A DRL approach for last-mile delivery with robot fleets

In recent years, the optimization of delivery systems in urban environments has gained increasing interest, particularly due to advancements in Reinforcement Learning (RL) and Deep Reinforcement Learning (DRL) techniques. This paper concerns the application of Deep Reinforcement Learning (DRL) algorithms within cooperative multi-agent frameworks for optimising urban delivery systems using realistic traffic simulations (e.g., SUMO). The existing research explores the use of RL or DRL for routing or planning individual vehicles but lacks a detailed evaluation of how cooperative multi-agent training affects system-wide efficiency, beyond the performance of individual agents. This work explores the application of DRL algorithms in a multi-agent simulation environment, with the goal of improving the efficiency of parcel deliveries. Using Simulation Urban Mobility Environment, a traffic simulation platform, a model is developed, specifically an autonomous delivery system where each vehicle cooperates to optimise the delivery tasks. The focus on the cooperative approach, examining in detail the effect of agents’ training on delivery time and operational costs. The selected approach has a significant impact on delivery efficiency and represents a promising solution to optimise autonomous transport systems.