FL-P4AV: A Federated Learning-Based Privacy-Preserving Personalized Path Planning Framework for Collaborative Autonomous Ground Vehicles

The growing deployment of autonomous ground vehicles in smart cities and logistics demands secure, efficient, and context-aware navigation systems. This paper proposes FL-P4AV, a Federated Learning-Enabled Privacy-Preserving Personalized Path Planning framework designed for collaborative autonomous vehicles. Unlike traditional centralized or static path planners, FL-P4AV allows each vehicle to train a lightweight local model that predicts navigation costs based on semantic features such as obstacle density and goal proximity. The models undergo refinement via federated learning, which maintains privacy by not sharing raw data and employing differential privacy mechanisms. The semantic weights obtained are incorporated into an enhanced A* algorithm to facilitate personalized and efficient route computation. Experimental evaluations in dynamic grid environments indicate that FL-P4AV results in shorter paths, fewer inflection points, reduced turning angles, and quicker planning times relative to baseline methods. Despite the existence of privacy-preserving noise, the system maintains steady convergence and adjusts dynamically to real-time environmental changes. FL-P4AV offers a scalable and secure framework for the coordination of decentralized autonomous vehicles in path planning, indicating substantial potential for real-world applications in smart transportation systems.