RB-LLM Control: An intelligent control framework with rule-based large language model decision-making

• This paper proposed an advanced intelligent control framework for autonomous systems. • The framework integrates a classical feedback controller, specifically an LQR, with a reasoning-based decision-making layer powered by the RB-LLM for robust autonomous operation. • A structured prompting strategy is developed to improve the logical reasoning of LLMs in autonomous control systems. • The RB-LLM controller enhances decision-making under environmental disturbances and model uncertainties. • RB-LLM interprets natural language and log/error data to adapt control inputs using structured prior knowledge. In this study, the long-term logical reasoning of Large Language Models (LLMs) is explored to enhance the decision-making of autonomous control systems operating in the presence of environmental disturbances and model uncertainties. Given that LLMs may exhibit limitations in mathematical and logical reasoning, a structured prompting strategy based on logical rules is designed in this paper to enhance reliability and improve response quality. In this paper, Rule-Based LLM (RB-LLM) is incorporated into the architecture of a control system to strengthen resilience in challenging scenarios that the system wasn’t specifically designed for, it interprets natural language and other log/error information for guidance and adaptation of control inputs using structured user-provided prior knowledge. Simulation results validate the effectiveness of our approach, emphasizing the potential of RB-LLM decision-making to advance the autonomy of control systems in complex and dynamic environments. The effectiveness of the proposed RB-LLM controller in terms of tracking accuracy, robustness to uncertainty, safety in harsh disturbances, and computational efficiency is evaluated. In this study, a comparative performance analysis of the proposed RB-LLM control framework, classical, and state-of-the-art approaches is conducted. Simulation results validate the advantages and limitations of each method, offering insights into the trade-offs between classical, optimization-based, and RB-LLM control techniques. For example, the RB-LLM framework achieves a reduction in maximum error ranging from 36% to 89% compared to existing approaches. RB-LLM control framework shows superior adaptability in the presence of modeling uncertainties, nonlinear dynamics, and emergency scenarios. The proposed framework is readily extendable to a broad range of control applications; in this study, its implementation is demonstrated using simulations for Unmanned Aerial Vehicles (UAVs).