A Federated Multi-Task Meta-Learning Framework for Collaborative Perception and Adaptation in Connected and Automated Vehicles

Connected and Automated Vehicles (CAVs) operate in dynamic environments influenced by traffic patterns and pedestrian behaviour, which complicates the development of real-time navigation algorithms with voluminous data communicated by CAVs, raising privacy concerns. To address these challenges, we propose Federated Learning (FL) for concurrent and collaborative learning across fleets to generate privacy-preserving personalised models that adapt to diverse environments. Combining graph neural networks (GNNs) enables the real-time modelling of vehicle interactions and captures spatial and temporal dependencies. Utilising a message-passing paradigm, GNNs facilitate dynamic communication among vehicles. By aggregating information from neighbouring nodes, GNNs learn meaningful feature representations that enhance perception in CAVs, improving their responsiveness and enabling route optimisation and traffic flow enhancement. In this work, Model Predictive Control (MPC) influences GNNs to improve vehicle state prediction. It optimises control actions that minimise a cost function, such as travel time, fuel consumption, or collision risk, while adhering to constraints. GNNs enable the system to adapt its predictive model based on evolving vehicle relationships. At the same time, MPCs reoptimise control actions in response to these changes, allowing the CAVs to manage trajectories and make informed decisions adaptively in dynamic environments. The Federated Multi-Task Meta-Learning Framework for Collaborative Perception and Adaptation in Connected and Automated Vehicles (FedCAV) model is deployed across Edge, Fog, and Cloud layers to optimise performance, with a total estimated latency of 210 ms for 10 vehicles, influenced by local model training. Its low first-byte latency of 25 to 34 ms enhances communication efficiency, facilitating real-time decision-making and adaptive interactions.