Federated Learning for End-to-End Naturalistic Driving: Generalization, Stability, and Knowledge Transfer

Abstract

This paper presents a systematic evaluation of generalization and knowledge-sharing capabilities in end-to-end naturalistic driving under Federated Learning. Using the Autonomous Naturalistic obstacle-Avoidance Interaction Dataset dataset, comprising four heterogeneous drivers performing real-world steering and obstacle-avoidance maneuvers, PilotNet is assessed across three training paradigms: isolated local training, centralized training on pooled data, and horizontal FL using FedAvg and FedProx aggregation protocols.