“Keeping up with changing customer demand”: An adaptive data-driven approach for storage and repositioning decisions in automated g warehouses

In warehouses, products are often not stored in their optimal positions, elongating retrieval and order picking time. A main reason is that storage assignment is based on historical demand frequency, whereas current demand patterns might just differ. However, as many warehouses are now automated or robotized, opportunities exist to dynamically and opportunistically reposition product loads based on real known demand and still reduce the makespan (the total time needed for retrieval, storage, and optional repositioning). We investigate the optimal retrieval of a known block of requests by explicitly additionally allowing in-between repositioning options. Surprisingly, in spite of the extra work and time involved, we show opportunistic repositioning may indeed be beneficial for reducing the makespan. We study the problem for two automated unit-load storage warehouses: automated storage and retrieval (AS/R) crane-based systems and robotic mobile fulfillment (RMF) systems, which have different travel metrics for the retrieval robots. The data-driven storage and repositioning (DDSR) problem, formulated as an integer linear program, leverages actual customer order data. The problem appears to be intractable for realistic systems due to the combinatorial nature of the possible repositions. We then reformulate the model, making it more tractable for moderate-sized problems. This model appears to beat real-life storage assignment heuristics like closest-open location assignment or demand-frequency class-based storage (even when these have full foresight of demand changes). The benefits appear to be around a 14%-30% shorter makespan, depending on the number of loads to be retrieved. For larger rack space utilization, the benefits decrease (since there are fewer options for repositioning). The method is sufficiently fast to be used in real warehouse systems, e.g. , by using a rolling horizon policy where repositions are calculated for the next block of requests while the current requests are executed. Our method offers managers an additional powerful tool to reduce system response time and thereby increase throughput capacity by smarter scheduling of their automated equipment and more efficient use of available storage space.