A Neuro-Symbolic Framework for LLM-Driven Task Planning and Execution in Industrial Assembly

Abstract

This paper proposes a neuro-symbolic framework that combines large language model (LLM)-driven task planning with formal symbolic verification and reactive Behavior Tree execution to support high-precision operations in industrial assembly. While Large Language Models (LLMs) demonstrate expertise in interpreting natural language instructions, they often struggle with processing structured manufacturing data and lack long-horizon planning capabilities and deterministic physical awareness required for complex robotic execution. Furthermore, their susceptibility to generating physically unfeasible sequences and their lack of robust error-recovery mechanisms limit their autonomous deployment. To address these limitations, the proposed pipeline decomposes the assembly process into sequential stages of planning, technological grounding, adaptive execution, and recovery. Initially, the framework leverages LLMs to synthesize multi-modal inputs—ranging from natural language directives to formalized product and process specifications—into a formal Hierarchical Task Network (HTN). To ensure physical feasibility, a deterministic rule-based compiler validates the HTN’s syntactic and structural consistency, rejecting undefined symbols and non-decomposable tasks, against established spatial and operational constraints. This verified logical sequence is subsequently compiled into a reactive Behavior Tree (BT) that guides execution through a fixed robotic skill library. During execution, a real-time object detection system and an atomic skill library dynamically translate the BT into grounded actions. Designed for human–robot collaborative environments, the architecture autonomously allocates standard operations to the robotic agent while safely deferring tasks requiring high dexterity to human operators. By combining the semantic adaptability of LLMs with the rigorous state validation of deterministic rule-based tools, this methodology is intended to reduce hallucination-driven failures and support safer manufacturing workflows, as evaluated through a system-level feasibility study rather than a component-wise ablation. The proposed architecture is evaluated against both physical and logical criteria using task-specific evaluation metrics. The framework was evaluated on four assembly tasks (T1–T4) using two Gemma model configurations, with 30 trials per task for each configuration (120 trials per configuration; 240 trials in total), achieving near-100% plan validity and near-zero hallucination rates under the higher-capacity configuration. Future integration of Vision-Language-Action (VLA) policies at the level of individual Behavior Tree leaves is discussed in Section Perception and Dynamic Adaptation as a promising direction to replace the current fixed robotic skill set.