Researchers introduce SPG-LLM, a novel approach that leverages large language models to optimize the grounding process in classical planning by identifying irrelevant objects and actions before computation. The method achieves significantly faster grounding times—often by orders of magnitude—across seven challenging benchmarks while maintaining or improving plan quality.
SPG-LLM addresses a fundamental computational challenge in automated planning where the grounding phase—converting abstract problem descriptions into concrete action instances—becomes exponentially expensive as task complexity increases. Traditional approaches struggle because the number of possible groundings grows rapidly with domain size, creating a severe bottleneck that limits scalability. The innovation lies in leveraging LLMs' natural language understanding capabilities to analyze PDDL (Planning Domain Definition Language) files semantically rather than relying solely on learned embeddings or relational features. This textual analysis enables the system to heuristically filter out irrelevant components before the computationally intensive grounding occurs.
The research builds on recent advances in partial grounding, which incrementally grounds only the most promising operators. However, most prior work treated domain knowledge implicitly through neural models rather than explicitly interpreting the semantic structure embedded in PDDL descriptions. SPG-LLM reverses this approach by treating LLMs as semantic analyzers that extract domain-specific insights humans would naturally recognize.
The practical implications are substantial for the planning community. Achieving orders-of-magnitude speedups in grounding directly enables solving larger, more complex problems that were previously intractable. This matters for applications spanning robotics, scheduling, and logistics optimization. The results suggest that hybrid approaches combining LLM reasoning with traditional planning algorithms may outperform purely learned or purely symbolic methods. Future research will likely explore extending this semantic analysis to other planning bottlenecks and investigating whether LLM-guided pruning strategies generalize across diverse domain types.
- →SPG-LLM uses LLMs to semantically analyze PDDL files and eliminate irrelevant objects and actions before grounding, reducing computational overhead.
- →The approach achieves grounding speedups of multiple orders of magnitude on seven hard-to-ground benchmarks while maintaining comparable or better plan quality.
- →This hybrid methodology combines explicit semantic analysis through LLMs with traditional symbolic planning rather than relying solely on learned representations.
- →The innovation has direct applications for scaling automated planning to larger problems in robotics, scheduling, and logistics domains.
- →Results demonstrate that leveraging textual and structural information in domain descriptions can be more effective than pure neural or relational approaches.