We analyze robotized warehouses, where robots move shelves between a storage area and output stations. At every output station there is a person – the picker – who takes items from the shelves and packs them into boxes according to customers' orders. The content of each shelf is different. The customer's orders are random. As a result, some shelves are used more frequently than the others. When the picker does not need the shelf any more, a robot carries this shelf back to the storage area. In the storage area, there are several free places and we need to decide where to put this shelf. Every decision influences the next one. And the sequence of these decisions influences the overall efficiency of the warehouse. We look for an optimal sequence. We develop a simple mathematical model in order to better understand the complex real-world warehouse and to develop optimization algorithms. In this model we neglect refilling of the shelves, because it happen much less frequently than emptying of the shelves at the output stations. We consider multiple scenarios: when we know exactly the future orders (offline), and when we do not know them (online). We test our results with a simulation framework.