Ontology-Grounded Llm Planning for Robust Table-Clearing Robot Manipulation

Robots operating in shared environments must do more than execute actions. They must ground task instructions in the current world state and remain responsive to changes during execution. Large language models (LLMs) offer strong reasoning capabilities for task planning, but their reliance on static, text-only scene descriptions often leads to brittle plans that fail once execution begins. To overcome this, we detail an ontology-grounded planning framework for robot manipulation that integrates an LLM with a symbolic world model and execution-time state updates. Rather than reasoning over free text alone, the planner operates on a grounded state representation built from ontology queries and simulator observations. We evaluate the proposed approach in a simulated PyBullet table-clearing setting under nominal conditions and execution-time disturbances. Within this setting, the ontologygrounded configuration achieves higher plan feasibility, task success, and failure recovery than the implemented text-only baseline.