Integrating LLMs and Classical Planning for Pallet Logistics: A Case Study

Robotic task planning requires generating structured and executable action sequences to achieve high-level goals. Large Language Models (LLMs) have demonstrated impressive abilities in natural language understanding and generation tasks, but their effectiveness in solving planning problems within the Planning Domain Definition Language (PDDL) remains an open question. In this work, we evaluate the feasibility of LLMs as planners by testing their performance on a novel pallet logistics domain, which introduces complex spatial reasoning and long-horizon planning challenges. Our results show that LLMs, even state-of-the-art models like GPT-4o and GPT-o1, struggle to generate reliable and executable plans, and their high computational cost makes them impractical for real-time robotic task planning. To address these limitations, we propose an agent-based framework that leverages the strengths of both LLMs and classical PDDL solvers. Instead of using LLMs for full-plan generation, we employ them to partially construct PDDL problem files by translating natural language task descriptions into structured goal definitions. The classical planner then ensures the optimality and feasibility of the final plan. This study provides an empirical evaluation of such hybrid planning in a realistic logistics setting, highlighting its robustness and potential for structured, language-guided robotic task execution.