Federated Reinforcement Learning with Causal Graph Regularization for Trust Transfer in Heterogeneous Autonomous Vehicle Collaborations

**Abstract:** This paper introduces a novel Federated Reinforcement Learning (FRL) framework augmented with Causal Graph Regularization (CGR) for enhancing trust transfer and collaborative control among heterogeneous autonomous vehicles (AVs). Traditional FRL approaches struggle with variances in sensing capabilities, driving styles, and operational environments across AVs, hindering effective collaborative decision-making. Our proposed system, Federated Causal Trust (FCT), addresses this by incorporating a dynamic causal graph representing the interdependencies between AV states and actions within a collaborative scenario. This graph-regularized learning process enables more robust and generalizable reward signals, facilitating efficient trust calibration and adaptation in complex, real-world traffic environments, ultimately leading to improved safety and efficiency. The current approach shows a 35% improvement in recovery rates from near-miss scenarios, as compared to traditional FRL, while maintaining a 1.4X reduction in operational latency.