Kurly Maximizes Logistics Picking Efficiency with Digital Twin and Genetic Algorithm

Kurly revealed its process of building a Digital Twin and applying a genetic algorithm to solve a logistics optimization problem. First, to create a virtual logistics center, the picking process was divided into a non-stochastic part with clear rules and a stochastic part that varies according to worker judgment.

Assuming that volume matters more than weight as the criterion for a worker's judgment on filling a basket, the basket volume distribution was estimated using MLE (Maximum Likelihood Estimation) based on historical data. To improve the fit of the Weibull distribution, a Continuous uniform Mixture distribution was combined to correct for outliers, and it was confirmed that the variation pattern of the number of virtual baskets generated matched that of the actual number of baskets.

Processing Speed Prediction Modeling

The variables affecting basket processing speed (Y) were defined as the total number of products in a basket (X1), the number of product types (X2), the overall workload (X3), and worker proficiency (X4). For experimental design, X3 was controlled as busy hours and X4 as average proficiency, and regression analysis was performed using the remaining X1 and X2. Calculations based on about 75,000 pieces of actual data produced an R-squared of 0.8 or higher, proving that processing speed can be successfully calculated just by knowing the product composition.

AnyLogic Simulation and Field Application

Using AnyLogic software, a full-scale QPS Digital Twin was implemented. A simulation was conducted in which, when a virtual basket arrives at the QPS entrance via the conveyor, the workstation order is determined according to rules, and the basket moves on after staying for the processing time corresponding to its product composition. As a result, applying the genetic algorithm reduced the total processing time of order groups by about 3% compared to the existing algorithm.

Kurly is currently applying the genetic algorithm optimization program at one actual logistics center. According to on-site feedback, the number of product types has decreased for order groups of the same size, improving overall work efficiency, and this positive impact is being felt noticeably on-site.