Researchers present an end-to-end framework that uses Large Language Models to convert natural language specifications into PDDL planning models, with iterative refinement through hardcoded and dynamic agents, then generates executable plans. The system demonstrates strong performance across multiple domains including classic planning problems where LLMs typically struggle, and integrates with established planning engines.
This research addresses a fundamental challenge in AI: bridging the gap between human intent expressed in natural language and formal computational planning. The framework tackles planning problems by decomposing the task into manageable stages—specification parsing, iterative refinement, validation, execution, and human-readable output translation. This represents meaningful progress in making formal planning systems more accessible to non-specialists.
The distinction between hardcoded agents (addressing syntax and constraint issues) and dynamic agents (adapting to domain-specific requirements) reflects practical engineering. By allowing LLMs to operate within structured workflows rather than purely generative contexts, the approach mitigates known limitations where language models struggle with logical reasoning and complex planning tasks. Testing across benchmarks like Google NaturalPlan and classic problems (Sokoban, Blocksworld, Tower of Hanoi) demonstrates the framework's versatility beyond narrow use cases.
The system's compatibility with multiple planning engines (Fast Downward, LPG, POPF, VAL) enhances its practical applicability for organizations with existing planning infrastructure. This modularity signals the work's maturity—it solves a real coordination problem rather than creating proprietary lock-in.
For the AI industry, this demonstrates how LLMs function best as components within larger systems rather than standalone solvers. The framework's success suggests growing viability for AI-assisted formal methods in logistics, robotics, and resource management. Future iterations may focus on scaling to more complex domains and reducing computational overhead of iterative refinement cycles.
- →LLM-powered orchestrator translates natural language to PDDL with iterative refinement from specialized agents addressing syntax, constraints, and domain-specific abstractions.
- →Hardcoded agents fix predictable issues while dynamic agents adapt to domain requirements, combining structured reliability with flexible reasoning.
- →Framework successfully handles planning domains where LLMs typically fail, including Tower of Hanoi and Sokoban through formal verification integration.
- →Compatibility with multiple established planning engines (Fast Downward, POPF, VAL) enables practical deployment in existing systems.
- →End-to-end natural language input-to-output pipeline eliminates human intervention, improving accessibility of formal planning systems.