AI-Driven Optimization of Urban Last-Mile Delivery A Multi-Agent Reinforcement Learning and Metaheuristic Approach.

Urban last-mile delivery presents significant challenges due to dynamic traffic conditions, stochastic customer demand, and operational constraints, often resulting in delayed deliveries, high operational costs, and increased environmental impact. This study investigates the application of artificial intelligence (AI) methods, including Reinforcement Learning (RL), Genetic Algorithm (GA), and Particle Swarm Optimization (PSO), to optimize last-mile delivery operations in a simulated urban environment. A multi-agent RL framework was developed, allowing delivery vehicles to dynamically adapt routes based on real-time traffic and customer demand. The performance of AI methods was evaluated against a conventional nearest-neighbor heuristic across metrics including delivery time, total distance traveled, operational cost, carbon emissions, and route completion rate. Results indicate that RL outperforms GA, PSO, and heuristics, achieving a 22% reduction in delivery time, 14% shorter travel distance, 18% lower operational cost, and 16.7% reduction in CO₂ emissions, while maintaining the highest route completion rate. GA and PSO also improved performance over heuristics but were less adaptive under dynamic conditions. The study demonstrates the practical feasibility and sustainability benefits of AI-driven last-mile delivery optimization. The proposed framework provides a foundation for scalable, real-time routing solutions, with implications for integrating AI with autonomous vehicles, electric fleets, and sustainable urban logistics.