A network-resilient and privacy preserving federated imitation learning framework for truck platooning

Abstract

The enhancement of Intelligent Transportation Systems (ITS) aims to revolutionize freight logistics through heavy-duty vehicle platooning, promising significant reductions in fuel consumption and enhancements in road safety. However, the operational reality of international logistics presents a connectivity paradox. Cooperative Adaptive Cruise Control (CACC) systems generally rely on centralized cloud architectures. This creates a critical limitation for international logistics operations. Cross-border transit introduces predictable network dead zones. Furthermore, stringent data sovereignty regulations, such as GDPR, make continuous raw data transmission technically and legally unfeasible. This thesis addresses these challenges by proposing a Network-Resilient Federated Imitation Learning (NR-FIL) framework designed specifically for the intermittent connectivity of international transit routes. The proposed framework synthesizes Federated Learning (FL) with Imitation Learning (IL). Instead of transmitting sensitive raw telemetry to a central server, the system treats each truck as an edge node. A local neural network on each vehicle approximates the safety-verified onboard CACC controller. This transformation of analytical control logic into a learnable neural policy is a key enabler, as it enables the deployment of Federated Learning—allowing the fleet to collaboratively improve and generalize the control strategy across diverse vehicle dynamics without sharing raw telemetry. To handle network partitioning, a novel "Store-and-Forward" mechanism is introduced. This architecture buffers local gradient updates during periods of disconnection and transmits them asynchronously once a secure connection is restored, ensuring learning continuity without data loss. The framework was validated using a simulation environment built with SUMO (Simulation of Urban MObility) and the TraCI interface, modeling the real-world geometry of the Turkey-Bulgaria (Kapıkule) border corridor. The experimental setup involved a platoon of 20 heavy-duty trucks subjected to stochastic packet loss and deterministic network blackouts. The results demonstrate that the NR-FIL framework significantly outperforms centralized baselines in terms of communication efficiency. While the system effectively maintains global model convergence, the restoration of connectivity following a dead zone triggers a "Soft Reconnection Shock," where the aggregation of divergent local models leads to increased variance rather than a sharp loss spike. However, the system recovers within approximately 200 simulation steps. Safety analysis reveals that while the system is robust, a model staleness threshold of 50 rounds exists, beyond which safety gaps may be compromised. Ultimately, this work confirms that privacy-preserving, decentralized learning with buffering mechanisms is a viable solution for ensuring operational continuity in the unpredictable infrastructure of international logistics.