LLM-Enabled Capability-Based Planning System: An Intelligent Assistant for Industrial Automation

This article introduces a hybrid auxiliary system that combines large language models (LLMs) with symbolic planning, aiming to address the interpretability and adaptability issues of capability-based planning in industrial automation scenarios. The system adopts a layered architecture: an SMT solver serves as the underlying layer to ensure planning correctness, while the LLM acts as a natural language interaction layer to handle user intentions and result interpretation. A human-in-the-loop mechanism is also introduced to ensure the controllability of knowledge model adjustments. This system has demonstrated good reliability and adaptability in modular production system tests, providing an intelligent assistant solution for industrial automation planning.

Modern industrial production faces challenges of dynamic demands and complex resources. Traditional capability-based planning methods have two major problems: first, solver feedback (especially when there is no solution) is difficult for humans to understand; second, there is a lack of intuitive interaction methods for adjusting knowledge models. The root cause lies in the "semantic gap" between symbolic planning systems and humans—machines process logic using formal languages, while humans are accustomed to natural language communication. How to balance formal rigor and user-friendly interaction has become an urgent issue to solve.

The system adopts a layered architecture, with the LLM as a natural language interaction layer superimposed on an SMT-based capability planner. Core components include: 1. Capability grounding (converting natural language requests into formal capability descriptions); 2. Symbolic planning (SMT solver performs logical reasoning to generate plans); 3. Result interpretation (LLM converts formal results into natural language, including analysis of reasons for no solution); 4. Planning adaptation (LLM proposes knowledge model adjustment suggestions, which are modified and re-planned after user approval). In addition, a central router coordinates the division of labor among five professional agents (natural language understanding, formal conversion, etc.).

The research team designed 4 types of scenarios with a total of 23 test cases on modular production systems: 1. Knowledge query (9 out of 10 cases handled correctly); 2. Satisfiable planning (all 4 cases successful); 3. Unsatisfiable repair (3 out of 4 cases generated valid repair suggestions); 4. Adaptive planning (all 5 cases converted to satisfiable through iterative adjustments). The results prove that this hybrid method is technically feasible, combining the rigor of symbolic methods with the ease of use of natural language interaction.

The main contributions of this research are: 1. Proving that LLMs can serve as effective natural language interfaces, improving system accessibility without sacrificing the correctness of underlying planning; 2. The human-in-the-loop mechanism provides a safe and controllable path for the dynamic evolution of knowledge models, with users holding the final decision-making power; 3. The router-based agent architecture provides a reference paradigm for modular design of complex industrial systems, with clear component responsibilities facilitating expansion and maintenance.

Current system limitations: The success rate of repair suggestions for unsatisfiable scenarios needs to be improved, and the ability to analyze complex conflicts is insufficient; the real-time performance of explanation generation needs optimization when the scale of the knowledge model expands. Future directions: Introduce stronger reasoning mechanisms to handle complex constraint conflicts; explore active learning to reduce reliance on user approval; expand to multi-agent collaborative planning scenarios. This hybrid architecture is expected to help humans cope with complex production environments in smart factories.