Deployment Oriented Federated Learning in Connected and Autonomous Vehicles: Challenges and Opportunities

Connected and Autonomous Vehicles (CAVs) operate in dynamic and heterogeneous environments where sensing, computation, communication, and control must continuously adapt to changing road, weather, and traffic conditions. Traditional centralized learning faces increasing limitations in such settings, since transmitting large sensor datasets to remote servers increases bandwidth usage, raises privacy concerns, slows adaptation, and limits scalability across diverse vehicle hardware and sensor configurations. Federated learning (FL) provides a decentralized alternative that enables vehicles to collaboratively improve models without sharing raw data, supporting local training that naturally accommodates heterogeneity and facilitates continuous adaptation to rare, long-tail driving scenarios. However, deploying FL in CAVs remains challenging because training and inference must coexist during vehicle operation under strict real-time, safety, and resource constraints. This paper discusses the unique challenges of deploying FL in CAVs that arise from the tightly coupled architectural integration of CAV systems and federated learning frameworks. We outline key deployment challenges and design principles for CAV systems and FL to ensure reliable, adaptive, and safe learning in autonomous mobility.