Generative AI meets classical control: hybrid few-shot PI tuning for autonomous robots

Foundation models have remarkable few-shot learning and data-generation capabilities. We harness these to adaptively tune Automated Guided Vehicle (AGV) Proportional–Integral–Derivative (PID) controllers with minimal real-world data. Our few-shot transfer learning strategy tackles the tedious trial-and-error retuning required for new conditions. We train an ensemble regression model on initial AGV data, then use a pre-trained foundation model to generate synthetic control samples from a few new trials, augmenting the dataset. Fine-tuning the ensemble on this combined real and synthetic data enables rapid convergence to effective Proportional–Integral (PI) gains for changing scenarios while ensuring precise, stable navigation. Real-world Automated Guided Vehicle (AGV) tests confirm robust tracking under varying speeds and reduce manual retuning effort, with lateral tracking Root Mean Square Error (RMSE) at 20 m/min reduced from 1.17 to 0.71 m over a few adaptation cycles. Offline, Generative Pre-trained Transformers (GPT)-augmented training maintains high predictive accuracy across 5–25 m/min (R 2 > 0.90), whereas training only on real logs causes R 2 to drop to approximately 0.33 at higher speeds. This hybrid of generative AI and classical control is novel: unlike methods requiring extensive data or manual tweaks, ours uses Artificial Intelligence (AI)-synthesized data for adaptive performance with minimal trials and few-shot measurements.