European Journal of Operational Research 288 (2021) 80–97

Contents lists available at ScienceDirect

European Journal of Operational Research

journal homepage: www.elsevier.com/locate/ejor

Production, Manufacturing, Transportation and Logistics

Introducing split orders and optimizing operational policies in robotic

mobile fulfillment systems
Lin Xie a,∗, Nils Thieme a, Ruslan Krenzler a, Hanyi Li b
a
Leuphana University of Lüneburg, Universitätallee 1, 21335 Lüneburg, Germany
b
Beijing Hanning Intelligent Tech Co., Ltd, Room 107–108, No.9 Kexing Road, Fengtai, Beijing, China

a r t i c l e i n f o a b s t r a c t

Article history: In robotic mobile fulfillment systems, human pickers don’t go to the inventory area to search for and
Received 26 December 2019 pick the ordered items. Instead, robots carry shelves (called “pods”) containing ordered items from the
Accepted 17 May 2020
inventory area to picking stations. At the picking stations, pickers put ordered items into totes; then these
Available online 26 May 2020
items are transported to the packing stations. This type of warehousing system relieves the human pick-
Keywords: ers and improves the picking process. In this paper, we concentrate on decisions about the assignment of
logistics pods to stations and orders to stations to fulfill picking for each incoming customer’s order. In previous
MIP models research for an RMFS with multiple picking stations, these decisions are made sequentially with heuris-
Integrated operational optimization tics. Instead, we present a new MIP-model to integrate both decision problems. To improve the system
Robotic mobile fulfillment systems performance even more, we extend our model by splitting orders. This means parts of an order are al-
Split orders lowed to be picked at different stations. To the best of the authors’ knowledge, this is the first publication
on split orders in an RMFS. And we prove the computational complexity of our models. We analyze differ-
ent performance metrics, such as pile-on, pod-station visits, robot moving distance and throughput. We
compare the results of our models in different instances with the sequential method in our open-source
simulation framework RAWSim-O. The integration of the decisions brings better performances, and al-
lowing split orders further improves the performances (for example: increasing throughput by 46%). In
order to reduce the computational time for a real-world application, we have proposed a heuristic.
© 2020 The Author(s). Published by Elsevier B.V.
This is an open access article under the CC BY license. (http://creativecommons.org/licenses/by/4.0/)

1. Introduction 2007). The unproductive searching and traveling times require the
picker-to-parts system to have a large workforce, especially for
The most important and time-consuming task in a warehouse companies which have millions of small items in large warehouses,
is the collection of items from their storage locations to fufill cus- such as e-commerce companies like Amazon or retailers like Zara
tomer orders. The process is called order picking, which may con- and Walmart, which provide both brick-and-mortar and online
stitute about 50–65% of the operating costs. Therefore, the order shops. Dependent on the type of retailers, they are facing many
picking is considered as the highest-priority area for productivity diverse customer orders each day (both single-line and multi-line
improvements (see De Koster, Le-Duc, & Roodbergen, 2007). Due orders). Also, the workforce of such companies is under high pres-
to the increasingly fast-paced economy, it is becoming more and sure due to the long traveling time (see Wulfraat, 2012). Kiva Sys-
more important that the orders are processed in a short time win- tems LLC, now Amazon Robotics LLC, came up with a unique solu-
dow. tion to avoid the unproductive times of human pickers in picker-to-
In a traditional manual order picking system (also called a parts systems; therefore, this solution accelerates the order picking
picker-to-parts system), the pickers spend 70% of their working time process (see Wurman, D’Andrea, & Mountz, 2008). In such system,
on the tasks of search and travel (see Tompkins, 2010; for an robots are sent to carry storage units, so-called ”pods,” from the in-
overview of manual order picking systems see De Koster et al., ventory area and bring them to human operators, who work only
at picking stations. At the stations, the items are picked according
to the customers’ orders. After picking, the robot transports the
∗
Corresponding author.
pod back to the storage area. There are also some other suppli-
E-mail addresses: xie@leuphana.de (L. Xie), nils.thieme@stud.leuphana.de (N. ers of such systems, such as Scallog, Swisslog (KUKA), GreyOrange
Thieme), ruslan.krenzler@leuphana.de (R. Krenzler), hli@hanningzn.com (H. Li). and Hitachi (see Banker, 2016). All of these systems may differ

https://doi.org/10.1016/j.ejor.2020.05.032
0377-2217/© 2020 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license. (http://creativecommons.org/licenses/by/4.0/)
 L. Xie, N. Thieme and R. Krenzler et al. / European Journal of Operational Research 288 (2021) 80–97 81

Table 1
Literatur overview in an RMFS.

decision problems publications

Strategic level:
storage area dimensioning, Lamballais, Roy, and De Koster (2017),
workstation placement Lamballais, Roy, and De Koster (2019)
Tactical level:
number of robots Yuan, Dong, and Li (2016), Yuan and Gong (2017),
Zou, Xu, De Koster et al. (2018), Otten, Krenzler, Xie, Daduna, and Kruse (2019)
Operational level:
decisions for each incoming order about
which robot carries Zhang, Yang, and Weng (2019), Roy, Nigam, de Koster, Adan, and Resing (2019)
which pod Boysen et al. (2017)
along which path Cohen, Uras, and Koenig (2015),Cohen, Wagner, Kumar, Choset, and Koenig (2017),
Merschformann et al. (2017)
to which station
decisions about where to
put the used pod back to Merschformann (2017),
Weidinger, Boysen, and Briskorn (2018),
Krenzler, Xie, and Li (2018)
decision rules Wurman et al. (2008),
Merschformann et al. (2019)
simulation Merschformann et al. (2018)
demonstration Xie et al. (2019)
performance characteristics Hanson, Medbo, and Johansson (2018)

technically in certain aspects, such as the lifting mechanism, but Table 2

The operational decision problems in an RMFS.
they share the same principle of the system (see the description
of such system in Section 2). This system is called robotic mobile Decision problem Description
fulfillment system (RMFS). Order Assignment:
Replenishment Order Assignment (ROA) assignment of replenishment
1.1. Research background orders to replenishment stations
Pick Order Assignment (POA) assignment of pick orders to
picking stations
In an RMFS environment, various optimization and allocation Task Creation:
problems have to be solved in real time. An overview of an Replenishment Pod Selection (RPS) selection of pods that will
RMFS and other automated warehousing systems can be found in move to a replenishment station
to store replenishment items
Azadeh, de Koster, and Roy (2017) and Boysen, de Koster, and Wei-
Pick Pod Selection (PPS) selection of pods to use for
dinger (2018). Here we give you an short literature overview of picking the pick orders assigned
RMFSs in Table 1, which also includes new researches that are not at a picking station
listed in both overview papers. The classification of three levels of Pod Repositioning (PR) assignment of an available
problems is based on Merschformann, Lamballais, de Koster, and storage location to a pod that
is not in the inventory area
Suhl (2019).
Task Allocation (TA) assignment of generated tasks for
In this work, we concentrate on operational problems. As robots in RPS, PPS, PR and
shown in Table 1, there are usually the following decision prob- additional tasks such as idling
lems in the order picking process in an RMFS (described in Path Planning (PP) planning of the paths for the
robots to execute
Wurman et al., 2008 and Merschformann, Xie, & Li, 2018): Each
time a new order arrives, we have to decide which robot carries
which pod along which path to which station to fulfill picking. So
the order is first assigned to a station (pick order assignment, in the assumption that orders have already been assigned to stations
short: POA), and then one or several pods are assigned to that sta- (in other words: without optimization of POA), such as in Boysen,
tion to fulfill that order (pick pod selection, in short: PPS). Robots Briskorn, and Emde (2017).
are assigned to deliver pods to that station (robot task allocation), Due to the importance of POA, POA and PPS should be opti-
while path planning plans the paths for the robots. After a pod is mized together; however, the complexity is increased due to con-
finished at a picking station, we have to decide where to put it sidering all possible pods and multiple picking stations (see ex-
back in the inventory area (pod repositioning). And there are two amples in Section 2.2). Therefore, the authors in Wurman et al.
decisions for the replenishment process, namely replenishment or- (2008) and Merschformann et al. (2019) provided heuristics to
der assignment and replenishment pod station. Table 2 describes solve POA and PPS sequentially, first POA and then PPS. We list
the operational problems in an RMFS. here the contributions of our work:

1.2. Contributions and paper structure • We develop a new mathematical model to solve integrated POA
and PPS for multiple stations.
We concentrate in this paper on the picking process, especially • We extend our integrated model to allow split orders, so not
POA and PPS. In the studies of the throughput performance of deci- all parts of an order are picked at the same station. To the best
sion rules for multiple online decision problems in Merschformann of the authors’ knowledge, this is the first publication on split
et al. (2019), they concluded that POA should be paid more at- orders in an RMFS. The literature review of split orders can be
tention in the literature and practice, since it affects the through- found in Section 2.3.
put performance of RMFS the most. In Boysen et al. (2018), they • We develop a heuristic for our new models to solve a real-
mentioned that the existing research into order picking is under world instance.
82 L. Xie, N. Thieme and R. Krenzler et al. / European Journal of Operational Research 288 (2021) 80–97

Retrieval

Storage

Storage
Retrieval Order picking
Replenishment

Order packing
Retrieval

Order picking
Transport to packing
staon

Fig. 1. The central process of an RMFS.

Table 3 The pods are transported by robots between the inventory area
Terms related to orders.
and workstations. Fig. 1 shows the central process in a simplified
Term Description RMFS from replenishment to packing:
SKU stock keeping unit • Retrieval process: After the arrival of a replenishment order,
order line one SKU with the ordered quantity
item a physical unit of one SKU
robots carry selected pods to a replenishment station to store
pick order a set of order lines from a customer’s order units in pods. Similarly, after receiving a pick order, robots carry
split order a pick order that is separated into several parts selected pods to a picking station, where the items for the order
replenishment order a number of physical units of one SKU lines are picked. Note that in order to fulfill pick orders, several
backlog all unfulfilled orders
pods may be needed, since orders may have multiple lines. The
items in (parts of) an order are picked into a tote.
Table 4 • Storage process: After a pod has been processed at one or more
The central components of an RMFS. stations, it is brought back to a storage location in the storage
Component Description area. The retrieval and storage processes are based on Hoffman
et al. (2013).
pods movable shelves, on which inventory is stored
storage area the inventory area where the pods are stored • Transport to packing stations: Once a tote is filled, it is trans-
workstations: ported by a conveyor to packing stations for packing.
picking stations where pickers pick the order items • Packing process: If all items in an order are contained in a tote,
packing stations where packers pack the pick order items and
packers are prompted by computer to select the correct-sized
the split orders are consolidated
replenishment stations where replenishment items are stored to pods box and pack the items. A split order has items delivered via
robots moving underneath pods and carrying them to multiple totes, since the items are picked by different pickers
workstations (picking stations). In this case, packers first sort items from a
conveyors transporting the pick orders between picking and tote to a correct-sized box on the shelf so that the items from
packing stations to finish packing
that order are grouped together. Packers are prompted by com-
puter to put the box for the split order on one given position
on the shelf, and later to find the box in the shelf to put the
• We extend our open-source simulation framework RAWSim-O rest of that order into the box. We use the term shelf to clar-
to evaluate the decisions made based on our new models. ify that they might be different to the pods, since they do not
need to be moved. Once all the items of a split order are in
This paper is organized as follows: In the next section, we de-
a box, the packer packs the box, and a space is open for the
scribe the RMFS and operational decision problems in detail. Af-
next split order. This packing process is based on the packing
ter that, we will describe our idea of integrated POA and PPS and
process in Amazon (see Toister, 2017).
split orders with examples. In Section 3, a mathematical model of
integrated POA and PPS and the extensions with split orders are
2.2. POA and PPS
described. We present simulation evaluations in Section 4.2. Fi-
nally, we draw conclusions and give pointers for further research
We concentrate in this paper on the online decisions for POA
in Section 5.
and PPS for multiple stations. For better clarification, online in our
case does not mean that we consider orders that have arrived over
2. Problem description with examples a number of time periods. In order to have enough optimization
potential, we assume that we know all orders in advance (for ex-
In this section, we first describe the RMFS, and the decision ample, we gather all orders that have come in the last hour in the
problems in an RMFS. After that, we will describe our idea of inte- backlog). But our methods can support the decisions for the new
grated POA and PPS and split orders with examples. incoming orders in each period (see Section 4.3). Instead, online
decisions in our case means that we have to make decisions for
2.1. RMFS both problems for each new period t. Period t − 1 is changed to
t if some jobs are finished at the stations while there are unful-
Firstly, we define some terms related to orders before explain- filled orders in the backlog. Also, situations such as the inventory
ing the processes in an RMFS in Table 3. of pods and the positions of pods in the queues at stations can and
The central components of an RMFS are listed in Table 4. will change from t − 1 to t. They are important for the POA and PPS
 L. Xie, N. Thieme and R. Krenzler et al. / European Journal of Operational Research 288 (2021) 80–97 83

decisions and are hard to calculate exactly in advance, since errors ders, both pods from the storage area, pods 3 and 4, are needed
and delays in previous time periods can affect them. For these rea- at each station. After items from pods 1 and 2 are picked, they
sons, we make the decisions for the integrated POA and PPS just are returned to the storage area. In the PPS, pod 3 visits station 1,
before the respective time period starts. This allows us to react to while pod 4 visits station 2 (Fig. 2(c)). After picking in both sta-
the current situation and take errors or delays from previous pe- tions, pods 3 and 4 switch stations so that the last item of each
riods into account. Furthermore, we test our results in a simula- order can be picked (Fig. 2(d)). In total, 6 pod-station visits were
tion framework, which provides us with the actual information for necessary to fulfill both orders in this example, therefore the pile-
each time period. In the literature, the online decisions for POA and on can be calculated as 12 picks/6 pods = 2 picks/pod.
PPS are usually solved sequentially for multiple stations (first POA, Integrated PPS and POA In the integrated PPS and POA approach,
then PPS; see Wurman et al., 2008, Merschformann et al., 2019). we have more information while assigning orders to stations, since
The sequential approach limits the information for both decisions. pods and orders are assigned to stations at the same time. This
Therefore, the decisions for POA and PPS should be made simulta- allows us to find optimal solutions that might not be intuitive at
neously to achieve optimum performance, so we can use informa- first glance and would not be found by the sequential POA and PPS.
tions about all pods and orders, including those that are assigned Note that we use information about all pods, including assigned
to stations and are unassigned. However, the complexity of solving ones at stations and unassigned ones in the storage area, but this
integrated POA and PPS is increased, and the integrated problem is way increases the complexity to solve both problems.
NP-hard (see the proof in Section 3). We will describe the differ- Using the same initial state as in the previous explanation of
ence between sequential and integrated problems with Example 1. the sequential POA and PPS in Example 1 (see Fig. 3(a)), we in-
To the best of the authors’ knowledge, this is the first publication tegrate these two decisions and assign orders and pods such that
about the integrated POA and PPS for multiple stations. the number of pod-station visits is minimized. This leads to the
Objective Before we explain POA and PPS with examples, we assignment of orders 1 and 2 and pod 3 to station 1, and orders 3
want to firstly explain our objective of POA and PPS. An RMFS and 4 and pod 4 to station 2 (see Fig. 3(b) and 3(c)). This results
with higher throughput is better. Based on the performance anal- in a pile-on of 3 (12 picks/4 pods) compared to 2 (12 picks/6 pods)
ysis of decision rules for multiple online decision problems in in the sequential example and only requires 4 pod-station visits to
Merschformann et al. (2019), they found that a high pile-on (the fulfill all orders instead of 6.
number of picks per handled pod) and a short distance traveled by Based on this example, we can see the benefit of integrating
robots together is a indicator for the success of a decision rule ap- POA and PPS by using information about the inventory of all pods
plied in an RMFS (to achieve high throughput). In order to achieve in these decisions. Therefore, we present a mathematical model in
that, we aim at minimizing the number of visits by pods to sta- this paper that integrates POA and PPS for multiple stations and
tions (in short: pod-station visits) for each decision of the POA and takes information about the inventory of all pods into account.
PPS problems. If fewer pod-station visits are needed for the given
orders, a shorter distance traveled by the robots can be achieved, 2.3. Allowing split orders in our integrated approach
and higher pile-on can be expected to finish the given orders. Fur-
thermore, a smaller number of pod-station visits also causes fewer In our integrated approach mentioned above, an order is only
changes of pods, and the waiting time of human pickers between allowed to be assigned to a single station. The second contribution
changes of pods is reduced (see Boysen et al., 2017; they also min- of this paper is to allow split orders in our integrated approach.
imize pod-station visits for the assignment of pods to orders). An- And we also prove that the models to allow split orders are NP-
other time component of human pickers is retrieving the items hard (see Section 3). A split order means that we divide an order
from their shelves, which is considered to be fixed. Due to the re- into two or more parts for picking (perhaps at different stations). A
duced waiting time, more orders can be handled within a mini- similar term, “splitting orders,” can be traced back to 1979, when
mal time. So, this matches the suggestion in Van Gils, Ramaekers, it was used by Armstrong, Cook, and Saipe (1979). They used split
Caris, and de Koster (2018) for an efficient order picking in picker- orders to keep batch sizes constant in batch picking. In Il-Choe and
to-parts systems. Sharp (1991) and De Koster, Le-Duc, and Zaerpour (2012) split or-
Example 1. Fig. 2(a) illustrates a small problem to fulfill four or- ders are used as part of the zoning in traditional picker-to-parts
ders 1, 2, 3 and 4. The different colors represent different SKUs warehouses, in which a storage area is split into multiple parts
(stock keeping units). For simplicity, the quantity of each SKU in (called zones), each with a different order picker. When an order
the orders is one. We have in total two picking stations. There is contains several SKUs that are stored in different zones, the SKUs
two empty totes at each station. In this example we assume that for the order are picked separately in each zone and merged later
each tote can hold three items. Pod 1 is currently at station 1 and for shipping. To the best of the authors’ knowledge, this is the first
pod 2 at station 2, while pods 3 and 4 are in the storage area. publication on split orders in an RMFS.
According to the following example, we expect allowing split
Sequential POA and PPS In the sequential POA and PPS, we use orders in an RMFS provide a better solution.
the same decision rule, Pod-Match, as in Merschformann et al.
(2019), which assigns the orders from the backlog to a station so Example 2. Fig. 4 illustrates the decision problem when assign-
that the items for the orders best match the pods that are already ing orders and pods to two picking stations. We have one empty
assigned to that station. Note that there is another more common tote at each station, while we have two identical orders 1 and 2 in
decision rule in Merschformann et al. (2019) (called Common-Lines) the backlog (in Fig. 4(a)). We assume that each tote can hold two
and Wurman et al. (2008), grouping similar orders at picking sta- items. These two orders contain SKUs shown in blue and orange.
tions in POA. However, the decision rule Pod-Match for POA is These two SKUs are located in two different pods, namely pod 1
shown to perform better in Merschformann et al. (2019), since this with the orange SKU and pod 2 with the blue SKU. Fig. 4(b) shows
rule uses information about assigned pods at stations in addition the optimal solution to the problem without split orders. We need
to information about orders in the backlog. pod 1 to visit station 1 and pod 2 to visit station 2; after that, pod
In Example 1, in the POA problem we assign orders 2 and 3 to 2 visits station 1 and pod 1 visits station 2. In total, we need four
station 1 (Fig. 2(b)), since two of their items can be picked from visits by pods to the stations to fulfill both orders. Instead, if we
pod 1 – the pod that is already at station 1. For the same reason, split orders 1 and 2 into blue and orange parts (see Fig. 4(c)), the
we assign orders 1 and 4 to station 2. To fulfill the assigned or- blue ones can be picked from pod 2 at station 2, while the orange
84 L. Xie, N. Thieme and R. Krenzler et al. / European Journal of Operational Research 288 (2021) 80–97

orders

1 2 3 4

2 3

pod 3 pod 1 station 1 pod 3 pod 1 station 1

1 4

pod 4 pod 2 station 2 pod 4 pod 2 station 2

(a) Initial state (b) POA: orders 2 and 3 → station 1,

orders 1 and 4 → station 2

2 3 3

pod 1 pod 3 station 1 pod 1 pod 4 station 1

1 4 1

pod 2 pod 4 station 2 pod 2 pod 3 station 2

(c) PPS I: pod 3 → station 1, pod 4 → (d) PPS II: pod 4 → station 1, pod 3 →
station 2 station 2
Fig. 2. An example for the sequential POA and PPS (best viewed in color).

ones can be picked from pod 1 at station 1. This allows both orders Pregenerated orders We store pregenerated orders in the back-
to be fulfilled with only two visits by pods to the stations instead log before the beginning of optimization, so we get the same or-
of four. Note that in this paper we don’t use one empty tote for ders for testing different approaches. From time to time, no new
exactly one order, but for several items, to enable comparison be- incoming orders are stored to the backlog for the optimization. Our
tween the solutions with and without split orders. We will explain approaches terminate if the backlog is empty.
more about this in Section 3.1 in the paragraph Capacity of a pick- Split order Splitting an order means separating the original or-
ing station. der into two or more parts (up to the number of SKUs in the or-
der). If an order is not split, we ensure that all order lines in that
However, the split orders might need additional consolidation order are assigned for picking at the same station (within a time
time in packing stations, where customers’ orders are packed and period). If an order is split, this constraint is relaxed by allowing
ready for shipping. We ignore this time in our study, we will de- order lines for that order to be assigned to more than one picking
scribe that in Section 3.1. station or more than one time period. There are two variants of a
split order:
3. Mathematical model
split among stations: all order lines for a pick order are as-
In this section, we describe the assumptions in Section 3.1 be- signed in the same period but may be assigned to different
fore we present our mathematical model of integrated POA and picking stations (see Example 1 in Section 1)
PPS (we call it the integrated model), and extend it with two vari- split over timesteps: order lines for a pick order may be as-
ants of allowing split orders. signed in different time periods and to different picking sta-
tions (see an example in Appendix A)
3.1. Assumption
Capacity of a picking station Commonly, the capacity of a pick-
SKUs All different SKUs in orders are available in pods. We as- ing station is defined as the number of orders that can be handled
sume that the quantity of the order line for each SKU is one. This at a time (order capacity). According to Wulfraat (2012), the typi-
assumption is consistent with common practice, since the number cal station can support 6 to 12 orders to be picked at a time. The
of items per order line is low. If a pod contains a SKU, then we introductory example of split orders shows that traditional order
assume that there are enough items in that pod to fulfill all orders capacity is incompatible with split orders, since simply counting
for that SKU. the number of assigned orders does not work anymore when only
 L. Xie, N. Thieme and R. Krenzler et al. / European Journal of Operational Research 288 (2021) 80–97 85

Fig. 3. Same example as in Fig. 2 (see Fig. 3(a)), but the decision is made by the integrated POA and PPS (see Fig. 3(b) and 3(c)) (best viewed in color).

parts of an order are assigned to the station. Instead, we introduce easily extended to support a limited packing capacity, as shown in
in this paper a new way to define the capacity of a picking station Section 4.3.
– limited by the number of items to be handled at a time – that Conveyor We assume that the conveyors between picking and
works for both, whole orders and split orders. We call this type packing stations are long enough to temporarily store orders and
of capacity item capacity. Another advantage of item capacity is a parts of them. The conveyors serve as a buffer to synchronize the
fairer distribution of workload among all stations, since the num- picking and packing stations.
ber of assigning items equals the number of picks. Note that the Consolidation time for split orders We consider the consolidation
number of items in each order differs and reflects the number of time as the additional time for the packer to pack the split orders.
picks. Based on the packing process in Amazon (described at the end of
Capacity of a packing station Orders that are not split can be Section 2.1, and the part of Fig. 4 on the right-hand side), shelves
packed directly by packers as soon as they arrive at the packing are used in the packing station to temporarily store split orders.
stations. Split orders require storage space on shelves at packing Each time a part of a split order arrives, the packer puts this part
stations to wait until all parts of the order are picked. Once all into the corresponding box on the shelf. Once all items of the split
parts of a split order are picked, it can be packed and one space order are in the box, the packer packs the box, and a space is open
on the shelf becomes free for the next split order. The capacity for the next split order. This process is almost the same as the pro-
of a packing station is therefore defined as the number of shelves cess for a normal order, i.e. folding a new box, putting all items in
multiplied by the number of boxes which can be stored on a shelf. the box and packing it. The additional time for packers caused by
We set the total capacity of all packing stations to a parameter C, the split orders is the searching time for the correct split-order box
and we assume it is large enough for all necessary split orders on the shelf, which we consider to be minimal since the position of
in this paper. This assumption is supported by the calculation in the box is stored on the computer. Therefore, we ignore this time
Appendix B. In our calculation, up to 78 split orders can be stored in our study.
on a shelf. And usually, in practice, there is more than one packing Maximal order size We assume that every order in the backlog
station. If more split orders are required, then additional shelves can fit into some picking station. That means the maximal item ca-
can be installed at packing stations. However, the situation might pacity of the largest picking station is not smaller than the number
differ from one company to another. Therefore, our model can be of items of the largest order.
86 L. Xie, N. Thieme and R. Krenzler et al. / European Journal of Operational Research 288 (2021) 80–97

orders
1 2

pod 1 station 1

pod 2 station 2

(a) Initial state

pod 1 station 1

pod 2 station 2

1/2
1

pod 1 station 1
pod 2 station 1

1/2
2

pod 2 station 2
pod 1 station 2
(c) With split orders: pod 1 → station
(b) Without split orders: first, pod 1 → 1 and pod 2 → station 2
station 1 and pod 2 → station 2 (upper
part); then, pod 2 → station 1 and pod
1 → station 2 (lower part)

Fig. 4. A solution for without and with split orders (best viewed in color).

Queue There is a queue at each station (for example, in Fig. 3(b), picking station. All pods are in the storage area, so there are no
pod 1 and pod 3 are in the queue at station (1). The space in a pods waiting at picking stations. At t = 1 we start to assign orders
queue is limited. Each time a pod leaves the queue, one pod can from the backlog and pods to picking stations. Each time t changes,
be added at the end of the queue. Pods leave a station once their the current situation in the warehouse (such as which pods are
inventory cannot be used anymore to fulfill any further assigned currently in storage or on their way to stations, free capacity at
orders. If pod 1 leaves the queue, pod 3 moves forwards to the stations, inventory of pods, decreasing order backlog) is updated
picker. and used to compute the next decisions. This way, we can handle
Period Once there is enough free item capacity at a station and errors or delays in the execution of previous decisions. The model
there are unfilled orders in the backlog, the time period is changed described in this section is solved in each period t using informa-
from t to t + 1 for all t ≥ 1. The required amount of free item ca- tion about the current state of the warehouse.
pacity is defined as the capacity that is needed to fit the smallest Shared storage policy Items of the same SKU are randomly
available order. In t = 0, no orders are assigned or picked at any spread over multiple pods. In Boysen et al. (2017), where this pol-
 L. Xie, N. Thieme and R. Krenzler et al. / European Journal of Operational Research 288 (2021) 80–97 87

icy is called mixed-shelves storage, the authors showed that this

policy is efficient. us ∈ Z ≥ 0, ∀s ∈ S (10)
Pods selection Our model computes the smallest possible set of
In the integrated model given above, we aim at minimizing the
pods to fulfill all assigned orders at each station in each period,
number of pods used in each period, while keeping the unused
without considering the distance between the selected pods and
capacity of stations as low as possible. Constraint set (2) sets the
picking stations.
value of yos to 1 for all assigned order/station tuples and ensures
Sequencing of pods During each period t we need to know the
that, if an order is assigned to a station, all order lines in the order
sequence of pods at each station. As our model only calculates the
are assigned to the same station as this order. And if an order is
optimal sets of pods for each station, we use the following policy
not assigned to any station, none of its order lines can be assigned
to create a sequence of pods: Robots begin immediately to carry
to a station. Constraint set (3) ensures that each order can be as-
all assigned pods to the respective stations and the sequencing of
signed to at most one station. Constraint set (4) ensures that the
pods is decided by the order of their arrival at the station. This
number of assigned items equals the amount of available capac-
ensures that station idle times are kept at a minimum.
ity minus the amount of unused capacity us at each station s ∈ S.
Without replenishment We assume that there is enough inven-
Each us increases the value of the cost function (assuming Wu > 0).
tory for all orders, so no replenishment is required.
Constraint set (5) ensures that for each order line of an order that
is assigned to a station, at least one pod p ∈ Pi is assigned to the
3.2. Integrated model
same station. Constraint set (6) ensures that pods that were as-
signed in previous periods and are currently on their way to a sta-
Firstly, we define the notation for the following model.
tion or in a station’s queue stay assigned to that station. Constraint
Sets:
sets (7)–(9) ensure that the respective variables can only have bi-
P Set of pods
nary values while (10) ensures that us can only have non-negative
S Set of currently available picking stations
Ps Set of pods Ps ⊆ P that are currently at station s
integer values.
PiSKU Set of pods PiSKU ⊆ P that include SKU i
O Set of current orders in the backlog
Proposition 1. The integrated model always has a feasible solution
Io Set of SKUs Io ⊆ I that constitutes an order o ∈ O which we can find in polynomial time.

Parameters: Proof. See Appendix C. 

Cs Current capacity of each picking station s ∈ S Proposition 2. The optimal solution of the integrated model is NP-
hard.
Decision variables:
 Proof. See Appendix C. 
1, pod p ∈ P is assigned to station s ∈ S
x ps
0, else
 3.3. Split-among-stations model
1, order o ∈ O is assigned to station s ∈ S
yos
0, else
yios  This model is an extension of the integrated model from the
1,
SKU i ∈ Io of order o ∈ O is assigned to station s ∈ S previous section. Now we allow splitting the SKUs in an order be-
0, else tween two or more stations. We need the following additional de-
us Amount of unused capacity for a station s ∈ S
cision variables.
The integrated model is invoked in the simulation each time the Additional decision variables:
time period t is changed. However, for simplicity the parameter t is 
1, order o ∈ O is assigned
not used in the model. Note that all sets, parameters and decision yo
0, else
variables may change for each time period t. eo the number of additional assigned picking stations for an order o ∈ O
 
Min x ps + Wu · us (1)
 
p∈P s∈S s∈S
Min x ps + Wu · us (1)
p∈P s∈S s∈S
s.t. yos = yios , ∀i ∈ Io, o ∈ O, s ∈ S (2)
s.t. (4 ) − (10 )
 yos ≥ yios , ∀i ∈ Io, o ∈ O, s ∈ S (2.1)
yos ≤ 1, ∀o ∈ O (3) 
s∈S yos ≥ yo, ∀o ∈ O (3.1)
 s∈S

yios + us = Cs , ∀s ∈ S (4) yios = yo, ∀i ∈ Io , o ∈ O (11)
o∈O i∈Io
s∈S

x ps ≥ yios , ∀i ∈ Io, o ∈ O, s ∈ S (5)
yo ≥ yos , ∀o ∈ O, s ∈ S (12)
p∈PiSKU

x ps = 1, ∀ p ∈ Ps , s ∈ S (6) 
yios ≥ yos , ∀o ∈ O, s ∈ S (13)
i∈Io

yos ∈ {0, 1}, ∀o ∈ O, s ∈ S (7)

yo ∈ {0, 1}, ∀o ∈ O (14)
yios ∈ {0, 1}, ∀i ∈ Io, o ∈ O, s ∈ S (8) The cost function (1) and constraints (4)–(10) are carried over
from the previous model. Constraint set (2) is relaxed, so that the
order lines for an order don’t have to be assigned to the same sta-
x ps ∈ {0, 1}, ∀ p ∈ P, s ∈ S (9) tion anymore (see constraint set (2.1)), but yos is still set to 1 for all
88 L. Xie, N. Thieme and R. Krenzler et al. / European Journal of Operational Research 288 (2021) 80–97

Framework
Start End
Agent

Update

Inialize Simulator Output results

Opmizer Visualizer

Opmize Render

Fig. 5. Overview of the simulation process (see Merschformann et al., 2018).

stations the order is assigned to. Constraint set (3) is also relaxed, not assigning all order lines of an order at once (see constraint set
to allow for the assignment of an order to more than one station (11.1)). The new constraint set (15) ensures that the value of ybio is
(see constraint set (3.1)). Constraint set (11) now ensures that if binary.
an order is active (at least one order line is assigned to a station),
all of its order lines have to be assigned to stations. Constraint set Proposition 5. Every solution of the split-among-stations model also
(12) sets the value of yo to 1 for each order that is assigned to solves the split-over-time model. There is always a feasible solution of
at least to a station. Constraint set (13) ensures that an order can the split-over-time model. The split-over-time model always provides
only be assigned to a station if at least one order line of the order a solution that is better than, or equally good as, the split-among-
is assigned to that station. Constraint set (14) ensures that yo is a stations model.
binary variable for each o ∈ O. Proof. See Appendix C. 
Proposition 3. Every solution of the integrated model also solves the
Proposition 6. The optimal solution of the split-over-time model is
split-among-stations model. There is always a feasible solution of the
NP-hard.
split-among-stations model. The split-among-stations model always
provides a solution that is better than, or equally good as, the inte- Proof. See Remark 1 in Appendix C. 
grated model.

Proof. See Appendix C.  4. Computational evaluation

Proposition 4. The optimal solution of the split-among-stations In this section, we describe the parameters and results of the
model is NP-hard. computational evaluation. We first describe the open-source simu-
lation framework used for this paper in Section 4.1. Next, we show
Proof. See Remark 1 in Appendix C. 
the results of the simulation in Section 4.2. In Section 4.3, we make
some remarks regarding the assumptions that were made in our
3.4. Split-over-time model computational evaluation from a practical point of view.

This model is an extension of the split-among-stations model

from the previous section. Here we also allow every order to be 4.1. Simulation framework
split over different time periods. This means some SKUs for an or-
der may be assigned in one period while the others will stay in the In the following evaluation we use RAWSim-O from
backlog to be assigned in later periods. We define one additional Merschformann et al. (2018), an open-source, agent-based and
binary variable. event-driven simulation framework providing a detailed view
Additional decision variable: of an RMFS. The source code is available at www.rawsim-o.de.
Fig. 5 shows an overview of the simulation process, which is

1, SKU i ∈ Io , o ∈ O is moved back to the backlog managed by the core simulator instance. The tasks of the simulator
ybio
0, else include:

 
• updating agents, which can resemble real entities, such as
Min x ps + Wu · us (1) robots and stations
p∈P s∈S s∈S • passing decisions to optimizers, which can either decide imme-
s.t. (2.1 ) − (3.1 ), (4 ) − (10 ), (12 ) − (20 ) diately or buffer multiple requests and release the decision later
 • exposing information to a visualizer, which allows optional vi-
yios + ybio = yo, ∀i ∈ Io , o ∈ O (11.1) sual feedback in 2D or 3D.
s∈S
The hierarchy of decision problems regarding the assignment
ybio ∈ {0, 1}, ∀i ∈ I o , o ∈ O (15)
of replenishment orders, pick orders and pods to station is illus-
All constraints from the previous model are carried over, ex- trated in Fig. 6. We develop a new optimizer, called POA & PPS,
cept for constraint set (11). Constraint set (11.1) is relaxed to allow to make integrated decisions for both POA and PPS. Furthermore,
 L. Xie, N. Thieme and R. Krenzler et al. / European Journal of Operational Research 288 (2021) 80–97 89

Requests:

Insert
Repl. order
received ROA RSP
Trips:

Extract
Pick order TA PP
received POA & PPS

PR

Store
Fig. 6. Order of decisions to be made after receiving a pick or replenishment order.

we extend the new optimizer to support split orders. And we de-

fine new type of capacity for each station, called item capacity, in
RAWSim-O. If a replenishment order needs to be assigned to a re-
plenishment station, the optimizers of ROA and RPS are responsi-
ble for choosing a replenishment station and a pod. This results in
an insertion request, i.e. a request for a robot to bring the selected
pod to the given workstation. If a picking order needs to be as-
signed to a picking station, the new optimizer POA & PPS submits
all necessary information in the simulator to our model and con-
verts its solution into extraction requests in the simulator. Extrac-
tion requests contain both, the item that needs to be picked and
the pod that it should be picked from. The model is in this case ei-
ther the integrated model in Section 3.2, the split-among-stations
model in Section 3.3 or the split-over-time model in Section 3.4.
Our models were implemented in the python API of Gurobi Opti-
mizer. Note that in Merschformann et al. (2018), first the POA op-
timizer is called and the extraction requests are generated; after
that, the PPS optimizer is called. In other words, pods are only as-
signed to stations after orders have already been assigned. More- Fig. 7. Probabilities of order lengths for |I| = 20.
over, in our new optimizer, extraction requests for an order can
be assigned to different stations to support split orders. We define
the item capacity as the sum of the extract requests at one station
for each (part of an) order that has at least one unfulfilled request
at that station. Furthermore, the system generates a store request
each time a pod needs to be transported back from a station to a
storage location, and the PR optimizer decides on the storage loca-
tion for that pod. The TA optimizer assigns robots to these tasks.
All tasks result in robot trips, which are planned by a PP optimizer.
The simulation framework conceptually consists of three differ-
ent inputs:

• instance configuration: contains orders, initial inventory and

available SKUs (see Section 4.1.1)
• layout configuration: determines the characteristics and dimen-
sions of the warehouse layout (see Section 4.1.2)
• optimizer configuration: specifies the decision rules for all op-
erational problems in an RMFS (see Section 4.1.3).

We describe these three inputs for our experiments in this pa-

per in the following sections. Fig. 8. Probabilities of SKUs for |I| = 20.

4.1.1. Instance generation

First, we describe how we generate SKUs and orders, and how Onal, Zhang, and Das (2017). The SKUs in the order are then cho-
to fill the pods. The set of SKUs is generated as I = {i1 , ..., i|I| }. For sen by sampling without replacement from I using the probabil-
each order o1 , ..., o|O| the number of different SKUs in it is deter- ity mass function of a geometric distribution with p = 5/|I|, to ac-
mined by a truncated (1 to |I|) geometric distribution with p = 0.4 count for the varying demand for different SKUs. Fig. 8 shows the
(see Fig. 7 for |I| = 20). And the number of items for a SKU is set probabilities of SKUs for |I| = 20.
to 1. It is typical in online retailing that most of the orders contain A shared storage strategy (as described in Bartholdi & Hackman,
very few line items, such as in Bozer and Aldarondo (2018) and 2017) is applied to fill the pods. To determine the initial inventory
90 L. Xie, N. Thieme and R. Krenzler et al. / European Journal of Operational Research 288 (2021) 80–97

Fig. 9. Simulation layout.

Table 5 Table 6
Instance parameters. Decision rules.

Symbol Description Values Decision Problem Sequential Integrated (without/with split orders)

|O| Number of orders 50, 150, 250 POA Pod-Match Model

|I| Number of SKUs 20, 100 PPS Demand Model
|P| Number of pods 50, 100 ROA not relevant not relevant
α SKUs per pod 2, 3 RPS not relevant not relevant
PR Nearest Nearest
PP WHCA∗n WHCA∗n

of the pods, lists of all SKUs in randomized order are concatenated

until the combined list includes at least |P| · α elements. Then, for
binations achieved the best throughput. The rule Pod-Match for the
each pod p1 , ..., p|P| , the inventory is determined by cutting off the
POA problem selects the pick order from the backlog that best
first α items in the list.
matches the pods already assigned to the station. Out of all pods
We generate instances with different parameters from Table 5.
that could fulfill at least part of an assigned order, the decision
By considering all possible combinations of the parameters, we
rule Demand for PPS chooses the pod whose inventory is in most
generate 24 instances. We test all methods in this paper with the
demand when combining all unfulfilled orders. The ROA and RPS
pregenerated instances to see the efficiency of the different algo-
problems are concerned with the replenishment process and thus
rithms. In a real RMFS, the orders would not be known at the
not relevant to our tests. The decision rule Nearest for PR assigns
start of the simulation but would instead be received during run-
the closest parking space to each pod leaving a picking station. For
time. Our methods are also compatible with this type of order-
the PP problem, non-volatile windowed hierarchical cooperative A∗
generation (see Section 4.3).
(called WHCA∗n ) from Merschformann, Xie, and Erdmann (2017) is
used. In this algorithm, A∗ is used to find the optimal path with-
4.1.2. Layout
out considering collisions for each robot, while a space-time reser-
The layout is identical for most of the test instances and is il-
vation table is used within a limited time window to avoid colli-
lustrated in Fig. 9: 428 pods (blue squares), 504 storage locations
sions between robots. Non-volatile means that the existing path
in 2 × 4 blocks, 4 picking stations (red squares) and 8 robots (gray
and reservation are stored for each robot within a limited time
circles). The length of station queues is 12. As we are focusing on
window. The model used for POA and PPS to test our integrated
the decisions made in the order picking process, we disregard the
approach is the integrated model described in Section 3 and both
replenishment process. When the number of pods is higher than
of its extensions for split orders.
the value of P, only |P| pods have an inventory and the empty
pods are not taken into consideration. The capacity of picking sta-
4.2. Simulation results
tions is similar to that used in Boysen et al. (2017). Boysen et al.
(2017) used an order capacity of 6 orders. Since we use item ca-
Each combination of method and instance is simulated ten
pacity instead of order capacity, we multiply 6 by 2.5, the mean of
times to lessen the effect of randomness caused by other optimiz-
the distribution used to determine the number of items per order
ers and simulation components. Testing was done on 12 Intel Xeon
(described in Section 4.1.1), which leads to an item capacity of 15.
X5650 Cores with 24 Gigabyte RAM. The following performance
metrics are tested:
4.1.3. Decision rules
Table 6 lists the decision rules for all operational problems in • the number of pod-station visits per order
the evaluation. The definitions of these can be found in Table 2. • the distance driven by robots per order
The POA and PPS decision rules selected for the sequential ap- • pile-on: the number of picks per pod-station visit
proach are based on Merschformann et al. (2019), since these com- • throughput: picks per hour.
 L. Xie, N. Thieme and R. Krenzler et al. / European Journal of Operational Research 288 (2021) 80–97 91

Fig. 10. Normalized pod-station visits per order.

Fig. 12. Normalized pile-on (best viewed in color).

Fig. 11. Normalized robot distances per order.

The parameter Wu is set to 2, since we aim to fully utilize Fig. 13. Picks per hour (best viewed in color).
the picking station’s capacity whenever possible. The highest pos-
sible cost of assigning an additional SKU is 1 (one additional pod
needed, just for this SKU), therefore any value of Wu greater than
1 ensures that as many orders as possible are assigned. Further pa- sequential approach, pods may have to be transported directly be-
rameters used in the simulation can be found in Appendix D, such tween stations, because the orders that share the same pods are
as parameters for robot movement. assigned to different stations. Therefore, there are pods moving be-
Fig. 10 shows the average number of pod-station visits per or- tween picking stations. In the integrated approach, we try to assign
der relative to the sequential approach for each instance set of orders which share the same pods to the same stations; therefore,
50, 150 and 250 orders. Compared to the sequential approach for it reduces the number of pods moving between picking stations.
the number of pod-station visits per order: the integrated model This explains why the distances per robot in the integrated ap-
improves this performance by 20% to 30% for different instance proach do not show as much of an improvement as pod-station
sets, the split-among-stations model (in short in figures and ta- visits per order compared to the sequential approach.
bles: split) shows improvements of about 50%, and the split-over- Fig. 12 shows the average pile-on of all methods for each in-
time model (in short in figures and tables: timesplit) improves on stance set of 50, 150 and 250 orders. Our approach with split or-
the sequential solution by 57% to 80% for different instance sets. ders causes more picks per pod-station visit (PSV), especially in
Fig. 11 shows similar improvements for the average distances the split-over-time model (up to 5.5 times as many picks per PSV)
driven by robots to complete an order for each instance set. The (best viewed in color).
correlation between pod-station visits and distances driven by In Fig. 13 we get the average picks per hour for each instance
robots confirms our assumption in Section 1, that distances driven set of 50, 150 and 250 orders. From 30% up to 46% more picks per
by robots can be reduced by minimizing the number of pod-station hours can be achieved by the split-among-stations and split-over-
visits. As the layout shown in Fig. 9 demonstrates, the distance be- time.
tween the inventory area and a picking station is in most cases Fig. 14 illustrates the total computing times for each instance
greater than the distance between any two picking stations. Both a set of 50, 150 and 250 orders. The total computing time for the in-
pod coming from the inventory area to a station and a pod com- tegrated approaches with and without split orders are much larger
ing from another picking station count as pod-station visits. In the than those for the sequential approach. We divide the total com-
92 L. Xie, N. Thieme and R. Krenzler et al. / European Journal of Operational Research 288 (2021) 80–97

Fig. 14. Computing time for the first period and others (best viewed in color).

Table 7
Computing time per simulation time for t > 1.

Instance set Sequential Integrated Split Timesplit

50 0.07% 0.17% 0.24% 0.61%

150 0.07% 1.4% 3.46% 6.41%
250 0.07% 3.39% 9.67% 14.67%

Table 8
Computing time t > 1 per order (in seconds).

Instance set Sequential Integrated Split Timesplit Fig. 15. Example of the heuristic method.

50 0.01 0.03 0.03 0.08

150 0.01 0.21 0.4 0.62
250 0.01 0.48 1.06 1.34 4.2.1. A Heuristic method for a real-world instance
The Canadian logistics company Think logistics (see
ThinkLogistics, 2012) owns an RMFS, which stores 20 0 0 SKUs
in 217 pods. And there are 15 robots used in the system. We
generated an instance of 20 0 0 orders and 25 SKUs per pod, and
puting time into the time at t = 1 and the time of the remaining tested it in the similar layout as in Fig. 9 (15 instead of 8 robots
periods t > 1. The period at t = 1 takes a lot of time, since no or- were used). SKUs and orders were generated as described in
ders or pods have previously been assigned to any stations and Section 4.1.1. The computational time of our integrated models
therefore more decisions are necessary than in the following pe- increased extremly for such instance. Therefore, we developed a
riods. As this usually happens at most once a day (when the sys- heuristic method to accelerate the computational time. The main
tem is stopped and then restarted), we can consider it a warm- idea is instead of considering all orders in the backlog as input for
up. It is more interesting to see the periods t > 1, since most deci- our models, we select only part of them (n ≤ |O|) in the following
sions are made in those periods. Note that the decisions at t = 1 way: First, all orders in the backlog are sorted according to the
in the split-over-time model are faster than those in the integrated percentage of orders lines in each respective order for which a
and split-among-stations models. The reason is that the split-over- pod containing the sku currently is in a stations queue (in short:
time model can easily find the best possible solution in the first match). Then the best n orders are selected. Fig. 15 shows an
timestep, where only one pod is needed at each station. Table 7 example to explain this heuristic method. We have two orders 1
shows the total computing time of the decisions in t > 1 in rela- and 2 in the backlog and we want to pick up n = 1 order. And
tion to the simulation time. We deem computing times acceptable pod 1 visits currently station 1. We have one open tote in station
as long as they are lower than the total simulation time. In our 1 and the match of order 1 to pod 1 is 100%, since all items of
tests, computing the decisions took at most around 15% of the sim- order 1 are stored in pod 1. The match of order 2 to pod 1 is 50%.
ulation time (split-over-time, 250-order instance set). The average So order 1 should be firstly picked from the backlog.
time for the assignment of one order is at most 1.5 seconds (see We test this method for different n in all instance sets (50-,
Table 8). 150-, 250-order) and model variants to analyze the effect n has on
 L. Xie, N. Thieme and R. Krenzler et al. / European Journal of Operational Research 288 (2021) 80–97 93

Fig. 16. Computational time vs. the number of pick-station visits for the 250-order instance set by prefiltering n orders (n is from 10 to 250) (best viewed in color).

computing time and the quality of the solution (number of pick- • Reduced size of queues: We set the size of the queue at each
station visits). The results of different n between 10 and 250 for picking station to be 12 (see Section 4.1.2). Based on our ob-
250-order instances in the split-among-stations model are illus- servation in the simulation, the queue is rarely full; one of the
trated in Fig. 16. The results of other instance sets show a simi- main reasons for this is that we minimize the number of pod-
lar distribution for the integrated approaches. By using the heuris- station visits. And allowing split orders reduces the number
tic method, we can save more than 55% in computing time while of pod-station visits by up to 80%. In general, fewer pods are
pod-station visits rise by 1% compared to the optimal results of brought to queues. So, allowing split orders might cause the
the split-among-stations model (see order count 50 in Fig. 16 and additional positive side-effect that the required space at pick-
order count 250 is the optimal solution). ing stations is reduced.
For testing the large instance, we chose n = 40 and solved • Capacity of packing stations: for the results shown above, we
with split-over-time method, since this method brings the best assumed the capacity of packing stations for split orders to be
performance, but it reqires more computational time. Compared to large enough to store all possible split orders. As this capacity
the sequential solution, we can reduce the number of pod-station might differ from one company to another, a situation where
visits by 60%, same is the reduced distances drived by robots. 47% not enough capacity is available to allow for the splitting of all
more picks per hours can be achieved, while the average compu- orders is conceivable. In order to consider this situation, we can
tational time of each order is 4.6 seconds, which is acceptable. We extend, for example, the models in Sections 3.3 and 3.4 with
simulated this instance ten times and the testing was done on Intel the following parameters, variables and constraints. We need
Xeon E5-2670 16 Cores with 24 GB RAM. two additional parameters: C for the total capacity of packing
stations (in other words: number of available split orders) and
N for the number of stations. The binary variable ylo is activated
4.3. Practical remarks
if order o ∈ O is split, while nl counts the number of currently
active split orders from the previous periods. Note that, this
Some assumptions in Section 3.1 might differ from real-world
number is decreased by one if one split order is picked com-
scenarios. In this section we discuss them from a practical point of
pletely.
view.

yos − eo = yo, ∀o ∈ O (3.2)
• Real-world orders: in this paper we use pregenerated instances
s∈S
to test the performance of different approaches. Instead, in a
real RMFS, new orders would constantly come in while the op- ylo ≥ eo/N, ∀o ∈ O (16)
timization algorithms are running. Even completely new SKUs
could be stocked during the optimization. To account for this,
our simulation RAWSim-O was recently extended by an inter-
ylo ≤ eo, ∀o ∈ O (17)
face to an ERP system to allow for its use as a robot control
system in real warehouses, as described in Xie, Li, and Thieme
nl + ylo ≤ C (18)
(2019). This new feature of the simulator could also be used in
conjunction with the content presented in this paper, to imple-
ment the model presented here in a real warehouse and use
ylo ∈ {0, 1}, ∀o ∈ O (19)
real instances instead of pregenerated testing instances.
• Considering orders with deadlines: in this paper we don’t gen-
erate orders with deadlines. The reason for that is we want to
eo ∈ Z ≥ 0, ∀o ∈ O (20)
see maximum effects of our methods. However, our heuristic
can be easily extended to select the orders with tight deadlines In constraint set (3.2), eo is counted as the number of addi-
first. tional stations to finish picking order o. Constraint sets (16) and
94 L. Xie, N. Thieme and R. Krenzler et al. / European Journal of Operational Research 288 (2021) 80–97

fill all customer orders. First, we introduced a new mathematical

model to integrate POA and PPS (in short: integrated). Second, we
extended the integrated model to allow for split orders. An order
is split when not all of its parts are assigned together to a sta-
tion. Two variations of split orders are considered: split-among-
stations (all order lines of a pick order have to be assigned in the
same time period, but may be assigned to different picking sta-
tions) and split-over-time (order lines of a pick order may be as-
signed in different time periods and to different picking stations).
In the instances we tested, the number of pod-station visits could
be reduced by up to 30% using the integrated model, 50% using
the split-among-stations model and up to 80% using the split-over-
time model compared to the state-of-the-art sequential approach
described in Merschformann et al. (2019).
Additionally, we analyzed the simulation results with regard to
three additional performance metrics, namely robot distances, pile-
on and picks per hour (throughput). According to our experiments,
a reduction in pod-station visits induces a reduction in robot dis-
tances and by definition it comes with higher pile-on. The picks
Fig. 17. An example of using separators in totes (LocusBot). per hour can be increased, due to the shorter waiting time of pick-
ers, caused by more efficient order assignments which require the
robots to drive less distance and therefore better supply the sta-
tions with pods to pick from.
(17) make sure that ylo is equal to one if eo ≥ 0, while constraint The running times for our test instances with 50, 150 and 250
set (18) makes sure that the number of split orders is less than orders were acceptable; however, we needed a heuristic method
C. Constraint set (19) defines ylo as binary variables for each or- to test a real-world instance with 20 0 0 orders. For instances of
der o ∈ O, while constraint set (20) ensures that eo can only that size, it is not practicable to consider all unfulfilled orders in
have non-negative integer values. the models at each period; moreover, not all orders are suitable in
To see the impact that a limited capacity of packing stations each period. Therefore, we selected 40 best orders according to our
has on the solution, we tested the 250-order instance set for heuristic in each period to submit to the model. Still, a 60% reduc-
a packing station capacity of 10. When applying this limited tion in pod-station visits and distances driven by robots could be
capacity to the split-among-stations model, pod-station visits achieved using the split-over-time model, and 47% more picks per
were reduced by 35% compared to the sequential approach, in- hours can be achieved with acceptable running times.
stead of nearly 50% when packing capacity is not limited. Sim- We additionally discussed some assumptions for the mathemat-
ilar effect can be achieved by the split-over-time model. But ical models from a practical point of view, such as the extension
the split-over-time model needs higher packing capacity, since of the split-among-stations model to support limited capacity of
some incomplete orders stay longer in packing stations. Note packing stations.
that the capacity limit of 10 is an extreme case and the solu- Since an RMFS is a new type of warehousing system, the con-
tion is still better than that of the sequential approach. In the cepts specific to RMFSs have not received much scholarly atten-
real world, higher capacities are possible. tion. For example, the implementation of priority zones (see Flipse,
• Item capacity: this capacity is used for comparing the sequen- 2011). These zones are located in the proximity of workstations to
tial and integrated approaches. The capacity of a picking station store pods which might be used in the near future. Determining
in the real world is limited by the number of totes, which de- the optimal size of priority zones would also be interesting.
pends on the size of a tote. There are the following possibilities
to apply our approaches without changing the layout of picking
stations: Acknowledgments
• using different sizes of totes, with smaller ones for split or-

ders The authors would like to thank two anonymous referees for
• using separators in totes for split orders (an example: Lo- their insightful comments and suggestions. Nils Thieme and Ruslan
cusBot in Fig. 17). Krenzler are funded by the industrial project “Robotic Mobile Ful-
• Reliability of the simulation: as described in Xie et al. (2019), fillment System”, which is financially supported by Ecopti GmbH
the simulation framework RAWSim-O was extended to connect (Paderborn, Germany) and Beijing Hanning Tech Co., Ltd. (Beijing,
to an ERP system and industry robots. So the optimizers in the China). We would like to thank the Paderborn Center for Parallel
simulation, including the newly developed ones in this paper, Computing for providing their clusters for our numerical experi-
can be applied directly in real-world scenarios. ments.

5. Conclusions Appendix A. Example of the split-over-time approach

In an RMFS, the decision on the assignment of orders to sta- We have in this example one open position for station 1,
tions (POA) affects the throughput of the whole system the most while we have two identical orders 1 and 2 in the backlog (in
(see Merschformann et al., 2019). Moreover, the decision on the as- Fig. A.18(a)). These two orders contain SKUs shown in blue and or-
signment of pods to stations (PPS) should be made together with ange. These two SKUs are located in two different pods, namely
POA to get better results (see Example 1 in Section 2.2). Therefore, pod 1 with the orange SKU and pod 2 with the blue SKU. By al-
we developed novel methods to solve both POA and PPS for multi- lowing the orders 1 and 2 to be split into blue and orange parts,
ple stations and make online decisions that minimize the number the orange ones can firstly be picked from pod 1 at station 1 in
of visits by pods to stations (in short: pod-station visits) to ful- period 1, while one position is open at station 2 (see Fig. A.18(b)).
 L. Xie, N. Thieme and R. Krenzler et al. / European Journal of Operational Research 288 (2021) 80–97 95

Fig. A1. An example of the split-over-time approach.

Proof of Proposition 1. By the assumption about Maximal order

size in Section 3.1, there exists an order o such that o has no more
items than capcity Cs at some station s ∈ S. Let us select this or-
der and the corresponding pods for our solution.
Formally this means: Set xps := 1 for all p ∈ Ps as required by
(6). Set yo := 1 and yo := 0 otherwise. Set yo s := 1 and set yos := 0
otherwise. Then (3) holds. Set yio s := 1 for all i ∈ Io and yios := 0
otherwise. Then (2) holds. By the choice of o and s constraint
(4) holds too. Finally, we set x ps := 1 for all p ∈ PiSKU with i ∈ Io ,
then (5) is fulfilled. Thus we have constructed a feasible solution.
To find order o and to set to set correspoding variables, we only
need polynomial time. 

Proof of Proposition 2. The main idea of the proof is that we can

Fig. B1. Positions of the boxes in size S (left) and M (right) within a tier of a shelf. select parameters for the integrated model with available stations
set S := {1, . . . , k} from Section 3.2 in such a way that Eqs. (1)–(10)
solve a given set covering problem. That means, given a set of sets
After that, the blue ones are picked from pod 2 at station 2 in pe- 
A := A1 , . . . , An , the universe U := ni=1 Ai , and a set B ⊂ U. Find a
riod 2 (see Fig. A.18(c)). This allows both orders to be fulfilled with
minimal index set J ⊂ {1, . . . , n} such that corresponding sets Ai
two visits of pods to the stations. 
cover the set B, or formally, B ⊂ nj∈J Ai .
Appendix B. Calculation example for the capacity of a packing We define an order o1 which has the same elements as the set
station B1 := B. We assume that the capacities of all the stations Cs , s ∈ S
are equal the size of the order o1 . Furthermore, we create disjoint
As mentioned before, the capacity of a packing station de- orders o2 , . . . , ok and define correspondsing disjoint sets B2 , . . . , Bk ,
pends on the number of parking shelves and their sizes. Ac- which do not share any elements from the universe U. All the or-
cording to Wulfraat (2012), the typical size of a shelf is ders form the set of current orders in the backlog O := {os |s ∈ S }.
99cm × 99cm × 244cm. And we have two common sizes of For every order oi , i ∈ {2, . . . , k} we also define additional set
boxes, namely 25cm × 17.5cm × 10cm and 37.5cm × 30cm × 13.5cm An+i−1 := Bi+1 . We define a set of pods P = {1, . . . , n, n + 1 . . . , n +
(see DHL Packsets in sizes S and M in https://www.dhl.de/ k} such that every pod p in P has the same elements as set Ap .
en/privatkunden/pakete-versenden/verpacken.html). As shown in Now we force the integrated model to assign all the orders
Fig. B.19(a) and B.19(b), we can store 12 and 6 boxes within a tier o1 , . . . , ok . In order to do it, we set the weight Wu := |B| + 1. Be-
of a shelf for the boxes in sizes S and M respectively. Then we as- cause every order o ∈ O needs maximaly |os | = |B| pods, it is
sume we store boxes in size S in half of the shelves, and boxes cheaper to use as many available capacities as possible. Because
in size M in the other half. By considering spaces for open boxes we have k orders, k stations, and the size of every order is the
and usable vertical spaces of a shelf, we assume 100centimetres of same as avaialble capacity of any station, the optimal solution will
vertical space is available for five tiers of boxes with sizes S, while assign all the orders and no capacity will be left. That means, for
another 100centimetres is available for three tiers of boxes with every optimal solution, the unused capacity is us = 0 for all s ∈ S.
size M. Based on these assumptions, we get to store 78 boxes in Therefore xps in (1) also solves
total on a shelf. In other words, we can store 78 split orders on a 
shelf. Min x ps . (C. 1)
p∈P s∈S
Appendix C. Omitted Proofs
Let xps be an optimal solution of the integrated problem. Be-
In this section we prove that the models in Section 3 always cause this solution minimizes (C. 1), it describes a minimal set of
have feasible solutions. pods to fulfill all the orders o1 , . . . , ok . Formally, we define an index
96 L. Xie, N. Thieme and R. Krenzler et al. / European Journal of Operational Research 288 (2021) 80–97

set References
 
J ∗ := p ∈ P : ∃s ∈ S : xps = 1 (C. 2) Armstrong, R. D., Cook, W. D., & Saipe, A. L. (1979). Optimal batching in a semi-au-
tomated order picking system. Journal of the Operational Research Society, 30(8),
then the index set J∗ is a mininal set with 711–720.
Azadeh, K., de Koster, M., & Roy, D. (2017). Robotized warehouse systems: Develop-

k  ments and research opportunities. Research paper (no. ERS-2017-009-LIS), ERIM
Bs ⊂ A p. (C. 3) Report Series Research in Management.
s=1 p∈J∗ Banker, S. (2016). Robots in the warehouse: It’s not just Amazon. http://www.forbes.
com.
Recall, B2 , . . . , Bk are the same as An+1 , . . . , An+k−1 with no ele- Bartholdi, J., & Hackman, S. (2017). Warehouse & distribution science: release 0.98.
ments from universe U. Thus, J := J ∗ \ {n + 1, . . . , n + k − 1} solves Supply Chain and Logistics Institute.
Boysen, N., Briskorn, D., & Emde, S. (2017). Parts-to-picker based order processing
our initial set covering problem in a rack-moving mobile robots environment. European Journal of Operational
  Research, 262(2), 550–562.
B = B1 ⊂ Ap = A p. (C. 4) Boysen, N., de Koster, R., & Weidinger, F. (2018). Warehousing in the e-commerce
p∈J ∗ \{n+1,...,n+k−1} p∈J era: A survey. European Journal of Operational Research.
Bozer, Y. A., & Aldarondo, F. J. (2018). A simulation-based comparison of two goods-
 to-person order picking systems in an online retail setting. International Journal
of Production Research, 56(11), 3838–3858. doi:10.1080/00207543.2018.1424364.
Remark 1. The proof of Proposition 2 also proves that the split- Cohen, L., Uras, T., & Koenig, S. (2015). Feasibility study: Using highways for bound-
among-stations model is NP-hard if xt p is the optimal solution for ed-suboptimal multi-agent path finding. In Proceedings of the eighth annual sym-
posium on combinatorial search.
(1), subject to constraints (2.1), (3.1), (4)-(10), and (11)-(14). This is Cohen, L., Wagner, G., Kumar, T. K. S., Choset, H., & Koenig, S. (2017). Rapid random-
because, after we found an optimal solution xps , we do not require ized restarts for multi-agent path finding solvers. arXiv preprint: 1706.02794
every order o ∈ O to be assigned to only one station to construct a De Koster, R., Le-Duc, T., & Roodbergen, K. J. (2007). Design and control of ware-
house order picking: A literature review. European Journal of Operational Re-
minimal index set J. search, 182(2), 481–501.
The proof of Proposition 2 also proves that the split-over-time De Koster, R. B., Le-Duc, T., & Zaerpour, N. (2012). Determining the number of zones
model is NP-hard, when we use the optimal solution of the split- in a pick-and-sort order picking system. International Journal of Production Re-
over-time model for xt p . The constant Wu forces any optimal so-
search, 50(3), 757–771.
Flipse, M. (2011). Altering and improving Kiva. Working Paper, Vrije Universiteit Am-
lution of the split-over-the time problem to use all available ca- sterdam.
pacities. Because we have capacity for all orders in the backlog, Hanson, R., Medbo, L., & Johansson, M. I. (2018). Performance characteristics of
robotic mobile fulfilment systems in order picking applications. IFAC-PapersOn-
no orders will be split in time. Thus, any optimal solution of the
Line, 51(11), 1493–1498.
split-over-the time model is simultaniously an optimal solution of Hoffman, A. E., Mountz, M. C., Barbehenn, M. T., Allard, J. R., Kimmel, M. E., San-
split-among-stations model, which is NP-hard. tini, F., Decker, M. H., D’Andrea, R., & Wurman, P. R. (2013). System and method
for inventory management using mobile drive units. https://www.google.com/
Proof of Proposition 3. We show that every solution of the inte- patents/US20130103552.
grated model also solves the split-among-stations. Let (xps ) p∈P,s∈S ,
Il-Choe, K., & Sharp, G. (1991). Small parts order picking: Design and oper-
ation. https://www2.isye.gatech.edu/∼mgoetsch/cali/Logistics%20Tutorial/order/
(yo )o∈O , (yos )o∈O,s∈S , and (yios )o∈O,io∈Io,s∈S be a solution of the in- article.htm.
tegrated model. From (2) follows (2.1). Set eo := 0 for all o ∈ Io, Krenzler, R., Xie, L., & Li, H. (2018). Deterministic Pod Repositioning Problem in
Robotic Mobile Fulfillment Systems. ArXiv e-prints,.
then from (3) follows (3.1), (12) and (20). Substitute (2) into (3),
Lamballais, T., Roy, D., & De Koster, M. (2017). Estimating performance in a robotic
then (11) follows. If we sum (3) on both sides of the equation over mobile fulfillment system. European Journal of Operational Research, 256(3),
 
i ∈ Io we get io ∈Io yios = io ∈Io yos ≥ yos and therefore (13) holds. 976–990.
From the first part of this proof and Proposition 1, it follows Lamballais, T., Roy, D., & De Koster, R. B. (2019). Inventory allocation in robotic mo-
bile fulfillment systems. IISE Transactions, 1–17.
that there is a feasible solution for the split-among-stations model. Merschformann, M. (2017). Active repositioning of storage units in robotic mobile
Finally, we have the same objective function (1). Because the fulfillment systems. In Proceedings of the operations research proceedings 2017
split-among-station problem is an optimization problem, its opti- (pp. 379–385). Springer.
Merschformann, M., Lamballais, T., de Koster, M., & Suhl, L. (2019). Decision rules for
mal solution is either robotic mobile fulfillment systems. Operations Research Perspectives, 6, 100128.
(xps ) p∈P,s∈S , (yo )o∈O , (yos )o∈O,s∈S , (yios )o∈O,io∈Io,s∈S or better.  Merschformann, M., Xie, L., & Erdmann, D. (2017). Multi-Agent Path Finding with
Kinematic Constraints for Robotic Mobile Fulfillment Systems. ArXiv e-prints,.
Proof of Proposition 5. The proof is analogous to the proof Merschformann, M., Xie, L., & Li, H. (2018). RAWSim-O: A simulation framework for
of Proposition 3. Let (xps ) p∈P,s∈S , (yo )o∈O , (yos )o∈O,s∈S , and robotic mobile fulfillment systems. Logistics Research, 11:8.
Onal, S., Zhang, J., & Das, S. (2017). Modelling and performance evaluation of explo-
(yios )o∈O,io∈Io,s∈S , (eo )o∈O , be a solution of the split-among-stations sive storage policies in internet fulfilment warehouses. International Journal of
model. Set all yi b := 0, then (11.1) holds. Furthermore, we have Production Research, 55(20), 5902–5915.
o Otten, S., Krenzler, R., Xie, L., Daduna, H., & Kruse, K. (2019). Lost-customers approx-
the same objective function (1). With the same argumentation as imation of semi-open queueing networks with backordering–An application to
in the proof of Proposition 3, we conclude all the statements of minimise the number of robots in robotic mobile fulfilment systems. ArXiv e-
Proposition 5.  prints,.
Roy, D., Nigam, S., de Koster, R., Adan, I., & Resing, J. (2019). Robot-storage zone
assignment strategies in mobile fulfillment systems. Transportation Research Part
Appendix D. Additional parameters in the simulation E: Logistics and Transportation Review, 122, 119–142.
ThinkLogistics (2012). Thinking robots – canadian logistics company deploys Kiva
robots in the warehouse. Materials Management & Distribution (Canada’s Sup-
ply Chain Magazine)https://www.thinklogistics.com/wp-content/uploads/2013/
Table D.9 02/MMD-ThinkLogistics.pdf.
Parameters of robot movement and time for picking units. Toister, J. (2017). An inside look at Amazon’s fulfillment cen-
ter operations. https://www.toistersolutions.com/blog/2017/4/10/
Parameter Value
an- inside- look- at- amazons- fulfillment- center- operations.
Robot acceleration/deceleration 1 sm2 Tompkins, J. A. (2010). Facilities planning (4th). Hoboken, NJ and Chichester: John
Robot maximum velocity 1.5 ms Wiley & Sons.
Time needed for a full turn of a robot 2.5s Van Gils, T., Ramaekers, K., Caris, A., & de Koster, R. B. (2018). Designing efficient
order picking systems by combining planning problems: State-of-the-art classi-
Time needed for lifting and storing a pod 2.2s
fication and review. European Journal of Operational Research, 267(1), 1–15.
Time needed for picking a unit 7s
Weidinger, F., Boysen, N., & Briskorn, D. (2018). Storage assignment with rack-mov-
Time needed for handling a unit at picking station 13s
ing mobile robots in kiva warehouses. Transportation Science, 52(6), 1479–1495.
Wulfraat, M. (2012). Is Kiva Systems a good fit for your distribution center? An
unbiased distribution consultant evaluation.http://www.mwpvl.com/html/kiva_
systems.html.
 L. Xie, N. Thieme and R. Krenzler et al. / European Journal of Operational Research 288 (2021) 80–97 97

Wurman, P. R., D’Andrea, R., & Mountz, M. (2008). Coordinating hundreds of cooper- Zhang, J., Yang, F., & Weng, X. (2019). A building-block-based genetic algorithm for
ative, autonomous vehicles in warehouses. AI Magazine, 29. doi:10.1609/aimag. solving the robots allocation problem in a robotic mobile fulfilment system.
v29i1.2082. Mathematical Problems in Engineering, 2019.
Xie, L., Li, H., & Thieme, N. (2019). From simulation to real-world robotic mobile Zou, B., Xu, X., De Koster, R., et al. (2018). Evaluating battery charging and swapping
fulfillment systems. Logistics Research, 12:9. strategies in a robotic mobile fulfillment system. European Journal of Operational
Yuan, R., Dong, T., & Li, J. (2016). Research on the collision-free path planning of Research, 267(2), 733–753.
multi-agvs system based on improved a∗ algorithm. American Journal of Opera-
tions Research, 06(06), 442–449. doi:10.4236/ajor.2016.66041.
Yuan, Z., & Gong, Y. Y. (2017). Bot-in-time delivery for robotic mobile fulfillment sys-
tems. IEEE Transactions on Engineering Management, 64(1), 83–93. doi:10.1109/
TEM.2016.2634540.