LLM-based User-Customizable Economic MPC for Consumer Unmanned Surface Vehicle Path Planning

Existing unmanned surface vehicle (USV) path planning methods based on model predictive control (MPC) frequently necessitate extensive manual parameter adjustments in response to changes in the control task, which limits their usability in consumer-grade unmanned electronic systems operated by non-expert users. When addressing the conflicting requirements of safety and economic efficiency in path planning, MPC typically prioritizes trajectory safety. Additionally, methods for adjusting controller performance based on personalized user requirements typically depend heavily on extensive user feedback or pretraining data. In this paper, an economic model predictive control (EMPC) approach assisted by a large language model (LLM) is proposed. Using an LLM agent, users can conveniently and efficiently update controller parameters via natural language to meet their personalized requirements. The LLM agent translates user instructions into mathematical expressions in real time, thereby enabling parameter adjustments for the EMPC controller and achieving performance consistent with user demands. Meanwhile, the EMPC approach balances obstacle avoidance safety with the USV’s operational energy consumption, effectively managing the trade-off between these two objectives. Experimental results demonstrate the effectiveness and feasibility of the proposed method, indicating its potential for real-time deployment in consumer unmanned surface vehicle electronic systems that require flexible user interaction and energy-efficient autonomous navigation.