An LLM-Based Assistance System for Intuitive and Flexible Capability-Based Planning

Researchers developed a hybrid system combining formal symbolic planning with large language models to improve capability-based planning in industrial automation. The system integrates natural-language interaction, explainability, and human-approved knowledge model adaptation, achieving high accuracy across planning and query tasks while maintaining formal correctness guarantees.

This research addresses a critical gap in industrial automation: making complex planning systems accessible and adaptable to non-expert users. Traditional capability-based planning solvers generate optimal process sequences from machine-readable knowledge models, but their opaque feedback—particularly when plans fail—limits practical adoption. By layering an LLM assistant atop a formal SMT planner, the system bridges symbolic reasoning and human communication, allowing operators to query system capabilities, understand failures, and iteratively refine constraints in natural language.

The architecture reflects broader trends in AI: hybrid systems that preserve formal guarantees while leveraging LLMs' language fluency and reasoning flexibility. The Human-in-the-Loop approval mechanism is particularly significant, maintaining human oversight as knowledge models evolve—a crucial safeguard in safety-critical industrial contexts. The four-component design (grounding, planning, interpretation, adaptation) demonstrates modular thinking that enables replication across domains.

For industry, this work reduces barriers to advanced planning adoption. Small and mid-sized manufacturers often lack domain experts to maintain complex planning models; natural-language interfaces democratize access. The 75% success rate on test cases (9 of 10 queries, all satisfiable cases, most unsatisfiable cases with repair proposals) suggests practical viability, though larger-scale validation is needed. The iterative adaptation mechanism enables systems to evolve with operational changes without developer intervention.

Looking forward, success depends on real-world deployment data—how systems perform with diverse industrial contexts, unexpected constraint violations, and noisy user inputs. Integration with Industry 4.0 platforms and digital twins could accelerate adoption, while the approach's generalizability to other formal-planning domains remains to be demonstrated.

• →Hybrid LLM-SMT planner improves industrial automation accessibility by enabling natural-language interaction with formal planning systems.
• →Human-in-the-Loop approval mechanism ensures formal correctness while allowing flexible, iterative knowledge model adaptation.
• →System achieved 75%+ success on planning queries and satisfiable cases, with concrete repair proposals for unsatisfiable scenarios.
• →Architecture decomposes into five specialized agents (grounding, planning, interpretation, adaptation, routing) enabling modular replication.
• →Reduces expertise barriers for manufacturers to adopt advanced planning, with potential for broader AI-symbolic-reasoning hybrid applications.