpret ambiguous instructions. These findings demonstrate the strong potential of RoboClarify to enhance the safety and practicality of robotic task execution.