Intelligent Path Optimization for Carbon-Constrained Last-Mile Delivery: A Reinforcement Learning and Heuristic Approach

The rapid expansion of e-commerce has intensified environmental challenges in urban logistics, particularly in last-mile delivery operations, leading to increased carbon emissions. This paper proposes a novel hybrid optimization framework that integrates deep reinforcement learning with ant colony optimization to address the carbon-constrained vehicle routing problem with time windows. The proposed approach employs a Deep Q-Network for intelligent action selection combined with adaptive ant colony refinement to achieve multi-objective optimization, balancing operational costs, delivery efficiency, and environmental sustainability. Experimental validation using real-world e-commerce datasets from major US urban areas demonstrates significant improvements across three metropolitan datasets (Chicago, Phoenix, Seattle), the proposed hybrid approach reduces CO₂ emissions by 21.6% and delivery costs by 17.3% on average relative to common heuristics (Clarke–Wright, GA, ACO), while maintaining delivery time compliance above 95.7%. Compared to the Clarke–Wright baseline alone, the approach yields substantially larger improvements, consistent with the trends shown in Table 5. The framework provides scalable decision support for sustainable logistics operations and contributes theoretical insights into hybrid intelligence systems for green transportation.