International Journal of Production Research, 2018

Vol. 56, No. 11, 3838–3858, https://doi.org/10.1080/00207543.2018.1424364

A simulation-based comparison of two goods-to-person order picking systems in an online

retail setting
Yavuz A. Bozer* and Francisco J. Aldarondo

Department of Industrial and Operations Engineering, University of Michigan, Ann Arbor, MI, USA
(Received 27 December 2016; accepted 5 December 2017)

Design and analysis of order picking systems continues to be an active topic of interest both in academia and practice,
especially in light of the significant increase in online retail sales. In this paper, we examine two types of well-known,
goods-to-person order picking systems, namely, a miniload system and a Kiva system. Using a simulation model, we
compare the performance of the two systems on the basis of expected throughput and expected container retrieval times
to process the same set of customer orders. We also discuss some of the advantages and limitations of the two systems.
Keywords: order picking methods; warehousing systems; E-commerce; simulation; end-of-aisle picking

1. Introduction
According to Pricewaterhouse-Coopers (2014), retail is the largest private employer in the United States (US); it directly
and indirectly supports 42 million jobs, and contributes $2.6 trillion annually to the US GDP. The US Census Bureau
(2016) reports that retail sales in the US continue to show steady growth, approaching $5 trillion in 2015. Similarly, the
percentage share of online retail sales has shown steady growth, reaching 7.26% (or $341 billion) in 2015. Equally
striking is the fact that there is ample room for growth in online retail sales as a percentage of overall retail sales.
Growth in the global retail market has been equally impressive. According to eMarketer (2014), retail sales worldwide
topped $22 trillion in 2014, with e-commerce sales accounting for $1.3 trillion, led by China and the US. The group
predicts that the ‘global retail market will see steady growth …, and in 2018, worldwide retail sales will … reach $28.3
trillion’ with the percent share of e-commerce increasing to 8.8% (or $2.5 trillion) by 2018.
The above growth in online retailing has fuelled not only the parcel delivery business but the order fulfilment busi-
ness as well, where thousands of orders are now picked in warehouses (or fulfilment centres) largely by human pickers.
As a result, order picking (OP) has become a focal point for many online retailers and academic researchers.
Nevertheless, OP is a labour-intensive activity and it accounts for 50%–55% of the operating cost in a warehouse
(Drury 1988; Frazelle 2002, 148, and Tompkins et al. 2010, 433). Searching for and extracting the items, including tra-
vel, may account for 80% of the time required to fill orders (Frazelle 2002, 154). OP is particularly challenging for
online retailers, which typically stock a large number of Stock Keeping Units (SKUs) and process many small orders as
measured by the number of line items/order. Using the right OP method and investing in the right equipment is a key
decision for many online retailers since it impacts their ability to make timely and accurate deliveries in a cost-effective
manner.
At the low-end of automation are walk-and-pick OP systems, which are classified as ‘person-to-goods’ (PTG) OP
since the pickers walk/travel to the pick face. At the higher end of automation are ‘goods-to-person’ (GTP) OP systems,
where a material handling system brings the containers to the pickers. (There are also fully automated OP systems with
no human pickers.) According to enVista (2017):

GTP systems have been around for decades, but recently, new technologies and software advancements have enabled signifi-
cant improvements in functionality, efficiency, inventory management, and space utilization. … (A)n increasing number of dis-
tribution and warehousing professionals are starting to believe that PTG technologies and order fulfillment methods are
somewhat outdated … Large warehouses, SKU proliferation and smaller order fulfillment windows are rendering PTG pro-
cesses less effective due to increased picking travel, mispicks, and increased labor requirements. … However, many companies
… are not been fully convinced that GTP integrated systems are reliable, and that they can effectively deliver acceptable ROI.

*Corresponding author. Email: [email protected]

© 2018 Informa UK Limited, trading as Taylor & Francis Group

 International Journal of Production Research 3839

In this study our objective is to compare the performance of two types of GTP OP systems; we are not concerned
with PTG vs. GTP. Nonetheless, increased picking rates and improved ergonomics are often stated as major advantages
of GTP OP systems. For a more comprehensive comparison, the reader may refer to enVista (2017).
GTP OP systems differ primarily by the type of material handling system used for bringing the containers to the
pickers. We focus on two specific types of such systems; namely, the Kiva system and the miniload automated storage/
retrieval (AS/R)-based OP system (or ‘miniload (OP) system’ for short). In the Kiva system, the SKUs are stored in
racks or ‘pods.’ Automated guided vehicles (AGVs) transport the racks to the pick stations. (Kiva Systems was acquired
by Amazon in 2012 and renamed ‘Amazon Robotics.’) One of the early industrial applications of Kiva involved a
600,000 sq.ft. Office Depot warehouse with 12,000 SKUs, 3500 pods and 430 AGVs, with separate pick and replenish-
ment stations (Wulfraat 2012). A more recent application of a Kiva-type of system is the one installed for Lekmer.com,
Scandinavia’s largest online toy retailer; it involves 1500 pods and 65 AGVs (Swisslog 2015).
In the miniload OP system, the SKUs are placed in containers (or trays) stored in racks. The containers are retrieved
by a storage/retrieval (S/R) machine (named ‘S/R’ for short) and delivered to the appropriate pick station via a conveyor
loop. In some miniload systems, one picker is dedicated to each aisle; no conveyor loop is provided. Also, some mini-
load systems contain multiple shuttles; see, for example, Guller and Hegmanns (2014) for a description of such systems.
Industrial applications of the miniload system can be found in VanDerLande (2017), where a four-aisle system with
38,720 locations is reported to achieve 600 lines/hour per picker and a six-aisle system with 13,000 locations is installed
in Netherlands for Tommy Hilfiger, a leading apparel and retail company.
In comparing the above two systems, we identify the design elements (i.e. the number of AGVs vs. the number of
miniload aisles) that yield the same throughput for picking the same set of orders. The primary tool used for the com-
parison is simulation. It is one of the few studies to directly compare two distinct OP systems. Whenever possible, the
same or similar parameters are used. The two systems are compared on the basis of throughput, picker and material han-
dling equipment utilisations, and order completion times as well as other relevant factors. While our primary purpose is
to compare the system configurations that yield the same throughput, we also present approximate cost measures.
In the next section we present a brief review of the OP literature, including analytical results and simulation studies
for miniload OP systems. A description of the two systems and the assumptions are presented in Section 3. Section 4
presents the numerical results, while Section 5 compares the two systems based on other/qualitative factors. The sum-
mary results and possible future research directions are presented in Section 6.

2. Literature review
OP has received considerable attention in the literature. For example, De Koster, Le-Duc, and Roodbergen (2007) cite
140 publications in OP, while Gu, Goetschalckx, and McGinnis (2007) cite 128 publications on warehousing, with 67
in OP. The selection of an OP method as a strategic decision is discussed in Gu, Goetschalckx, and McGinnis (2010)
and Dallari, Marchet, and Melacini (2009). The latter study, derived from the statistical analysis of 67 warehouses, pre-
sents an OP system selection methodology that uses two key parameters: the number of SKUs in the picking area and
the number of line items picked/day. The authors emphasise the need for design methodologies that require less detailed
data and allow shorter ‘time-to-design’ through analytic models. Another important aspect in the design stage is technol-
ogy selection. As evidenced by the 155 publications cited in Roodbergen and Vis (2009), among the OP technologies
available, AS/R systems have received significant attention.
Analytical results derived for GTP OP include Bozer and White (1990), who developed a two-server closed queue-
ing model for a miniload system with two pick positions and one picker/aisle. Subsequent work by Bozer and White
(1996) extended the above model to multiple pick positions/aisle and multiple aisles assigned to each picker. The above
models estimate the expected throughput of the system and the expected utilisation of the picker. Approximate through-
put bounds were developed by Foley and Frazelle (1991) and by Foley, Frazelle, and Park (2002), Foley, Hackman, and
Park (2004) for a miniload system with two pick positions/aisle.
A miniload system with a U-shaped conveyor was studied by Park, Frazelle, and White (1999), who proposed a
capacitated two-stage cyclic queueing model. The U-shaped conveyor provides additional buffer between the picker and
the S/R. Cycle times are derived and closed-form expressions for multiple performance measures are presented, assum-
ing exponentially distributed pick times. An extension to the case where one picker serves multiple aisles is presented
by Koh, Kwon, and Kim (2005). Both of the above models are used for determining the steady-state behaviour of the
U-shaped conveyor and for buffer sizing.
A miniload system with multiple pick stations on a conveyor loop is analysed by Claeys, Adan, and Boxma (2016).
The authors assume that new orders are always available to be released into the system so they can focus on the maxi-
mum throughput capability of the pick stations. The individual totes for an order are assumed to arrive randomly at each
3840 Y.A. Bozer and F.J. Aldarondo

pick station, and they do not necessarily arrive in the sequence in which the orders were released (sometimes called
‘out-of-sequence arrivals’). Each pick station is modelled with multiple, short input buffers since it is assumed that each
tote is within the reach of the picker. The authors assume that each pick station handles only one order at a time, mean-
ing the picker must wait for a tote of the current order being processed even if totes for other orders are present in the
buffer. Modelling each pick station as a single-server polling system (where the server represents the picker), the authors
derive bounds for the mean order flow time, which yields bounds on the maximum throughput. Out-of-sequence arrivals
are also considered by Andriansyah et al. (2014), who study an automated pick station that uses an integrated carousel
mechanism to deal with such arrivals. Unlike Claeys, Adan, and Boxma (2016), they assume that multiple orders can
be processed simultaneously.
Hwang, Moon, and Gen (2002) examine an integrated system where the containers retrieved by a miniload AS/RS
are delivered to an assembly line via an AGV system. The authors develop an analytic model to concurrently design the
miniload AS/RS and the unit load size of the AGVs. Component parts are assumed to be delivered in kits prepared at a
kitting station. We refer the reader to Andriansyah (2011) and Roodbergen and Vis (2009) for further reading on analyti-
cal results for miniload OP systems.
Several researchers have used simulation models to study miniload OP systems. Medeiros, Enscore, and Smith
(1986) developed a simulation model for two pick positions/aisle, while Pulat and Pulat (1989) modelled the case with
a U-shaped conveyor. Perry, Hoover, and Freeman (1984) modelled a miniload system with a conveyor loop. The
authors present a heuristic design approach aimed at meeting the required throughput by manipulating several physical
configuration variables and operating rules. They conclude that the conveyor loop is the bottleneck. Guller and
Hegmanns (2014) also simulate a miniload system with a conveyor loop; however, they focus on a multishuttle system.
Assuming Poisson arrival of orders, the authors study the expected order completion time and throughput as a function
of the order profile (ranging, on average, from 1.9 to 2.5 line items/order). They observe that the completion time for a
single-line order may be greater than that of a four-line order.
Raghunath, Perry, and Cullinane (1986) and Takakuwa (1996) developed modular simulation models that can repre-
sent multiple system configurations or operating policies. The latter paper covers the case where an AGV-loop serves as
the interface between the AS/R system and the pick stations. Assuming rail-guided vehicles (RGVs) for the interface
loop, Lee, de Souza, and Ong (1996) use simulation to estimate the system throughput as they vary the number of
RGVs. Orders are assumed to arrive randomly. The authors examine buffer sizing and RGV deadlocking as well.
Another modular simulation model for a miniload system with a conveyor loop and pick stations is presented by
Andriansyah et al. (2011), who use a process algebra-based language. The authors demonstrate the modularity of the
model by varying the control system logic and the number of miniload aisles. Their primary argument is that ‘model
architecture and control structure are crucial for simulation studies of industrial scale AS/RS.’ They assume that a tote
box is allowed to enter the conveyor loop only if it has ‘reserved’ space on the loop. (This implies that simply waiting
in the buffer without reserved space is not considered.) Also, each S/R machine is assumed to hold up to four tote
boxes.
Onal, Zhang, and Das (2017) also use simulation; however, they focus on PTG OP. They study an online retailing
warehouse with 400 SKUs and 3240 bins. The authors show the impact of the ‘explosive storage’ policy (where incom-
ing materials are broken-up and distributed to multiple storage points in the warehouse) and other factors such as bins
with comingled SKUs.
Despite numerous trade/commercial articles written about Kiva systems, there are relatively few refereed publications
on the subject. In an earlier paper by Wurman, D’Andrea, and Mountz (2008), in addition to content aimed at the artifi-
cial intelligence community, the authors discuss the challenges associated with operating large OP systems and the
drawbacks of carousels and AS/R-based systems. The authors, who were affiliated with Kiva Systems (the company),
point out a number of advantages of the Kiva system, which we return to in Section 5. Although the paper is informa-
tive, it generally paints the Kiva system in a favourable light and the limitations of the system are not addressed ade-
quately. No analytic or simulation results are presented. A similar paper on Kiva systems, highlighting their advantages,
is presented by Guizzo (2008).
An analytic model for Kiva systems, which is called ‘robotic mobile fulfilment’ (RMF) by some researchers, is pre-
sented by Lamballais, Roy, and De Koster (2017). Modelling the system as a queueing network, the authors present
analytic results to estimate the expected system throughput and the average AGV utilisation for single-line-item and
multi-line-item orders. They also show that the throughput capacity of the system is impacted by the location of the pick
stations but not by the shape of the forward area. Another study on RMF systems is presented by Zou et al. (2017).
Using an analytic model, the authors show that the performance of the system can be improved by assigning the pick
stations to the AGVs so that the expected sojourn time of the retrieval transactions is minimised. The improvement is
significant if the pick times at the stations have large differences.
 International Journal of Production Research 3841

A similar study for RMF systems is presented by Boysen, Briskorn, and Emde (2017), who focused on order pro-
cessing and rack (or pod) retrieval sequencing for a single-pick station. The authors consider the case where multiple
orders are picked concurrently at the pick station, and the objective is to minimise the pod retrievals, given that multiple
SKUs are located in each pod and the same SKU may appear in more than one of the orders being picked at the station.
They present the problem as a formal decision problem and propose solution procedures, which are shown to reduce the
number of AGVs required compared to simple decision rules that the authors claim are used in real-world warehouses.
(Since the details of Kiva systems are not published, it is not clear how much improvement would be obtained over the
retrieval sequencing logic used in Kiva systems.)
Another Kiva-based OP study is presented by Stowe (2016). Assuming that the downstream processes are not the
bottleneck, the study focuses on maximising the pick rate at the pick stations by segregating the SKUs by velocity (a
well-known technique known as slotting) and through policies that increase picker retention. The latter scheme is based
on the author’s observation that the number of person-hours worked by inexperienced pickers explains practically all of
the variability in the pick rates, which argues for picker retention. Combined, the above two strategies are estimated to
increase the throughput by 10%.
As the literature review indicates, a large majority of the refereed papers perform a detailed analysis of a specific
type of OP system, while papers that directly and quantitatively compare two competing OP systems are very rare in
the refereed literature. (By direct comparison we mean that the same set of orders and the same pick times are used.)
While detailed analyses of specific types OP systems will continue to enrich the OP literature, studies that compare two
systems to better understand their relative strengths/weaknesses are needed as well. Such studies are also valuable from
a practical standpoint since decision-makers must choose between competing systems when they instal a new OP
system.

3. System description
We next present the details of the two systems.

3.1 The Miniload system

The main components of the miniload system are described below (Figure 1).
Miniload aisles: Each aisle contains an aisle-captive S/R and single-deep storage racks. Each aisle is attached to the
loop by input/output (I/O) conveyors (see ‘miniload buffer’ in Figure 1). The I/O point for the S/R is located at the
lower front corner of the rack.
Pick stations: The pick stations are attached to the loop by I/O conveyors (see ‘pick station buffer’ in Figure 1).
Each (active) pick station is assigned one picker. All the line items in an order are picked by the same picker. Contain-
ers arrive at a pick station through its input buffer. Once the item(s) are picked, the containers depart through its output
buffer. No picker travel is necessary since the containers are automatically indexed, one at a time, towards the picker,
and then taken away when picking is completed. Two examples are depicted in Figure 2. In some systems, at each pick
station, additional conveyors attached to the loop deliver empty containers to the pick stations and/or transport the
picked items to a sortation/packing area. Such conveyors are beyond the scope of our study.
Conveyor loop: It transports the containers between the miniload aisles and the pick stations. The loop also provides
additional buffering – if a container finds a full I/O buffer at a miniload aisle or pick station, the container recirculates
on the loop.
Container flow: A retrieval request queue is maintained for each miniload aisle. When an order is opened
(Section 3.3.1), the appropriate retrieval request(s) are placed in the retrieval request queue of each S/R, which serves
the requests on a FCFS basis. The request is completed when the S/R deposits the container in the miniload output buf-
fer, where it waits to merge onto the loop. Once on the loop, the container is transported to the input buffer of the
appropriate pick station. After the item(s) are picked, the container is moved to the pick station’s output buffer where it
waits to merge onto the loop. Once on the loop, the container is returned to the same aisle via the miniload input buffer
and stored back in the rack by the S/R. The container flow is depicted in Figure 3.

3.2 The Kiva system

The Kiva system (Figure 4) employs AGVs. Its main components are described below.
Forward area and the pods: The forward area consists of single-wide aisles and pods. Cross-aisles are added to
minimise blocking and give the AGVs a clear path to/from the pick stations (Figures 4 and 5(a), (b)). (The shading in
3842 Y.A. Bozer and F.J. Aldarondo

Figure 1. The miniload system (miniload AS/RS with a conveyor loop supplying the pick stations).

(a) (b)

Figure 2. Two examples of a pick station. (Images are property of VanDerLande).

the forward area is explained later.) The pods are simple, multi-tier racks (Figure 5(b)) that are fairly short, which
allows safe transportation by the AGVs while giving the pickers access (Figure 5(c)).
 International Journal of Production Research 3843

Figure 3. Flowchart of container flow in the Miniload system.

AGVs: The pods are moved, one at a time, between the forward area and the pick stations by the AGVs (Figure 5(b)
and (c)), which navigate over a grid of two-dimensional barcodes (or QR Codes) embedded in the floor. The AGV spins
in place to lift/lower a pod.
Pick stations: Each (active) pick station holds one picker. All the line items in an order are picked by the same
picker. Once it is moved by an AGV to the appropriate pick station, the pod waits in the pick station buffer located
ahead of the picker (Figures 4 and 5(c)). The AGV stays with the pod while it waits in the buffer.
3844 Y.A. Bozer and F.J. Aldarondo

Figure 4. The Kiva system (the forward area, the AGVs, and the pick stations).

(a) (b) (c)

Figure 5. Examples of the Kiva system (a) The forward area, (b) Kiva AGV travelling with a pod in the forward area, (c) A pick
station (images are property of Kiva systems).

Container flow: A single retrieval queue is maintained for all the requests. When an order is opened (see Sec-
tion 3.3.1), the appropriate retrieval requests are placed in the retrieval request queue. The requests are served, one at a
time, on a FCFS basis by one of the AGVs. To serve a request, the AGV travels empty from its current location to the
pod. It then picks up the pod and travels to the appropriate pick station. Once the item(s) are picked, the AGV returns
the pod to the forward area and becomes available to serve another request. Two rules are considered to select an AGV
when serving a request; the closest available (CA) rule and the first available (FA) rule. With CA, when there’s a retrie-
val request, if multiple AGVs are available, the AGV that is closest to the pod (to be retrieved) is used. If no AGVs are
available, the first AGV to become available is used. With FA, the AGVs are numbered; when there’s a retrieval
request, the AGVs are scanned starting with AGV number one, and the first available AGV is used. If no AGVs are
available, the first AGV to become available is assigned to the request. The container flow is depicted in Figure 6.
 International Journal of Production Research 3845

Figure 6. Flowchart of container flow in the Kiva system.

3.3 Assumptions and parameters

We aim for as much a balanced comparison as possible, while keeping the two simulation models simple. Our intent is
a reasonably accurate but basic comparison of the two systems.

3.3.1 General assumptions for both systems

(1) SKUs: There are 6000 SKUs. The SKUs are equally partitioned into three size categories (small, medium, and
large). Each category experiences the same level of demand. We assume no stock-outs.
(2) Orders: A batch of 2400 customer orders is generated randomly for each replication. Both systems are simulated
with the same set of orders. Each order may contain up to 10 line items, with an average of 2.325 line items/order
3846 Y.A. Bozer and F.J. Aldarondo

Table 1. Distribution of the number of line items/order.

Number of line items 1 2 3 4 5 6 7–10

Probability 0.45 0.25 0.10 0.07 0.06 0.04 0.03

(Table 1). A majority of the orders contain very few line items, which is typical in online retailing; see, for example,
Lasgaa (2010), Napolitano (2013), and Onal, Zhang, and Das (2017). The number of pieces/line item is beyond the
scope of our study. The pick time implicitly assumes that some line items may require multiple pieces.
To generate an order, we first sample the number of line items from Table 1. For each line item, we randomly
choose an SKU size category (small, medium, or large). Given the size category, we then randomly sample an
SKU. Each size category, and each SKU within a category, are equally likely to be selected.
(3) Trays: The results are not limited to a specific type of container but to achieve sufficient detail, we assumed the
SKUs are stored in ‘trays,’ which is common for miniload systems. Three tray configurations are assumed: tray type
A holds four small-size SKUs, tray type B holds two medium-size SKUs and tray type C holds one large-size SKU.
Each tray measures 4 ft. (length) × 2 ft. (width) × 1 ft. (height).
(4) Picking: There are four pick stations. Each pick station may have multiple orders open simultaneously. Picks are
performed in the same sequence the containers are delivered at a pick station. The pick time/line item is uniformly
distributed between 10 and 30 s, which includes the transition time between consecutive trays/pods.
(5) Order release: To avoid flooding the system, we employ a pull-type mechanism like CONWIP (Spearman, Woo-
druff, and Hopp 1990), which limits the number of orders that are opened and assigned to a pick station at one time.
When an open order is completed (i.e. the last line item is picked), a new order is opened and assigned to the picker.
To open a new order, the unpicked orders are scanned, and the first eligible order is selected. (An order is eligible if
all the containers needed for that order are in the rack/forward area, and none of them are needed for orders that are
already open.) If too many (or too few) orders are opened, it leads to congestion and it impacts the availability of
the containers (or it results in picker idle time).

3.3.2 Miniload system assumptions

(1) Storage rack and trays: There are five aisles, with identical racks. Each rack is divided into three zones
(Figure 7). A tray is equally likely to be located in any opening within its zone. Each opening measures 1.25 ft.
(high) × 2.5 ft. (wide) × 4 ft. (deep). Since the number of bays and tiers are integers, the number of tray posi-
tions slightly exceeds the requirement. For example, the zone for small SKUs is seven tiers high and eight bays
long, which yields 56 positions/rack, or 56 × 2 × 5 = 560 positions across the five aisles. With 2000 small
SKUs, and four small SKUs/tray, there are actually 500 trays, which leaves space for future SKUs. (We also
simulated the system with four aisles. When the fifth aisle is removed, the rack dimensions in the remaining four
aisles are proportionally increased since the total number of trays is fixed.)
(2) S/R machines: There are five single-shuttle S/Rs (one in each aisle). Each S/R travels at a speed of 400 fpm
(vertical) and 800 fpm (horizontal), which represent approximately 68% of the top speed for state-of-the-art
miniload S/Rs (Swisslog 2016). No acceleration/deceleration effects are considered. The S/R container pickup/
deposit (P/D) times are equal to 3 s each. At the start of the simulation, each S/R is idle at its I/O point.
(3) S/R control logic: The trays are retrieved on a FCFS basis. In retrieving and later storing the trays, the S/R fol-
lows an ‘opportunistic interleaving’ policy (Bozer and Cho 2005); i.e. having completed a storage operation, if
there is a retrieval request in the queue, the S/R performs a retrieval operation. If the retrieval request queue is
empty, the storage request queue is checked; if it’s empty also, the S/R idles in the rack; if the storage request
queue is not empty, the S/R performs a storage operation. Likewise, having completed a retrieval operation, if
there is a storage request in the queue, the S/R performs a storage operation. If the storage queue is empty, the
retrieval request queue is checked; if it’s empty also, the S/R idles at the I/O point; if the retrieval request queue
is not empty, the S/R performs a retrieval operation.
(4) Conveyor loop: The conveyor travels at 200 fpm and has a maximum capacity of 54 trays. The trays are moved
on/off the loop by automated 90-degree transfer mechanisms, which are not simulated explicitly.
(5) Buffers: Each I/O buffer holds up to five trays.
 International Journal of Production Research 3847

Figure 7. Miniload rack with small, medium and large SKU zones (not to scale).

3.3.3 Kiva system assumptions

(1) Pods: Each pod consists of four shelves. Each shelf is equivalent to one miniload tray. A pod is assumed to hold
SKUs of only the same size category; i.e. a pod holds either 16 small-size SKUs (four SKUs/shelf ) or eight
medium-size SKUs (two SKUs/shelf ) or four large-size SKUs (one SKU/shelf).
(2) Forward area: Pods in the forward area are zoned by size (Figure 4). A pod is equally likely to be stored in
any position within its zone. To minimise congestion, cross-aisles are placed every five pods. There are 134
pods of small-size SKUs, 272 pods of medium-size SKUs and 518 pods of large-size SKUs in the forward area.
(Note that, as in the miniload system, we slightly exceed the minimum requirements. For example, we need
2000/4 = 500 trays for small SKUs but the pods provide 134 × 4 = 536 positions.)
(3) AGVs: 50 AGVs (with no downtime or battery charging) forms the baseline. Also, congestion is not modelled,
although it is likely. This assumption may favour the Kiva system but it is partially rectified by the cross-aisles.
Also, in travelling between two points, an AGV follows the shortest rectilinear path. (This assumption may
again favour the Kiva system since congestion-avoidance may force a vehicle to take a different route. However,
congestion-avoidance is beyond the scope of our study.) Each AGV travels at a speed of 260 fpm or 3 mph
(Wulfraat 2012). No acceleration/deceleration effects are considered. The pod P/D times are equal to 5 s each.
(The AGV spins in place to pick-up/deposit a pod.) At the start of the simulation, all the AGVs are idle and
located immediately outside the front of the forward area.
(4) AGV control logic: Each AGV follows a dual command (DC) cycle; i.e. when picking is completed, the AGV
travels from the pick station to the forward area to store the pod; it then travels to the next pod to be retrieved
(which is termed ‘travel between’ or TB), and subsequently returns to the pick station with the next pod.
(Matching an AGV with the next pod, i.e. FA vs. CA, was described in Section 3.2.) If there are no requests in
the retrieval queue, the AGV idles at the last location where it stored a pod.

4. Simulation model and results

A simulation model was developed and validated for each system with ProModel, where model development starts with
adding elements to the physical layout and then identifying the entity that flows through the layout. For the miniload
system, the elements in the layout are: the retrieval queue, the miniload aisles, the miniload input and output buffers,
the conveyor loop, the pick station input and output buffers, and the pick stations. For the Kiva system, the elements in
the layout are: the retrieval queue, the forward area, the pick station input buffers and the pick stations. The AGVs are
modelled as a dynamic resource (since each AGV moves through the system but is not an entity/customer). The cus-
tomer orders were created outside of Promodel as an excel file.
An extensive number of runs were made to validate the model and ensure that it runs according to the assumptions
and descriptions stated in Section 3. These runs include but are not limited to: (1) Runs made with and without infinite-
capacity buffers at the miniload and the pick stations, (2) Runs made with and without infinite speed for the conveyor
loop and infinite speed for the S/R machines, (3) Comparison of the analytic value of the expected analytic S/R machine
cycle time (Bozer and White 1984) with the expected S/R machine cycle times obtained from the simulation model, (4)
Comparison of the expected and simulated travel time values for the Kiva system for different sizes of the forward area,
3848 Y.A. Bozer and F.J. Aldarondo

(5) Tracing the flow of individual trays (and pods) through each system to verify the container flow logic, and (6) Veri-
fying conservation of flow by checking the number of trays (or pods) entering and exiting each element in the layout.
Five sets of orders were generated to simulate five replications for each system. Each system was evaluated with
four pick stations, and five or eight open orders/pick station. The miniload system was evaluated with four and five
aisles, while the Kiva system was evaluated with 50 and 70 AGVs, using FA or CA dispatching.
For each system, we report the (1) Total elapsed time (in hours) to process all 2400 orders (the simulation is stopped
when the last tray/pod is returned to storage); (2) Throughput of the system (line items picked/hour); (3) Expected utili-
sation of the pickers and the S/Rs or the AGVs; (4) Expected time (in mins) to deliver a tray/pod to the pick station
(measured from the time an order is opened to the time individual trays/pods arrive at the pick station buffer); and (5)
Expected time (in mins) to complete an order (measured from the time an order is opened to the time the last line item
for that order is picked). We also measured the contents of the I/O buffers, the conveyor loop and the retrieval queue(s).
Since throughput is a key measure, we focused on those configurations that yield a high expected picker utilisation.
However, equipment utilisation is also relevant since the S/Rs and the AGVs are capital-intensive. The results for the
miniload system are shown in Tables 2 and 3. Providing a fifth miniload aisle does not appear to be justified; it
increases the throughput by at most 2%, while the expected S/R utilisation decreases from about 92 to 73%. As antici-
pated, the number of open orders/pick station is an important parameter. With five open orders/pick station, we obtain
satisfactory results; if it’s increased to eight, the throughput increases by only 1.7%, while the order completion time
increases by about 55% and the loop content increases by 66%. The tray delivery time increases as well. (If the number
of open orders/pick station is less than five, the system throughput decreases noticeably.)
All the buffers and the conveyor loop appear to have adequate capacity (Table 3). The pick station input buffers
reach capacity (five trays) but their average content is below five trays, provided we use five open orders/pick station.
With four miniload aisles and five open orders/pick station, the loop is, on average, only 38% full, and its maximum
content reaches 38 trays (70% full). That is, there is no indication that the conveyor loop is a bottleneck. Owing to

Table 2. Miniload system simulation results.

Open Total Throughput Tray Order

No. of orders/ elapsed (line S/R machine Picker delivery completion
ML aisles picker time (h) items/h) utilisation utilisation time (mins) time (mins)

4 5 Avg. 8.014 700.706 0.913 0.977 2.218 3.965

95% CI 7.943, 8.086 698.307, 703.105 0.906, 0.920 0.975, 0.980 2.195, 2.241 3.930, 4.000
8 Avg. 7.883 712.395 0.925 0.995 3.602 6.213
95% CI 7.804, 7.961 709.675, 715.115 0.921, 0.930 0.994, 0.996 3.568, 3.637 6.152, 6.275
5 5 Avg. 7.866 713.921 0.733 0.995 2.015 3.891
95% CI 7.790, 7.942 712.200, 715.642 0.731, 0.735 0.994, 0.996 1.987, 2.043 3.851, 3.931
8 Avg. 7.854 714.993 0.736 0.997 3.492 6.190
95% CI 7.779, 7.929 713.014, 716.971 0.732, 0.740 0.996, 0.997 3.446, 3.538 6.123, 6.257

Table 3. Miniload system simulation results for the buffers.

No. of Open PS input PS output ML input

ML orders/ No. of buffer buffer buffer ML output ML retrieval Conveyor loop
aisles picker containers content content content buffer content queue content content (util.)

4 5 Average 3.22 0.18 1.42 0.18 2.30 20.62 (0.38)

95% CI 3.17, 3.27 0.18, 0.18 1.36, 1.48 0.18, 0.18 2.26, 2.34 20.24, 20.99
Maximum 5 2 5 2 15 38
8 Average 4.35 0.24 1.64 0.29 2.91 35.20 (0.65)
95% CI 4.32, 4.38 0.24, 0.24 1.62, 1.66 0.28, 0.29 2.78, 3.04 34.72, 35.69
Maximum 5 4 5 5 13 54
5 5 Average 3.76 0.19 0.50 0.17 0.68 24.47 (0.45)
95% CI 3.74, 3.78 0.19, 0.20 0.49, 0.52 0.16, 0.17 0.66, 0.70 24.23, 24.71
Maximum 5 2 5 3 15 48
8 Average 4.50 0.30 0.52 0.40 0.69 40.69 (0.75)
95% CI 4.49, 4.51 0.29, 0.31 0.51, 0.53 0.39, 0.42 0.66, 0.72 40.66, 40.72
Maximum 5 5 5 5 13 54
 International Journal of Production Research 3849

technological advances, the conveyor in our study is faster than the conveyor used by Perry, Hoover, and Freeman
(1984). With eight open orders /pick station, however, the average loop occupancy increases to 65%, and its maximum
content reaches its capacity (54 trays). The average content of the retrieval request queue (across the four miniload
aisles) is 2.30 and 2.91 requests, respectively, for five and eight open orders/pick station. For validation purposes, we
also simulated the case with infinite buffer capacities, which means no recirculation occurs on the loop due to blocking.
The expected travel time on the loop from the simulation model is equal to 36.52 s. Assuming each miniload aisle and
each pick station is equally likely to be used, the analytic expected travel time on the loop is equal to 37.50 s.
The results suggest that balancing the workload between the pickers and the S/Rs is important. In our particular
case, balance is achieved when the number of miniload aisles is equal to the number of pickers. In general, the appro-
priate ratio depends on the cycle time of the S/Rs vs. the pick times. (The picker may pick multiple line items from the
same tray but given the total number of SKUs, the small order sizes, and the number of SKU’s/tray, that is a small
probability.)
With four aisles, the average simulated S/R cycle time is 10.27 s and 17.97 s, respectively, for a single command
(SC) and dual command (DC) trip, including the P/D times. Using Bozer and White (1984), and assuming randomised
storage over the entire rack, the analytic expected values of SC and DC are equal to 10.33 s and 19.08 s, respectively.
The simulation values are slightly smaller since there’s a higher probability of visiting the zones with small and medium
SKUs. Recall that all the SKUs have equal velocity but the number of SKUs per tray depends on the SKU size. As
explained in Bozer and Cho (2005), a SC trip is I/O − (storage or retrieval) − I/O, while a DC trip is I/O − (store old
tray) − (retrieve new tray) − I/O. Given ‘opportunistic interleaving’ (see Section 3.3.2), about 42% of the S/R trips are
SC, and 58% are DC, yielding an average S/R cycle time of 9.523 s/tray. That is,
EðS/R cycle time/trayÞ ¼ ð% SC tripsÞ EðSC cycle timeÞ þ ð%DC tripsÞEðDC cycle time=2Þ (1)

EðS/R cycle time/trayÞ ¼ ð0:42Þð10:27Þ þ ð0:58Þð17:97=2Þ ¼ 9:523 s/tray (2)

Note that the DC cycle time is divided by 2 because two trays are handled on each trip. Since each tray is moved twice
per pick, the E(S/R workload/pick) ¼ 2ð9:523Þ ¼ 19:05 s/pick and the total expected workload per S/R, say, WSR, is
given by:
EðS=R workload=pickÞðN Þ ð19:05ÞðN Þ
WSR ¼ ¼ ¼ 4:7625ð N Þ; (3)
nSR 4
where N = number of picks (or line items) in the batch and nSR = number of S/R machines.
Likewise, let WP be the total expected workload per picker. That is,
Eðpick time=line itemÞðN Þ ð20ÞðN Þ
WP ¼ ¼ ¼ 5:00ð N Þ; (4)
nP 4
where nP = number of pickers. A balanced workload implies that WRATIO = WSR/Wp = 1, which holds true for our case
since 4.7625/5 ≅ 1.0.
The results for the Kiva system are shown in Tables 4 and 5. Providing more than 50 AGVs does not appear to be
justified. With five open orders/pick station, if the number of AGVs is increased from 50 to 70, the expected AGV utili-
sation decreases from 0.953 to 0.739, while the expected picker utilisation improves by only 0.3%, which is not statisti-
cally significant at 5%. Also, with 50 AGVs, increasing the number of open orders/pick station from five to eight
slightly decreases the expected picker utilisation.
A closer examination reveals that this is due to the manner in which the AGVs are initially allocated to the pick sta-
tions. When an order is opened by the first pick station, all the AGVs needed for that order are dispatched. With eight
open orders, and 2.325 line items/order, the first pick station seizes the first 18 AGVs or so. (The number may be
slightly less than 18 since, with a very small probability, some pods may contain two or more picks). Hence, the first
three pick stations combined seize approximately all 50 AGVs, which forces the fourth station to wait. (With five open
orders/pick station, we would dispatch 5 × 2.325 × 4 ≅ 46 AGVs.) When 70 AGVs are used, as expected, 5 vs. 8 open
orders/pick station has no significant impact on the expected picker utilisation. An alternative AGV allocation scheme
could have avoided penalising the last pick station, but regardless of the allocation scheme used, if too many orders are
opened at once, it would create an AGV shortage.
The expected DC travel time is equal to approximately 80 s for the FA rule. Assuming randomised storage over the
forward area, it is straightforward to obtain the expected travel time analytically, which is equal to 84.07 s. As before,
the simulation value is slightly smaller since there’s a higher probability of visiting the zones with small and medium
3850 Y.A. Bozer and F.J. Aldarondo

Table 4. Kiva system simulation results.

No. Open Total Tray Order

of orders/ elapsed Throughput AGV Picker delivery completion
Dis-patch AGVs picker time (h) (line items/h) utilisation utilisation time (mins) time (mins)

CA 50 5 Avg. 7.868 713.721 0.955 0.988 1.091 3.877

95% CI 7.789, 711.015, 0.952, 0.958 0.987, 0.990 1.077, 3.840,
7.947 716.427 1.106 3.914
8 Avg. 7.951 706.255 0.989 0.978 3.251 6.249
95% CI 7.878, 703.312, 0.986, 0.992 0.975, 0.981 3.204, 6.188,
8.025 709.198 3.297 6.310
70 5 Avg. 7.853 715.096 0.744 0.990 0.736 3.870
95% CI 7.779, 712.165, 0.740, 0.748 0.987, 0.993 0.728, 3.835,
7.927 718.026 0.744 3.905
8 Avg. 7.868 713.697 0.976 0.988 1.623 6.164
95% CI 7.808, 713.170, 0.973, 0.979 0.986, 0.991 1.590, 6.106,
7.929 714.224 1.655 6.223
FA 50 5 Avg. 7.878 712.826 0.953 0.987 1.279 3.893
95% CI 7.803, 710.993, 0.949, 0.957 0.984, 0.990 1.257, 3.855,
7.953 714.659 1.301 3.930
8 Avg. 8.038 698.624 0.988 0.967 3.457 6.301
95% CI 7.977, 696.474, 0.985, 0.991 0.965, 0.970 3.397, 6.232,
8.099 700.775 3.516 6.371
70 5 Avg. 7.851 715.243 0.739 0.990 0.929 3.872
95% CI 7.777, 712.944, 0.735, 0.743 0.988, 0.992 0.925, 3.837,
7.926 717.543 0.932 3.907
8 Avg. 7.856 714.872 0.978 0.990 1.794 6.165
95% CI 7.773, 711.698, 0.977, 0.980 0.987, 0.992 1.754, 6.105,
7.938 718.047 1.835 6.225

Table 5. Kiva system simulation results for the buffers.

Dis-patch No. of AGVs Open orders/picker No. of pods PS input buffer content Pod retrieval queue content

CA 50 5 Average 7.10 3.90

95% CI 7.05, 7.15 3.764, 4.04
Maximum 24 32
8 Average 7.59 28.87
95% CI 7.56, 7.61 28.38, 29.37
Maximum 25 71
70 5 Average 8.13 0.06
95% CI 8.05, 8.21 0.03, 0.10
Maximum 30 14
8 Average 12.26 10.02
95% CI 12.20, 12.31 9.62, 10.41
Maximum 35 44
FA 50 5 Average 6.45 3.64
95% CI 6.39, 6.51 3.40, 3.88
Maximum 23 36
8 Average 6.99 28.43
95% CI 6.96, 7.02 27.78, 29.09
Maximum 29 67
70 5 Average 7.45 0.04
95% CI 7.38, 7.53 0.02, 0.05
Maximum 27 10
8 Average 11.64 9.52
95% CI 11.61, 11.68 9.07, 9.97
Maximum 32 44
 International Journal of Production Research 3851

SKUs. (The expected DC travel time with CA is equal to 67 s/trip. The CA rule reduces the expected TB time from
27.60 s to 14.85 s.)
Since each pod is handled twice, adding a P/D time of 10 s, we obtain a total cycle time of 90 s (FA) and 77 s
(CA) per trip. However, this reduction has no noticeable impact on the expected picker utilisation because it is already
quite high under FA, and the time gained by the AGV is essentially lost waiting at the pick station buffer. For example,
as shown in Table 5, with 50 AGVs and five open orders/pick station, when we switch from FA to CA, there is no sig-
nificant increase in the expected picker utilisation, while the average content of the pick station buffer increases from
6.45 to 7.10 AGVs. With 70 AGVs, the expected picker utilisation remains at 0.990, while the average buffer content
increases from 7.45 to 8.13 AGVs.
We also consider balancing the workload in the Kiva system. With 50 AGVs, five open orders/pick station, and the
FA rule, the average queue of 6.45 AGVs at each pick station corresponds to approximately 129 s of expected waiting
time in the pick station buffer. Hence, assuming one pick per pod, we have,
EðAGV workload/pickÞ ¼ EðDC cycle time/podÞ þ Eðwait timeÞ ¼ 90 þ 129 s/pick: (5)
(Interestingly, the expected time an AGV waits in the pick station buffer exceeds the expected DC cycle time/pod.)
Hence, the total expected workload per AGV, say, WAGV, is given by:

EðAGV workload=pickÞðN Þ ð90 þ 129ÞðN Þ

WAGV ¼ ¼ ¼ 4:38ð N Þ; (6)
nAGV 50
where N = number of picks (or line items) in the batch and nAGV = number of AGVs. As before, Wp = 5.00(N). Thus,
WRATIO = WAGV/Wp = 4.38/5 = 0.876, which is close to 1.0. However, the expected wait time at the pick station (which
appears in the numerator) depends on the number of AGVs, with fewer AGVs generally resulting in a shorter queue.
Therefore, the relationship between nAGV and WAGV is more complicated than it appears, which suggests that the notion
of a balanced workload for the Kiva system needs further investigation.
We next compare the two systems. Although the expected picker utilisation is slightly better for the Kiva system
(0.977 vs. 0.987, which is statistically significant at 5%), the expected throughput of the Kiva system with 50 AGVs,
FA dispatching, and five open orders/pick station is essentially equal to the expected throughput of a four-aisle miniload
system with five open orders/pick station. That is, one miniload aisle yields about the same expected throughput as 12–
13 AGVs.
One may also compare the order completion times and the container (tray or pod) delivery times. (The delivery time
is measured from the instant an order is opened to the instant the tray/pod enters the pick station buffer. Waiting time in
the buffer is not included.) Although the average order completion times are comparable (3.965 min for the miniload
system vs. 3.877 min for the Kiva system), the Kiva system provides considerably shorter average container delivery
times (2.218 vs. 1.279 min), which would be a desirable feature for high-priority orders; however, it does not yield a
higher expected throughput since both systems maintain high expected picker utilisations by queueing the containers in
front of each picker.
In the miniload system, each tray spends, on average, 1.186 min (out of 2.218 min) on the loop, including possible
recirculation. This value drops to 0.726 min if the pick station input buffer capacity is unlimited and no recirculation
takes place. The above result suggests that the miniload container delivery time can be reduced if the input buffer capac-
ity is increased but given that the loop is, on average, 38% full (with a maximum value of 70%), we do not see a com-
pelling reason to increase the buffer sizes. Further insight is gained for both systems by examining the percentage
breakdown of the container delivery times (see Table 6).
The two systems were further evaluated by increasing the number of SKUs from 6000 to first 12,000 and then
24,000. As before, the SKUs are equally partitioned into three size categories, and each category experiences the same

Table 6. Percentage breakdown of the average container delivery time.

Miniload System Kiva System

Waiting in the retrieval queue 36.10% Waiting in the retrieval queue 29.90%
Retrieved by the S/R machine 7.00% Retrieved by the AGV 70.10%
Waiting to enter the conveyor 2.90%
Travelling on the conveyor loop 54.00%
3852 Y.A. Bozer and F.J. Aldarondo

Table 7. Miniload system simulation results with additional SKUs.

Open Total Throughput S/R Tray Order

No. No. of orders/ elapsed (line machine Picker delivery completion
SKUs ML aisles picker time (h) items/h) utilisation utilisation time (mins) time (mins)

6000 4 5 Avg. 8.014 700.706 0.913 0.977 2.218 3.965

95% CI 7.943, 8.086 698.307, 703.105 0.906, 0.920 0.975, 0.980 2.195, 2.241 3.930, 4.000
8 Avg. 7.883 712.395 0.925 0.995 3.602 6.213
95% CI 7.804, 7.961 709.675, 715.115 0.921, 0.930 0.994, 0.996 3.568, 3.637 6.152, 6.275
5 5 Avg. 7.866 713.921 0.733 0.995 2.015 3.891
95% CI 7.790, 7.942 712.200, 715.642 0.731, 0.735 0.994, 0.996 1.987, 2.043 3.851, 3.931
8 Avg. 7.854 714.993 0.736 0.997 3.492 6.190
95% CI 7.779, 7.929 713.014, 716.971 0.732, 0.740 0.996, 0.997 3.446, 3.538 6.123, 6.257
12,000 4 5 Avg. 8.641 649.868 0.951 0.903 2.675 4.274
95% CI 8.574, 8.708 646.283, 653.453 0.947, 0.955 0.897, 0.910 2.662, 2.688 4.248, 4.300
8 Avg. 8.353 672.324 0.977 0.937 4.320 6.579
95% CI 8.261, 8.445 666.730, 677.918 0.970, 0.984 0.927, 0.947 4.287, 4.353 6.515, 6.643
5 5 Avg. 7.905 710.335 0.818 0.991 2.068 3.902
95% CI 7.849, 7.961 708.800, 711.870 0.816, 0.820 0.990, 0.993 2.046, 2.091 3.875, 3.929
8 Avg. 7.865 714.001 0.823 0.996 3.520 6.199
95% CI 7.805, 7.925 711.144, 716.858 0.821, 0.826 0.996, 0.997 3.500, 3.539 6.151, 6.247
24,000 5 5 Avg. 8.107 692.641 0.914 0.964 2.302 4.005
95% CI 8.045, 8.170 688.427, 696.856 0.911, 0.917 0.960, 0.969 2.292, 2.313 3.977, 4.034
8 Avg. 7.861 714.320 0.936 0.993 3.666 6.202
95% CI 7.798, 7.925 711.479, 717.161 0.934, 0.939 0.992, 0.995 3.637, 3.695 6.150, 6.253
6 5 Avg. 7.874 713.204 0.773 0.993 2.030 3.888
95% CI 7.812, 7.935 711.799, 714.610 0.770, 0.775 0.992, 0.994 2.023, 2.036 3.857, 3.919
8 Avg. 7.836 716.615 0.776 0.997 3.491 6.191
95% CI 7.768, 7.904 714.231, 718.999 0.774, 0.779 0.996, 0.997 3.455, 3.527 6.136, 6.246

level of demand. For the miniload system, the rack size is increased proportionally to accommodate the additional
SKUs. For the Kiva system, however, we increased the number of trays per pod from four to five trays, and then we
proportionally increased the size of the forward area. The results are shown in Table 7 for the miniload system. With
12,000 SKUs, four miniload aisles are sufficient to yield an expected picker utilisation of 90% with five open orders per
picker. However, five miniload aisles are required to obtain a near-100% picker utilisation. Likewise, with 24,000 SKUs,
a 96% picker utilisation is achieved with five miniload aisles but a sixth miniload aisle is needed to achieve a near-
100% picker utilisation. Hence, as the number of SKUs is increased, we need additional miniload aisles to maintain a
near-100% picker utilisation, although it’s interesting to note that increasing the number of open orders per station from
five to eight also yields a high picker utilisation without requiring an additional aisle.
For the Kiva system, increasing the number of trays/pod helps but the number of trays/pod is not doubled when the
number of SKUs is doubled, leading to an increase in the size of the forward area. However, the results for the Kiva
system (Table 8) show that additional AGVs are not required due to the larger forward area. Rather, the AGVs spend
more time travelling, which results in less time waiting in the input buffer of the pick station as shown in Table 9.
Nonetheless, the retrieval time per pod shows a slight increase as the number of SKUs is increased. The results depicted
in Figure 8 show the change in the system throughput as a function of the number of AGVs, the dispatching rule (FA
vs. CA), and the number of SKUs.
A related and interesting question is the methodology we used for sizing the two systems for comparison purposes.
We basically fixed the number of pick stations, and then adjusted in each system the number of miniload aisles and the
number of AGVs to reach an expected picker utilisation of nearly 100%. For the miniload system, this resulted in a sim-
ple search since the desired number of miniload aisles (four in our case) is quickly identified. For the Kiva system, we
experimented with the number of AGVs to determine the appropriate number (50 AGVs), which is the basis of the
result shown in Figure 8. More work is needed to develop a more systematic and perhaps more comprehensive method-
ology to compare OP systems.
In terms of cost, there are no detailed cost models published to methodically compute the cost of either system as a
function of the design elements. However, as a ballpark comparison, according to Wulfraat (2013), a miniload system
costs $500,000–$750,000 per aisle, plus the cost of the warehouse management/control system, and an additional
$100,000–$250,000 for ‘services.’ A warehouse with a Kiva system of 50–100 AGVs, on the other hand, is estimated
Table 8. Kiva system simulation results with additional SKUs.

Open
No. No. of orders/ Total elapsed Throughput AGV Picker Tray delivery Order PS input
SKUs AGVs Dis-patch picker time (h) (line items/h) utilisation utilisation time (mins) completion time (mins) buffer content (max)

6000 50 CA 5 Avg. 7.868 713.721 0.955 0.988 1.091 3.877 7.101 (24)
95% CI 7.789, 7.947 711.015, 716.427 0.952, 0.958 0.987, 0.990 1.077, 1.106 3.840, 3.914 7.053, 7.149
8 Avg. 7.951 706.255 0.989 0.978 3.251 6.249 7.587 (25)
95% CI 7.878, 8.025 703.312, 709.198 0.986, 0.992 0.975, 0.981 3.204, 3.297 6.188, 6.310 7.561, 7.613
FA 5 Avg. 7.878 712.826 0.953 0.987 1.279 3.893 6.452 (23)
95% CI 7.803, 7.953 710.993, 714.659 0.949, 0.957 0.984, 0.990 1.257, 1.301 3.855, 3.930 6.395, 6.510
8 Avg. 8.038 698.624 0.988 0.967 3.457 6.301 6.993 (29)
95% CI 7.977, 8.099 696.474, 700.775 0.985, 0.991 0.965, 0.970 3.397, 3.516 6.232, 6.371 6.965, 7.022
12,000 50 CA 5 Avg. 7.879 712.750 0.961 0.988 1.241 3.887 6.589 (23)
95% CI 7.814, 7.943 710.805, 714.695 0.959, 0.963 0.986, 0.990 1.221, 1.262 3.855, 3.919 6.557, 6.620
8 Avg. 8.069 695.917 0.987 0.966 3.418 6.283 7.037 (27)
95% CI 8.019, 8.119 691.492, 700.342 0.984, 0.990 0.961, 0.970 3.368, 3.467 6.226, 6.341 7.011, 7.062
FA 5 Avg. 7.939 707.336 0.953 0.980 1.467 3.907 5.699 (20)
95% CI 7.874, 8.003 706.436, 708.236 0.950, 0.956 0.979, 0.981 1.450, 1.484 3.876, 3.938 5.671, 5.727
8 Avg. 8.085 694.582 0.990 0.963 3.686 6.347 6.273 (25)
95% CI 8.014, 8.155 692.588, 696.576 0.988, 0.992 0.961, 0.965 3.647, 3.724 6.297, 6.396 6.243, 6.303
24,000 50 CA 5 Avg. 8.003 701.690 0.967 0.977 1.578 3.946 5.460 (21)
95% CI 7.935, 8.070 700.803, 702.577 0.966, 0.968 0.975, 0.978 1.565, 1.590 3.918, 3.974 5.437, 5.482
International Journal of Production Research

8 Avg. 8.182 686.376 0.988 0.955 3.857 6.406 5.884 (24)

95% CI 8.099, 8.264 682.910, 689.841 0.987, 0.989 0.951, 0.960 3.799, 3.916 6.337, 6.475 5.851, 5.916
FA 5 Avg. 8.162 688.038 0.962 0.958 1.921 4.024 4.420 (22)
95% CI 8.091, 8.232 686.152, 689.924 0.959, 0.964 0.955, 0.960 1.911, 1.932 3.994, 4.055 4.375, 4.465
8 Avg. 8.360 671.794 0.989 0.935 4.245 6.554 4.936 (30)
95% CI 8.246, 8.474 667.100, 676.487 0.987, 0.991 0.928, 0.941 4.188, 4.302 6.472, 6.636 4.890, 4.981
3853
3854 Y.A. Bozer and F.J. Aldarondo

to cost between $2 and $4 million. The cost of interfacing the system to the warehouse management system, or the cost
for ‘services,’ is not indicated. Other factors that impact the usability of both systems and potentially their long-term
performance/cost are presented in the next section.

5. Other/qualitative factors
Wurman, D’Andrea, and Mountz (2008) and Wulfraat (2012) point out a number of advantages of the Kiva system over
PTG OP systems such as improved picking rates and picker productivity, increased order accuracy. Many of the advan-
tages they state apply to other GTP OP systems as well. Also, some limitations of the Kiva system are not covered ade-
quately. In this section, we compare the miniload system and the Kiva system with respect to qualitative (or hard to
quantify) factors that we believe most decision-makers will take into account, beyond those results presented in the pre-
vious section. (Since both systems represent GTP OP, we will not state their advantages over PTG OP systems. For
such a comparison, the interested reader may refer to enVista [2017]).
Flexibility and scalability: Both systems offer flexibility in the number of pick stations that can be activated (or
staffed). However, both systems are limited by the capacity of the material handling system that retrieves/stores the con-
tainers. Although the conveyor loop can be designed to accommodate a future miniload aisle (Figure 1), the Kiva sys-
tem is more advantageous since it provides incremental scalability by letting the user add more AGVs and/or pods as
needed, with minimal interruption. (Too many AGVs, however, may cause congestion.) Incremental scalability may be
particularly attractive for an e-tailer on a growth path. Also, if growth forces relocation, the Kiva system is essentially
portable while the miniload system would be costly to relocate.
Parallel processing and built-in redundancy: The Kiva system provides substantial parallel processing since a pod
can be transported by any one of the AGVs, and at any given time, there could be a large number of pods being moved
to/from the pick stations. (This advantage is reflected in the faster container delivery times in the Kiva system as shown
in the previous section.) Multiple AGVs also provide redundancy against breakdowns and scheduled maintenance. If an
S/R breaks down, however, all the trays in that aisle become inaccessible. If the conveyor loop breaks down, all the
pick stations would be impacted, unless the trays are moved manually. In contrast, if an AGV breaks down, the impact
on the Kiva system would be minimal. (Of course, frequent breakdowns or extended downtimes in either system would
be unacceptable.)
Ergonomic factors: The Kiva system requires the picker to bend down or reach up to pick from the lower and upper
levels of a pod (Figure 9). In contrast, each tray in the miniload system is presented to the picker at an appropriate
height. Furthermore, the tray is tilted towards the picker, and the platform on which the picker stands can be adjusted to
the picker’s height, making the picking process ergonomically more operator-friendly. Also, with the miniload system,
the picker has an unobstructed view of the tray. To reduce picking errors, both systems can be equipped with pick-to-
light (or similar) technologies. An excellent paper comparing the ergonomics of a Kiva-based and AS/RS-based pick
station is presented by Lee, Chang, and Choe (2017), who used digital human modelling to assess the risk factors for
work-related musculoskeletal disorders for the pickers. The authors conclude that OP systems based on AS/RS pose
lower risk factors for human pickers, while the pick station in Kiva-based systems is in critical need of posture change.
For further information on ergonomics in OP systems, the reader may refer to Grosse et al. (2015), who stress that, in
the process of increasing the efficiency of OP systems, one must integrate human factors into the model in order to
enhance performance and reduce the long-term cost.
Footprint and cube utilisation: The Kiva system results in poor cube utilisation and a large footprint for the forward
area since the pods cannot be stacked. Building a mezzanine would alleviate this limitation but it would also require

Table 9. Breakdown of AGV time per pod retrieval (in seconds).

Open
No. of No. of orders/ Ave. storage Ave. empty Ave. retrieval Ave. time in PS Ave. total
Dispatch SKUs AGVs picker travel time travel time travel time input buffer time per pod

CA 6000 50 5 29.53 14.85 29.54 145.44 219.36

12,000 50 5 33.84 18.07 33.84 134.85 220.60
24,000 50 5 43.15 26.06 43.14 113.16 225.51
FA 6000 50 5 29.28 27.70 29.29 128.29 214.56
12,000 50 5 33.89 34.73 33.89 117.70 220.21
24,000 50 5 43.02 48.77 43.01 94.49 229.29
 International Journal of Production Research 3855

Figure 8. System throughput as a function of the number of AGVs.

(a) Picker reaching up to pick from top (b) Picker bending to pick from bottom
shelf of the pod (Photo: Quiet Logistics) shelf of the pod; note the step ladder
(Photo: Amazon Fulfillment)

Figure 9. Kiva system picker ergonomics.

elevators. This would increase the container delivery times and the cost of the system, especially if there is considerable
weight on the mezzanine. Also, the elevators may become a bottleneck. The miniload system, on the other hand, pro-
vides high cube utilisation since the storage rack can be as tall as the building allows. A tall structure may be especially
advantageous in areas with expensive/limited land. According to enVista (2017), ‘The real benefit of GTP systems … is
unparalleled vertical space utilisation. AS/RS systems can be constructed up to 100 feet high.’
Security of items: Both systems offer item security since access to the forward area is restricted. However, the mini-
load system offers additional security since the trays are stored in tall racks and each aisle is occupied by an S/R. Fur-
thermore, with the appropriate tray design, items remain secure in the tray. With the Kiva pods, on the other hand, since
3856 Y.A. Bozer and F.J. Aldarondo

the sides are open, items may fall out. Elastic bands can secure the items (see Figure 9(b)) but they may also interfere
with picking, forcing the picker to pull the band with one hand while picking with the other.
Battery charging: According to Wulfraat (2012), the AGVs

run on rechargeable lead-acid batteries that are charged at frequent intervals … such that there is no battery change-out process
required. The (AGVs) simply travel to designated charge stations every couple of hours where they receive a 5-min battery
recharge … For the purposes of budgeting, assume that at any time, 5% of the (AGVs) will be (unavailable) … When develop-
ing the business case for this technology, it is important to note that the batteries … require periodic replacement. For batteries
that undergo consistent daily usage, the typical battery life cycle is in the order of 1.5–2 years …

6. Conclusions and future research

Using simulation, we compare the performance of a miniload AS/R system and a Kiva system, in an online retail setting
where most orders have very few line items. Given the same set of orders, and the same number of pick stations and
picking parameters for both systems, we compare the configuration/size of each system that yields approximately the
same throughput in line items picked/hour. Given the assumptions and parameters of the study, our primary conclusion
is that a miniload system with four aisles and a conveyor loop yields approximately the same throughput as a Kiva sys-
tem with 50 AGVs. We also show numerically that, for both systems, a favourable balance between picker and equip-
ment (miniload or AGV) utilisation is achieved when the expected container retrieval-storage cycle time is
approximately equal to the expected pick time/container. The results also indicate that the number of open orders/pick
station is an important control parameter to ensure high picker utilisation, while avoiding congestion. Although we var-
ied the number of SKUs and the number of miniload aisles/number of AGVs, our conclusions are limited to the config-
urations/parameter values we tested. Additional studies are needed to compare the above two systems and other OP
systems under alternative parameter values. Last, we compare the two systems on other factors such as flexibility/scala-
bility, parallel processing/redundancy, ergonomics, footprint/cube utilisation and the security of items.
We used three categories of SKUs but assumed randomised storage within each category. Future studies may con-
sider slotting based on SKU velocities. Both systems would benefit from storing high-velocity SKUs closer to the pick
stations (although storing the fast-moving SKUs in the same area may create AGV congestion in the Kiva system).
Another direction to explore is the impact/timing of replenishment in both systems. To increase the cube utilisation, one
may also study the impact of installing a mezzanine and elevator(s) for the Kiva system. Another configuration to inves-
tigate would be the miniload system with the conveyor loop replaced by unit load AGVs to transport individual trays.
A majority of the studies in OP examine or optimise certain aspects of a specific type of system. While such studies
are very valuable, more studies that directly and quantitatively compare two or more types OP systems would enhance
the literature and at the same time provide additional insights for decision-makers and potential users. While the
throughput capacity of the system and capital cost/return-on-investment are key factors in selecting an OP system, other
factors (in particular flexibility/scalability and ergonomics at the pick stations) should also be taken into account.

Disclosure statement
No potential conflict of interest was reported by the authors.

References

Andriansyah, R. 2011. “Order-picking Workstations for Automated Warehouses.” Doctoral diss., Technische Universiteit Eindhoven.
Andriansyah, R., W. W. H. de Koning, R. M. E. Jordan, L. F. P. Etman, and J. E. Rooda. 2011. “A Process Algebra Based Simulation
Model of a Miniload-workstation Order Picking System.” Computers in Industry 62 (3): 292–300.
Andriansyah, R., L. F. P. Etman, I. J. B. F. Adan, and J. E. Rooda. 2014. “Design and Analysis of an Automated Order-picking
Workstation.” Journal of Simulation 8 (2): 151–163.
Boysen, N., D. Briskorn, and S. Emde. 2017. “Parts-to-picker Based Order Processing in a Rack-moving Mobile Robots Environ-
ment.” European Journal of Operational Research 262 (2): 550–562.
Bozer, Y. A., and M. Cho. 2005. “Throughput Performance of Automated Storage/Retrieval Systems Under Stochastic Demand.” IIE
Transactions 37 (4): 367–378.
Bozer, Y. A., and J. A. White. 1984. “Travel-time Models for Automated Storage/Retrieval Systems.” IIE Transactions 16 (4):
329–338.
 International Journal of Production Research 3857

Bozer, Y. A., and J. A. White. 1990. “Design and Performance Models for End-of-aisle Order Picking Systems.” Management Science
36 (7): 852–866.
Bozer, Y. A., and J. A. White. 1996. “A Generalized Design and Performance Analysis Model for End-of-aisle Order-picking Sys-
tems.” IIE Transactions 28 (4): 271–280.
Claeys, D., I. Adan, and O. Boxma. 2016. “Stochastic Bounds for Order Flow times in Parts-to-picker Warehouses with Remotely
Located Order-picking Workstations.” European Journal of Operational Research 254 (3): 895–906.
Dallari, F., G. Marchet, and M. Melacini. 2009. “Design of Order Picking System.” The International Journal of Advanced Manufac-
turing Technology 42 (1–2): 1–12.
De Koster, R., T. Le-Duc, and K. J. Roodbergen. 2007. “Design and Control of Warehouse Order Picking: A Literature Review.”
European Journal of Operational Research 182 (2): 481–501.
Drury, J. 1988. Towards More Efficient Order Picking, IMM Monograph 1. Cranfield: Institute of Materials Management.
enVista Corp. 2017. “Goods-to-Person or Person-to-Goods? 10 Factors to Consider When Determining the Right Order Fulfillment
Solution.” Accessed September 30, 2017. http://www.mhi.org/media/members/16455/131129730897732486.pdf
Foley, R. D., and E. H. Frazelle. 1991. “Analytical Results for Miniload Throughput and the Distribution of Dual Command Travel
Time.” IIE Transactions 23 (3): 273–281.
Foley, R. D., E. H. Frazelle, and B. C. Park. 2002. “Throughput Bounds for Miniload Automated Storage/Retrieval Systems.” IIE
Transactions 34 (10): 915–920.
Foley, R. D., S. Hackman, and B. C. Park. 2004. “Back-of-the-envelope Miniload Throughput Bounds and Approximations.” IIE
Transactions 36 (3): 279–285.
Frazelle, E. H. 2002. World-Class Warehousing and Material Handling. New York: McGraw Hill.
Grosse, E. H., C. H. Glock, M. Y. Jaber, and W. P. Neumann. 2015. “Incorporating Human Factors in Order Picking Planning Mod-
els: Framework and Research Opportunities.” International Journal of Production Research 53 (3): 695–717.
Gu, J., M. Goetschalckx, and L. F. McGinnis. 2007. “Research on Warehouse Operation: A Comprehensive Review.” European Jour-
nal of Operational Research 177 (1): 1–21.
Gu, J., M. Goetschalckx, and L. F. McGinnis. 2010. “Research on Warehouse Design and Performance Evaluation: A Comprehensive
Review.” European Journal of Operational Research 203 (3): 539–549.
Guizzo, E. 2008. “Three Engineers, Hundreds of Robots, One Warehouse.” IEEE Spectrum 45 (7): 26–34.
Guller, M., and T. Hegmanns. 2014. “Simulation-based Performance Analysis of a Miniload Multishuttle Order Picking System.” In
Proceedings of the 47th CIRP Conference on Manufacturing Systems, Procedia CIRP, vol. 17, Windsor, CA, 475–480.
Hwang, H., S. Moon, and M. Gen. 2002. “An Integrated Model for the Design of End-of-aisle Order Picking System and the Deter-
mination of Unit Load Sizes of AGVs.” Computers & Industrial Engineering 42 (2–4): 249–258.
Koh, S. G., H. M. Kwon, and Y. J. Kim. 2005. “An Analysis of the End-of-aisle Order Picking System: Multi-aisle Served by a Sin-
gle Order Picker.” International Journal of Production Economics 98 (2): 162–171.
Lamballais, T., D. Roy, and M. B. M. De Koster. 2017. “Estimating Performance in a Robotic Mobile Fulfillment System.” European
Journal of Operational Research 256 (3): 976–990.
Lasgaa, M., 2010. “How E-retailing Can Influence Your Logistics? A White Paper on E-fulfillment.” Groenewout, February 11.
https://www.groenewout.com/wp-content/uploads/2016/04/20111206162121_9024D198-WP-How-E-retailing-can-influence-your-
logistics.pdf
Lee, S. G., R. de Souza, E. K. Ong. 1996. “Simulation Modelling of a Narrow Aisle Automated Storage and Retrieval System (AS/RS)
Serviced by Rail-guided Vehicles.” Computers in Industry 30 (3): 241–253.
Lee, J. A., Y. S. Chang, and Y. H. Choe. 2017. “Assessment and Comparison of Human-Robot Co-work Order Picking Systems
Focused on Ergonomic Factors.” In Advances in Safety Management and Human Factors. AHFE 2017. Advances in Intelligent
Systems and Computing, edited by P. Arezes, 516–523. vol. 604, Los Angeles, CA: Springer.
eMarketer. 2014. “Retail Sales Worldwide Will Top $22 Trillion This Year.” eMarketer, December 23. https://www.emarketer.com/Arti
cle/Retail-Sales-Worldwide-Will-Top-22-Trillion-This-Year/1011765
Medeiros, D. J., E. E. Enscore, Jr., and A. Smith. 1986. “Performance Analysis of Miniload Systems.” In Proceedings of the 1986
Winter Simulation Conference, December 8–10, edited by J. Wilson, J. Henriksen, and S. Roberts, 606–612. Washington, DC.
Napolitano, M. 2013. “Warehouse and Distribution Center Management: Omni-Channel Distribution – Moving at the Speed of
‘Now’.” Logistics Management, June 1. http://www.logisticsmgmt.com/article/warehouse_dc_operations_omni_channel_distribu
tion_moving_at_the_speed_of_no
Onal, S., J Zhang., and S. Das. 2017. “Modelling and Performance Evaluation of Explosive Storage Policies in Internet Fulfilment
Warehouses.” International Journal of Production Research. Published online.
Park, B. C., E. H. Frazelle, and J. A. White. 1999. “Buffer Sizing Models for End-of-aisle Order Picking Systems.” IIE Transactions
31 (1): 31–38.
Perry, R. F., S. V. Hoover, and D. R. Freeman. 1984. “An Optimum-seeking Approach to the Design of Automated Storage/Retrieval
Systems.” In Proceedings of the 1984 Winter Simulation Conference, November 28–30, edited by S. Sheppard, U. Pooch, and
D. Pegden, 348–354. Dallas, TX.
3858 Y.A. Bozer and F.J. Aldarondo

PricewaterhouseCoopers LLP (PWC). 2014. “The Economic Impact of the U.S. Retail Industry.” National Retail Federation (NRF),
September. https://nrf.com/sites/default/files/Documents/The%20Economic%20Impact%20of%20the%20US%20Retail%20Indus
try%20REV.pdf
Pulat, B. M., and P. S. Pulat. 1989. “Throughput Analysis in an Automated Material Handling System.” Simulation 52 (5): 195–198.
Raghunath, S., R. Perry, and T. Cullinane, 1986. “Interactive Simulation Modeling of Automated Storage Retrieval Systems.” In Pro-
ceedings of the 1986 Winter Simulation Conference, December 8–10, edited by J. Wilson, J. Henriksen, and S. Roberts, 613–
620. Washington, DC.
Roodbergen, K. J., and I. F. Vis. 2009. “A Survey of Literature on Automated Storage and Retrieval Systems.” European Journal of
Operational Research 194 (2): 343–362.
Spearman, M., D. Woodruff, and W. Hopp. 1990. “CONWIP: A Pull Alternative to Kanban.” International Journal of Production
Research 28: 879–894.
Stowe, J. D. 2016. “Throughput Optimization of Multi-Agent Robotic Automated Warehouses.” M.S. Thesis. Massachusetts Institute
of Technology.
Swisslog. 2016. “Tornado Miniload Crane, FactSheet.” Swisslog. Accessed June 23, 2016. http://www.swisslog.com/tornado
Swisslog. 2017. “Swisslog to Deliver First CarryPick® Solution for DB Schenker Logistics.” Swisslog. Accessed November 12,
2017. http://www.swisslog.com/dbschenker
Takakuwa, S. 1996. “Efficient Module-based Modeling for a Large-scale aS/RS-AGV System.” In Proceedings of the 1996 Winter
Simulation Conference. December 8–11, edited by J. M. Charnes, D. J. Morrice, D. T. Brunner, and J. J. Swain, 1141–1148.
Coronado, CA.
Tompkins, J. A., J. A. White, Y. A. Bozer, and J. M. A. Tanchoco. 2010. Facilities Planning. 4th ed. New York: Wiley.
US Census Bureau News. 2016. “Quarterly Retail E-commenrce Sales.” US Department of Commerce, August 16. https://www2.cen
sus.gov/retail/releases/historical/ecomm/16q2.pdf
VanDerLande Industries. 2017. “Automated Storage and Retrieval Systems (aS/RS).” Accessed November 12, 2017. http://www.mhi.
org/members/42059
Wulfraat ,M. 2012. “Is Kiva Systems a Good Fit for Your Distribution Center? An Unbiased Distribution Consultant Evaluation.”
MWPVL Intl., Inc. Accessed May 5, 2015. http://www.mwpvl.com/html/kiva_systems.html
Wulfraat, M. 2013. “How to Select a Split Case Picking System.” 28th Annual Material Handling & Logistics Conference, September
8–11, Park City, UT.
Wurman, P. R., R. D’Andrea, and M. Mountz. 2008. “Coordinating Hundreds of Cooperative, Autonomous Vehicles in Warehouses.”
AI Magazine 29 (1): 9–19.
Zou, B., Y. Gong, X. Xu, and Z. Yuan, 2017. Assignment Rules in Robotic Mobile Fulfilment Systems for Online Retailers. Interna-
tional Journal of Production Research. Published online.
Copyright of International Journal of Production Research is the property of Taylor & Francis
Ltd and its content may not be copied or emailed to multiple sites or posted to a listserv
without the copyright holder's express written permission. However, users may print,
download, or email articles for individual use.