Heterogenous Multi-agent Reinforcement Learning for Last-mile Delivery Optimization

Session Number

Project ID: CMPS 31

Advisor(s)

Ankit Agrawal, Saint Louis University

Discipline

Computer Science

Start Date

17-4-2024 9:40 AM

End Date

17-4-2024 9:55 AM

Abstract

The surge of e-commerce demands innovative solutions to streamline last-mile delivery logistics. Autonomous delivery vehicles (ADVs) offer a promising avenue, also combatting the lack of delivery drivers. However, their success hinges on effectively managing the complexities arising from diverse delivery modes (e.g., aerial and ground-based) in obstructed or constrained environments. Traditional Multi-Agent Reinforcement Learning (MARL) approaches may not optimally coordinate heterogeneous agents, impacting delivery efficiency and cost. This research addresses this gap by investigating Heterogeneity-aware MARL (H-MARL) techniques to optimize delivery routes and resource allocation for mixed-fleet ADVs. Through computational benchmarking, we hypothesize that H-MARL algorithms can significantly outperform traditional MARL algorithms in delivery speed and cost reduction due to their ability to consider the aerial and land agents' unique capabilities and constraints. This work has the potential to revolutionize urban goods delivery by informing the design of intelligent, multi-modal autonomous logistics systems.

Share

COinS
 
Apr 17th, 9:40 AM Apr 17th, 9:55 AM

Heterogenous Multi-agent Reinforcement Learning for Last-mile Delivery Optimization

The surge of e-commerce demands innovative solutions to streamline last-mile delivery logistics. Autonomous delivery vehicles (ADVs) offer a promising avenue, also combatting the lack of delivery drivers. However, their success hinges on effectively managing the complexities arising from diverse delivery modes (e.g., aerial and ground-based) in obstructed or constrained environments. Traditional Multi-Agent Reinforcement Learning (MARL) approaches may not optimally coordinate heterogeneous agents, impacting delivery efficiency and cost. This research addresses this gap by investigating Heterogeneity-aware MARL (H-MARL) techniques to optimize delivery routes and resource allocation for mixed-fleet ADVs. Through computational benchmarking, we hypothesize that H-MARL algorithms can significantly outperform traditional MARL algorithms in delivery speed and cost reduction due to their ability to consider the aerial and land agents' unique capabilities and constraints. This work has the potential to revolutionize urban goods delivery by informing the design of intelligent, multi-modal autonomous logistics systems.