Scholarlink

<p>The accelerating growth of e-commerce and the emergence of smart urban infrastructure demand transformative approaches in last-mile delivery. Traditional systems are constrained by inefficiencies in energy consumption, limited adaptability, and poor integration with existing transportation networks. To address these challenges, this study introduces a digital twin–driven intelligent delivery framework that leverages cyber-physical synchronization for autonomous operations in smart urban infrastructure. The system integrates real-time sensor data with predictive modelling to build a dynamic digital replica of the delivery environment. This digital twin enables precise coordination between Autonomous Delivery Robots (ADRs) and public transportation systems. Central to this framework is a Multi-Agent Reinforcement Learning (MARL) engine, which enables decentralized task allocation and adaptive route optimization across dynamic delivery nodes. A time-expanded directed graph (TEDG) models key infrastructure, including depots, transit stops, and consumer destinations, under various constraints such as battery levels, service windows, and traffic congestion. Digital Twins reflect operational states in real time to inform decision layers, while Adaptive Graph Neural Networks (GNNs) process spatiotemporal patterns to support localized autonomy. Additionally, Large Language Models (LLMs) forecast demand trends and optimize delivery scheduling. Experimental evaluations demonstrate the model’s robust scalability and responsiveness. The average delivery time rises moderately from 110 milliseconds to 150 milliseconds when the node count increases from 10 to 50, while battery usage ranges from 45 mAh to 65 mAh. Despite environmental complexities, the successful delivery rate remains above 90%, adaptability drops only slightly from 94% to 85%, and decision latency extends from 12 milliseconds to 22 milliseconds. The system maintains high operational throughput, with the makespan increasing from 175&#xa0;s to 215&#xa0;s under stress conditions. For <i>logistics providers</i>, the framework reduces routing costs, improves battery efficiency, and enhances predictive demand fulfillment. For <i>city officials</i>, the results demonstrate measurable improvements in congestion management and emission reduction through integration with public transport networks. For <i>consumers</i>, the system ensures faster, more reliable deliveries, personalization options, and real-time adaptability with ecological responsibility. These outcomes underscore the transformative potential of integrating Digital Twins, MARL, GNNs, and LLMs for developing agile, energy-efficient, and stakeholder-driven decision-support systems, positioning this approach as a promising enabler of autonomous logistics in digitally connected smart manufacturing and urban mobility ecosystems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimizing Last-Mile Logistics with Digital Twins and Multi-agent Learning: Enhancing Efficiency, Customer Satisfaction, and Urban Integration

  • 期刊论文

  • 发表日期：

    2025年11月25日

查看全文
  • Redha J. Shaker,
  • Fadheela Hussain,
  • Vishwas Chakranarayan,
  • Fayzeh Abdulkareem Jaber

摘要

The accelerating growth of e-commerce and the emergence of smart urban infrastructure demand transformative approaches in last-mile delivery. Traditional systems are constrained by inefficiencies in energy consumption, limited adaptability, and poor integration with existing transportation networks. To address these challenges, this study introduces a digital twin–driven intelligent delivery framework that leverages cyber-physical synchronization for autonomous operations in smart urban infrastructure. The system integrates real-time sensor data with predictive modelling to build a dynamic digital replica of the delivery environment. This digital twin enables precise coordination between Autonomous Delivery Robots (ADRs) and public transportation systems. Central to this framework is a Multi-Agent Reinforcement Learning (MARL) engine, which enables decentralized task allocation and adaptive route optimization across dynamic delivery nodes. A time-expanded directed graph (TEDG) models key infrastructure, including depots, transit stops, and consumer destinations, under various constraints such as battery levels, service windows, and traffic congestion. Digital Twins reflect operational states in real time to inform decision layers, while Adaptive Graph Neural Networks (GNNs) process spatiotemporal patterns to support localized autonomy. Additionally, Large Language Models (LLMs) forecast demand trends and optimize delivery scheduling. Experimental evaluations demonstrate the model’s robust scalability and responsiveness. The average delivery time rises moderately from 110 milliseconds to 150 milliseconds when the node count increases from 10 to 50, while battery usage ranges from 45 mAh to 65 mAh. Despite environmental complexities, the successful delivery rate remains above 90%, adaptability drops only slightly from 94% to 85%, and decision latency extends from 12 milliseconds to 22 milliseconds. The system maintains high operational throughput, with the makespan increasing from 175 s to 215 s under stress conditions. For logistics providers, the framework reduces routing costs, improves battery efficiency, and enhances predictive demand fulfillment. For city officials, the results demonstrate measurable improvements in congestion management and emission reduction through integration with public transport networks. For consumers, the system ensures faster, more reliable deliveries, personalization options, and real-time adaptability with ecological responsibility. These outcomes underscore the transformative potential of integrating Digital Twins, MARL, GNNs, and LLMs for developing agile, energy-efficient, and stakeholder-driven decision-support systems, positioning this approach as a promising enabler of autonomous logistics in digitally connected smart manufacturing and urban mobility ecosystems.