This research addresses a critical weakness in LMM-based robotic planning: the limited ability to interpret complex scenes in domain-specific contexts.

• Proposes a structured state grounding approach using domain-conditioned scene graphs
• Improves scene understanding compared to current LMM-based methods
• Enhances robotic planning capabilities through better environment perception
• Particularly valuable for engineering applications requiring precise task execution

For engineering teams, this approach offers more reliable robotic systems capable of understanding their operational environments with greater domain-specific precision, potentially reducing errors in automated workflows.

Domain-Conditioned Scene Graphs for State-Grounded Task Planning