Highly dynamic topology, rapid link breakage, and fluctuating density limit the reliability of conventional reactive routing in vehicular ad hoc networks (VANETs). This paper presents a digital twin-driven multi-agent reinforcement learning framework for predictive and multi-objective next-hop selection. The digital twin estimates communication-link lifetime from vehicle position and relative mobility, while each vehicle operates as a Q-learning agent. A weighted reward jointly considers link stability, residual packet lifetime, distance, and congestion. The method was implemented as a custom IPv4 routing module in NS-3 and compared with AODV under IEEE 802.11 communication, 20-40 vehicular nodes, a 300 m radio range, and a 60 s simulation. The proposed framework reduced average end-to-end delay from 0.30 s to 0.10 s and increased packet delivery ratio from 91.4% to 94.8%. The improvement was accompanied by a lower measured throughput of 201.8 kb/s compared with 285.9 kb/s for AODV, indicating a reliability-latency trade-off. The findings support predictive decentralized routing for intelligent transportation networks while identifying throughput and overhead optimization as important extensions.

Vehicular ad hoc networks, digital twin, multi-agent reinforcement learning, multi-objective routing, AODV, NS-3.