FedAGV: Federated Meta-Learning for Rapid Adaptation of AGV Trajectory Planing

Automated Guided Vehicles (AGVs) are foundational to modern logistics and smart manufacturing, where the shift from mass production to flexible, short-series production demands frequent and rapid adaptation of AGV routes to individualized logistics tasks. In such dynamic industrial settings, precise trajectory prediction is essential for operational efficiency and safety. However, state-of-the-art deep learning models like Transformers are highly data-hungry and suffer from a severe "cold-start"problem, making it prohibitively expensive and disruptive to collect sufficient data and test new AGV route variants in real environments. To overcome this bottleneck, we propose FedAGV, a novel federated meta-learning framework that efficiently instills generalized motion knowledge from multiple existing AGV deployments in a privacy-preserving manner. This approach enables the creation of a foundational model that can be quickly and effectively adapted to new routes using minimal data, thereby significantly reducing the need for costly real-world testing. Experimental results demonstrate that FedAGV outperforms models trained from scratch in both data efficiency and prediction accuracy, paving the way for more agile and scalable AGV deployment in flexible manufacturing environments.