The deployment of sidewalk robots or automated guided vehicles (AGVs) offers significant potential to reduce traffic congestion and operational costs in last-mile logistics. Delivery systems using a two-layer structure, with trucks in the first layer and sidewalk robots at stations in the second, are modeled as Two-Tier Vehicle Routing Problems with Robot Stations (2T-VRP-RS). In this paper, we extend the 2T-VRP-RS to introduce the Two-Tier Multi-Depot Vehicle Routing Problem with Robot Stations and Time Windows (2T-MDVRP-RS-TW). We propose a Mixed-Integer Linear Programming (MILP) formulation and a Multi-Start Iterated Local Search metaheuristic with Crossover and Repair (MS-ILS-CR) operators. The MS-ILS-CR combines three strategies: a Multi-Start mechanism, a Crossover operator, and a Repair operator adapted from Variable Neighborhood Search, Genetic Algorithms, and Adaptive Large Neighborhood Search, respectively. We introduce new instance sets to study the 2T-MDVRP-RS-TW. Results showed that the MILP formulation is competitive for small instances of 10 customers. However, the metaheuristic consistently achieved better solution quality and lower computational times, with improvements of over 20% in the average objective values for instances of up to 30 customers. The MS-ILS-CR outperformed similar metaheuristic methods, improving its initial solutions by 66%. Furthermore, a sensitivity analysis showed a limited impact of time-window variations on the objective values. Nevertheless, a significant sensitivity was found to the depot configurations. Hence, increasing the number of depots from 1 to 3 improved the average objective values by 56%. Additionally, the MS-ILS-CR performance was further enhanced with more crossover iterations, but at the cost of higher computational times.

Drone delivery is a fast and innovative method for delivering parcels, food, and medical supplies. Furthermore, this low-contact delivery mode contributes to reducing the spread of pandemic and vaccine-preventable diseases. Focusing on the delivery of medical supplies, this paper studies optimizing the distribution operations at a drone hub that dispatches drones to hospitals located at different geographic locations. Each hospital generates stochastic demands for medical supplies to be covered. This paper classifies stochastic demands based on the distance between hospitals and the drone hub. Satisfying the demands requires flying over different ranges, which is directly related to the amount of charge of the drone batteries. We develop a stochastic scheduling and allocation problem with multiple classes of demand and model the problem using a finite Markov decision process approach. We provide exact solutions for the modest sizes instances using backward induction and discuss that the problem suffers from the curses of dimensionality. Hence, we provide a reinforcement learning method capable of giving near-optimal solutions. We perform a set of computational tests using realistic data representing a prominent drone delivery company. Finally, we analyze the results to provide insights for managing drone hub operations and show that the reinforcement learning method has high performance compared with the exact and heuristic solution methods.

Nowadays, many companies face shorter product life cycles, increasing the need to properly forecast demand for newly introduced products. These forecasts allow them to support operational decisions, such as procurement and inventory control. However, forecasting the demand of new products is challenging compared to existing products, since historical sales data is not available as an indicator of future sales. Moreover, little attention has been paid in literature to quantitative methods for new product forecasting, especially with respect to quantifying the uncertainty in demand. In this paper, we present a novel demand forecasting method denoted by DemandForest, which combines K-means, Random Forest, and Quantile Regression Forest. This machine learning-based approach combines the historical sales data of previously introduced products and product characteristics of existing and new products to make prelaunch forecasts and support inventory management decisions for new products. DemandForest clusters and predicts demand patterns, and predicts the quantiles of the total demand during an introduction period. We validate and illustrate our approach for forecasting and inventory management using real-world data sets of several companies. Compared to several benchmark methods, DemandForest provides the most accurate predictions, resulting in potential inventory savings of around 15% depending on lead times and service levels.