Abstract
<title>Abstract</title> <p>The rapid electrification of urban last-mile logistics has created an acute need for charging strategies that are simultaneously proactive, stakeholder-aware, and sensitive to real-time operating context. We propose an Agentic AI framework integrating Context-Aware Multi-Agent Deep Reinforcement Learning (CAMA-DRL) to optimise charging infrastructure for last-mile delivery e-bikes in urban micro-mobility ecosystems. Four autonomous stakeholder agents—delivery operators, charging station operators, fleet planners, and environmental regulators—coordinate via Deep Q-Networks (DQN) conditioned on real-time contextual signals including delivery demand, battery state-of-charge, renewable energy availability, and traffic conditions. Unlike reactive rule-based controllers, each agent plans across an extended delivery horizon and shares a common reward structure that internalises the system-wide consequences of locally optimal decisions, enabling coordinated rather than competing charging behaviour. Evaluated in a high-fidelity London urban simulation, CAMA-DRL achieves a 28% increase in fleet utilisation, 35% reduction in station congestion, 32% improvement in delivery punctuality, and a 42% CO2 emission reduction compared to rule-based and single-agent baselines. A component-wise ablation further isolates a 5–8% contribution attributable specifically to real-time context encoding beyond multi-agent coordination alone, confirming that the two mechanisms are synergistic rather than merely additive. These results establish Agentic CAMA-DRL as a scalable, deployment-ready solution for sustainable urban last-mile logistics, and the modular architecture suggests a clear migration path towards mixed e-bike and electric-van fleets sharing common charging hubs.</p>