LLM-Based Hierarchical Control Architecture for Robotic Operation

Abstract

Robotic operation in unstructured environments remains a long-standing critical challenge, driven by inherent environmental uncertainty, task complexity, and the conflicting requirements for high-level semantic reasoning and low-level real-time reactive control. While large language models (LLMs) demonstrate exceptional natural language understanding and common-sense reasoning capabilities, directly deploying them as end-to-end robot controllers suffers from severe limitations, including high inference latency, insufficient real-time responsiveness, and potential safety risks induced by model hallucinations. To address these issues, this paper proposes a novel LLM-based three-layer hierarchical control architecture, which synergizes the cognitive advantages of LLMs with the real-time reliability of traditional control systems, and designs a closed-loop feedback mechanism for dynamic replanning. Extensive experiments on multiple representative robotic tasks verify that the proposed approach achieves a task success rate of up to 94%, outperforming state-of-the-art baselines by 7-12% while strictly complying with real-time performance and safety constraints.