This research introduces a Graphormer-guided framework that enables robots to perceive and respond to safety risks during task execution, moving beyond static safety rules.

• Combines LLMs with graph neural networks for structured risk perception
• Enables low-latency hazard adaptation for safety-critical robotic applications
• Provides a novel approach for dynamic safety compliance during task execution
• Addresses significant gaps in robotic safety for industrial environments

This advancement is crucial for security applications where robots must operate safely alongside humans while adapting to changing risk conditions in real-time.

Graphormer-Guided Task Planning: Beyond Static Rules with LLM Safety Perception