Delay-aware multi-agent reinforcement learning for mixed-autonomy platoon control

It is recognized that controlling mixed-autonomy platoons comprising connected and automated vehicles (CAVs) and human-driven vehicles (HDVs) can enhance traffic flow. While multi-agent reinforcement learning (MARL) is a promising real-time control paradigm, existing MARL-based platoon controllers rarely account for communication and execution delays, despite their inevitability in practice and their critical impact on safety and stability. Moreover, most delay-compensation approaches are model-based, which becomes unsuitable when HDV dynamics are unknown and model-free coordination among CAVs is required. To address this gap, we formulate mixed-autonomy platoon control with delays as a delayed Markov game and develop a delay-aware learning framework supported by a delay-dependent theoretical analysis. Specifically, we provide a theoretical analysis that establishes explicit performance bounds between delayed and undelayed tasks under smoothness conditions, highlighting that the delay-induced performance gap is governed by the policy smoothness and the belief uncertainty under delayed observations. Motivated by this insight, we propose a multi-agent transformer (MAT) that exploits the disturbance-propagation structure of platoons to learn coordinated and regularized policies, serving as effective undelayed experts for reliable transfer to delayed environments. Finally, we validate the proposed approach through extensive simulations and human-in-the-loop experiments, demonstrating the control performance and sample efficiency of the proposed method.