A Safety-Shielded Deep Reinforcement Learning Approach for Virtual Coupling Train Following under Uncertainty and Communication Delay Conditions

Virtual Coupling (VC) aims to reduce inter-train following distances by relying on relative braking and vehicle-to-vehicle information exchange. Although VC train-following control is commonly addressed using model-based approaches such as Model Predictive Control, maintaining safe operation under communication delay, adhesion degradation, and positioning uncertainty remains challenging. This study proposes a safety-shielded Deep Reinforcement Learning (DRL) architecture for VC train following. A Soft Actor-Critic policy generates continuous nominal control commands for speed tracking, headway regulation, comfort, and actuation effort, while a braking-distance-based safety shield supervises the hard minimum-distance constraint through a one-step action-projection mechanism whenever the projection remains feasible. Therefore, the DRL policy is not treated as a standalone safety-certifying controller; safety is provided by the combined policy–shield architecture. The proposed method is evaluated under nominal operation, 10% adhesion loss, a communication-delay sweep (τ ∈ {0.0, 0.5, 1.0, 2.0, 5.0, 10.0} s), and a combined-uncertainty stress test involving delay, adhesion loss, and position noise. All scenarios are assessed over 50 randomized seeds using worst-case and distributional statistics. Additional diagnostics, including safety-shield activation, raw-action feasibility, correction magnitude, and shield-off evaluation, are used to distinguish the role of the learned policy from that of the safety shield. The results show safe operation under nominal conditions, 10% adhesion loss, and communication delays up to 2.0 s in the considered setting. In contrast, delays of 5.0 s and 10.0 s and the combined-uncertainty stress test lead to safety violations and early termination, identifying the feasibility boundary of the considered VC formulation. Overall, the study provides a reproducible safety-shielded DRL baseline for VC train following and clarifies the delay regimes under which the proposed architecture remains applicable.