MAPPO boosts C-V2X delivery by 4% and halves training time

The paper, submitted to arXiv by Moritz Schaffenroth and colleagues, tackles the real-time decision problem of which radio access technology (RAT) a connected vehicle should use: the cellular Uu link, the NR-V2X PC5 sidelink, or both simultaneously. As V2X apps move beyond simple Cooperative Awareness Messages to bandwidth-hungry, latency-sensitive tasks like shared perception and cooperative driving, a single static channel choice no longer suffices. The authors formulate this as a multi-agent reinforcement learning problem and apply MAPPO (Multi-Agent Proximal Policy Optimization), comparing it against a single-agent deep RL baseline, a static decision tree, and fixed channel strategies.

In an urban test scenario, MAPPO achieved a 0.535 on-time delivery ratio for a single controlled vehicle (vs. 0.508 for DRL) and 0.567 when all vehicles followed the learned policy (vs. 0.548). Training time dropped by half because the multi-agent setup naturally models vehicle-to-vehicle interactions. The improvement was most pronounced in advanced application scenarios, suggesting that future autonomous driving systems will rely on adaptive, multi-agent communication strategies to meet heterogeneous reliability and latency requirements.