Prompt-Based PID Tuning for Adaptive AGV Control via Few-Shot and MultiSpeed Strategies

Automated Guided Vehicles (AGVs) are central to material handling in modern warehouses, where precise motion control is critical for ensuring high throughput, safety, and energy efficiency. At the core of this control lies Proportional-Integral-Derivative (PID) tuning, which remains a persistent challenge in dynamic, data-scarce, and resource-constrained environments. This study introduces a prompt-driven PID tuning framework that leverages large language models (LLMs) to infer optimal gains through structured in-context learning, eliminating the need for system identification or iterative tuning. The proposed framework implements two novel strategies: (1) <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MultiSpeed (MS)</i> prompting, which interpolates gain values based solely on the target speed, and (2) <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Few-Shot + MultiSpeed (FS+MS)</i> prompting, which incorporates exemplar-based conditioning for improved generalization. Experiments conducted on a two-wheeled AGV across five discrete velocities (5–25 m/min) demonstrate that both methods achieve PID gains closely aligned with simulation-optimized baselines, while FS+MS consistently yields the most accurate and stable results. The inferred gains preserve physically consistent trends, such as increasing <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">K_p</i> and decreasing <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">K_i</i> with speed, and maintain trajectory tracking RMSE within 5–10% of optimal values. Our framework achieves over 80% reduction in calibration effort and supports millisecond-scale inference on embedded hardware, making it well-suited for real-time deployment in edge robotics.