Kelly Builds Digital Twin-Based Simulation of Logistics Center Picking Process

Kelly's Data Service Development Team built a digital twin of the logistics center picking process using M&S (Modeling and Simulation) technology. To overcome the constraints of real-world field experiments and the limitations of historical data analysis, they developed a system that tests scenarios and seeks optimal solutions in a virtual environment.

Simulation Modeling Structure

The simulation defines workers' state changes and transitions based on the DEVS formalism, and was implemented using AnyLogic, a Java-based tool. The model is divided into a Process layer, which is generated for the number of workers, and a Main layer that manages it. Workers move along the shortest path while accounting for physical obstacles, and wait when their destination is congested, reflecting real on-site constraints. Parameters such as movement speed and task time are dynamically determined based on probability distributions.

Validation and Use Cases

A Kolmogorov-Smirnov test comparing actual data from a specific day with the simulation results showed no statistical difference between the two distributions (p=.99), confirming the model's fidelity. This led to the following insights:

• Path Comparison: When the number of SKU visits is low, a Z-shaped path is advantageous, but when it is high, a U-shaped path results in a shorter distance.
• Inventory Placement: Placing high-frequency SKUs near the entrance reduces average task time by 5% when 6 workers are deployed. However, when the number of workers increases to 10 or more, bottlenecks near the entrance intensify, requiring proper distribution.

Kelly plans to apply reinforcement learning and other techniques going forward to further advance production plan optimization, and aims to predict and improve productivity through digital transformation across its logistics centers.