You are on page 1of 20

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/342364113

Robotic mobile fulfilment systems considering customer classes

Article  in  International Journal of Production Research · December 2020


DOI: 10.1080/00207543.2020.1779370

CITATIONS READS

10 461

3 authors, including:

Yeming Gong Zhe Yuan


emlyon business school EMLV Business School
155 PUBLICATIONS   1,530 CITATIONS    14 PUBLICATIONS   190 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Warehousing View project

IT and Management View project

All content following this page was uploaded by Zhe Yuan on 08 September 2020.

The user has requested enhancement of the downloaded file.


International Journal of Production Research

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tprs20

Robotic mobile fulfilment systems considering


customer classes

Yeming Gong , Mingzhou Jin & Zhe Yuan

To cite this article: Yeming Gong , Mingzhou Jin & Zhe Yuan (2020): Robotic mobile fulfilment
systems considering customer classes, International Journal of Production Research, DOI:
10.1080/00207543.2020.1779370

To link to this article: https://doi.org/10.1080/00207543.2020.1779370

Published online: 22 Jun 2020.

Submit your article to this journal

Article views: 12

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=tprs20
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH
https://doi.org/10.1080/00207543.2020.1779370

Robotic mobile fulfilment systems considering customer classes


Yeming Gonga , Mingzhou Jin b and Zhe Yuan c

a EMLYON Business School, Ecully, France; b Department of Industrial Engineering, The University of Tennessee at Knoxville, Knoxville, TN, USA;
c Research Center, EMLV Business School, Leonard de Vinci Pôle Universitaire, Paris La Defense, France

ABSTRACT ARTICLE HISTORY


This paper studies a Robotic Mobile Fulfilment System (RMFS), featured by a number of robots lift- Received 5 September 2019
ing and transporting movables storage shelves from storage grids to order pickers. In such systems, Accepted 31 May 2020
online retailers often classify their customers by two major classes ‘expedited shipping’ and ‘stan- KEYWORDS
dard shipping’. We build high-dimension Markov models to describe this system with customer Warehouse management;
classes, calculate the throughput of this system given the number of robots and provide design facility planning and design;
rules to determine the optimal number of robots and their capacities considering the trade-off order picking; supply chain
between capacities of picker stations and robots. We verify the analytic results of Markov models delivery; robotic mobile
with simulations. We further consider multiple-picker RMFS and study its optimal design. fulfilment system

1. Introduction
of 67.2 billion USD in 2018, is cooperating with Shanghai
This research is inspired by a new Robotic Mobile Ful- Banwi IT Company (see banwi.cn) to build an RMFS sys-
filment System (RMFS), installed by many online retail- tem, and implementing robotic picking in a warehouse of
ers such as Amazon and Zappos’s distribution centre 5000 m2 in Shanghai.
in Kentucky. The new system significantly improves the RMFS applies affordable robots, very simple shelves,
performance of warehouse operations (Hoyt, Lee, and basic conveyers for picker workstations, and centralised
Marks 2009). RMFS is a relatively new order fulfilment control software to provide a so-called ‘Mobile Fulfil-
system featured by a number of robots transporting mov- ment System’. Mobile robot is an important application
able shelves from storage positions, called storage ‘grid’ field for a number of control techniques, mainly dealing
by industry, to order picker workstations (Enright and with constraints (González et al. 2011). Using hundreds
Wurman 2011; Lamballais, Roy, and De Koster 2017a). of autonomous mobile robots and sophisticated control
Robotic warehouse techniques were researched before, software, the intelligent storage system can provide a
but they have not been economically feasible (par- complete order fulfilment system for retailers. In a tra-
tially from the decreasing of robots price) and got real ditional warehouse, goods are moved using conveyors or
implementation until recently (Azadeh, De Koster, and by human operators with tow racks or forklifts (Yuan
Roy 2019). The robotic mobile fulfilment system has been and Gong 2017). In RMFS, products are stored in some
applied to warehouses in firms like Zappos, Walgreens, light, simple and portable storage shelves in a parallel-
Staples, Diapers, Gap, and Boston Scientific (Bell 2019). aisle storage zone. These storage shelves are not fixed to
In 2012, Amazon bought Kiva system for 775 million the floor and can be lifted and moved by robots. When
USD and implemented a number of RMFS systems in an order is released, a robot receives the order infor-
their fulfilment centres (Turban et al. 2018). Some other mation wirelessly and uses cameras to read navigational
material handling companies and robot companies are barcode stickers on the warehouse floor (Scanlon 2009).
developing different robotic picking systems. For exam- The robot then navigates the warehouse to find the appro-
ple, RMT Robotics, a high-tech company in Toronto, priate storage shelves containing the requested item(s),
Canada is developing robotic picking systems and robotic lifts up the storage shelf after sliding under it, and brings
case picking systems (see RMTrobotics.com). ADAM it to a picker workstation. A picker will pick out item(s)
(autonomous delivery and manipulation) robots of RMT and place them in a carton. Completed orders in cartons
companies are similar to Kiva’s robots. Jingdong (see will be moved by conveyors to packers or storage zone
jd.com), one of the largest online retailers with revenue for further handling and then to the docks for shipping.

CONTACT Zhe Yuan zhe.yuan@devinci.fr Research Center, EMLV Business School, Leonard de Vinci Pôle Universitaire, Paris La Defense 92 916, France

© 2020 Informa UK Limited, trading as Taylor & Francis Group


2 Y. GONG ET AL.

Warehouse robots can improve the labour and space effi- 2. Literature review
ciency, reduce the operational cost, and improve both
This research is related to parts-to-picker systems,
investment and operational flexility (Vincent 2018). The
robotic picking, automated guided vehicle system, and
Amazon Robotics operating in fulfilment centres in the
robotic mobile order fulfilment system.
United States can improve nearly 50% capacities of han-
dling the inventory compared with traditional systems
(Welch 2015).
2.1. Parts-to-picker systems
RMFS has been applied in two application settings in
retailing: traditional and online retailers (Razumov 2005; There are two types of warehouse systems: parts-to-
Zou et al. 2017; Hitchcock et al. 2018). For traditional picker and picker-to-parts systems (De Koster, Le-Duc,
retailers, the productivity or throughput is one important and Roodbergen 2007). RMFS is closer to parts-to-picker
measure (Hulten, Dean, and Harper 2007). For online system if we view robots with products as parts. The lit-
retailers, apart from the throughput, the response time erature on parts-to-picker systems, includes automated
from order receiving to home delivery is important (Yuan storage and retrieval systems (AS/RS) (Roodbergen and
and Gong 2017). A typical online retailer may classify Vis 2009), carousels (Gu, Goetschalckx, and McGin-
customers into two or more classes by priorities. For nis 2010), and modular vertical lift modules (VLM)
example, Amazon classifies its customers by two major (Dukic, Opetuk, and Lerher 2015).
classes ‘expedited shipping’ and ‘standard shipping’. This AS/RSs are storage systems using arrays with fixed
paper considers the order fulfilment system with two path storage and retrieval machines running between
classes. fixed storage racks. Providing the access to all stored
Thereby, our research question is: ‘How to describe items, AS/RSs have been widely used in supply chain
RMFS and optimise its operations and designs to improve and logistics industries (Roodbergen and Vis 2009).
its throughput, with considering customer classes?’ To AS/RSs can save floor space and labour costs, reduce
answer the research question, this paper will first con- error rates, and realise high throughput (Rosenblatt, Roll,
sider two customer classes for the RMFS using Markov and Zyser 1993). Obvious drawbacks of AS/RSs are the
models and evaluate its performance for one picker. We low flexibility and high investment costs. Researchers
then build high-dimension Markov models to describe have studied different optimal operation and design
the RMFS with multiple picker stations. We further deter- problems of AS/RSs, including travel-time models (Xu
mine the optimal number and capacity of robots, the et al. 2016; Schenone et al. 2019), request sequencing
optimal number of pickers to maximise throughput. problems (Yang et al. 2017), space estimation model
Finally, we verify analytic results through simulation and (Eldemir, Graves, and Malmborg 2004), throughput opti-
numerical experiments. misation (Yu and Yu 2019), the storage assignment prob-
Our contributions are as follows: lem (Bozer and Cho 2005), and the class-based storage
problem (Roshan, Shojaie, and Javadi 2019).
(1) This is the among the earliest papers to study A carousel system is used for storing products in dis-
RMFSs considering customer classes. Compared tribution centres and manufacturing facilities (Lee and
with single-class system, this is a more realistic appli- Hwang 1988). The literature on carousels is extensive
cation setting for online retailers. (Hassini 2008), and horizontal or vertical carousel sys-
(2) We make methodological contribution with high- tems are suitable for different levels of storing small-
dimension Markov models that are able to capture sized and medium-sized products in distribution centres
more operational details. and manufacturing facilities (Pazour and Meller 2013).
(3) We provide optimal facility designs to balance the Researchers have studied different optimal operation and
capacity of robots, the number of robots, and the design problems of carousels, including throughput per-
number of pickers. Our optimal design can signifi- formances (Litvak 2006), acceleration and deceleration
cantly improve the productivity. problems (Hwang, Song, and Kim 2004).
A new system based on AS/RS, called Vertical Lift
The remainder of the paper is organised as follows: Module (VLM) is another type of parts-to-picker sys-
In the following section, we conduct a literature review. tem with vertically arranged trays stored on two sides
Section 3 presents modelling and analysis for RMFS of the unit, an extractor device placed in the middle,
models with customer classes considering a picker and and an automated control system that delivers items to a
multiple picker stations. Section 4 shows computational workstation (Dukic, Opetuk, and Lerher 2015). A variety
results and experiments. We conclude with our contribu- of configurations and applications change VLM systems
tion and suggestion for further research in Section 5. from manual picking of small items to automated picking
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 3

(Calzavara, Sgarbossa, and Persona 2019). Meller and of AGVs in warehouse systems have considered the
Klote (2004) study throughput models for carousel/VLM transporting capacity (Soukhal, Oulamara, and Mar-
systems with multiple carousels by modelling an approx- tineau 2005), warehouse picking operations using mul-
imate two-server queuing system. Nicolas, Yannick, and tiple AGVs (Farahvash and Boucher 2004), and self-
Ramzi (2018) optimise the practical order batching prob- organising approach for single AGV routing (Umar
lem in an automated warehouse with vertical lift mod- et al. 2015). RMFS is an application using AGVs in ware-
ules and use the real data to validate the model. Li, house systems, but our main research subject is about
Huang, and Dai (2017) optimise the order batching and warehousing operation systems rather than the AGV
picker routing for a warehouse using parts-to-picker sys- itself.
tem. They formulate an integer programming model
with the objective of minimising the total travelling
2.3. Robotic mobile order fulfilment systems
distance.
RMFS fulfils customer orders with mobile technique.
Mobile order fulfilment has been applied in warehouse
2.2. Automated guided vehicle system
operations. Gong and De Koster (2008) study a ware-
An automated guided vehicle (AGV) system is equipped house system that employs labour and a mobile order ful-
with non-contact guided unmanned vehicles, and used filment system to achieve time-based competitive advan-
for repeating transportation tasks in areas of manufac- tage for online retailers. Different from manual picker-to-
turing systems, internal and external transportation sys- parts system, RMFS employs robots and a parts-to-picker
tems, and warehouses (Vis 2006; Sabattini et al. 2017). system with mobile communication and control. Wur-
AGV’s control and design of transport routes have man, D’Andrea, and Mick (2008), a pioneering researcher
become the focus of AGV researches (Baker and in RMFS, studies coordination of hundreds of cooper-
Canessa 2009). The control and design involve the lay- ative autonomous vehicles in warehouses. Enright and
out of the flowpath, including fixed guided paths and Wurman (2011) further study optimisation and coordi-
unfixed guided paths (Aiello, Enea, and Galante 2002), nated autonomy in a mobile fulfilment system. RMFS
vehicle routing and scheduling (Fazlollahtabar and Saidi- mainly uses the centralised control (Merschformann,
Mehrabad 2015), traffic management (Xu et al. 2003), Xie, and Erdmann 2017; Shabanov and Ivanov 2019).
vehicle dispatching, location of pick-up and deliv- Luna and Bekris (2011) propose a centralised algorithm
ery points (Kiran and Tansel 1989), battery manage- using two primitives to describe robots’ activities for solv-
ment (McHaney 1995), vehicle requirements (Bechtsis ing multi-robot path planning problems. Milutinovi and
et al. 2017), and position of idle vehicles (Gademann Lima (2006) model a centralised control with a number
and van de Velde 2000). Our research is related to two of robots considering uncertainty. This control problem
research topics of AGV: the unfixed guided path design is to optimise the cost function. Researchers also study
and vehicle application in warehousing. In guided path other problems for RMFS, such as simulation framework
design with fixed routes, an AGV simply moves along (Merschformann, Xie, and Li 2017), path planning (Mer-
the path determined in advance (Ventura and Lee 2001), schformann, Xie, and Erdmann 2017; Lee et al. 2019),
with the objective to minimise transport time includ- decision rules (Merschformann et al. 2018), and repo-
ing AGVs’s travel time and waiting time. However, other sitioning of storage units (Merschformann 2018). As far
AGVs use a non-oriented path guided system (Singh and as we know, RMFS is one of the earliest business imple-
Tiwari 2002) and can relatively easily change their paths mentations of robotic picking in a general background
(Tompkins et al. 2010). Our research is closer to unfixed of warehouses or distribution centres and for general
guided paths. products.
In the automotive industry, AGV systems are com- There exists the researches of applying robots in ware-
monly used in assembly production lines to increase houses (Guizzo 2008) with real implementation. Knust
the flexibility of transport, such as engine or chas- and Xie (2019) analyse the nurse rostering problem for
sis assembly lines or robotic welding lines (Malsch RMFS considering horizontal and vertical rules. They
and Dohse 1993). AGV is an important equipment propose a mixed integer programming model and use
for the flexible transshipment of work-piece or pal- computational experiments to solve the problem based
lets and transportation in storage systems and auto- on data in Germany and Austria. Azadeh, De Koster,
mated factories (Lu et al. 2017). In manufacturing fields, and Roy (2019) review robotic handling systems which
AGV systems are applied to transport all materials include robotic mobile fulfilment systems. They analyse
related to the manufacturing process (Fazlollahtabar, models and methods to make optimal decisions. Yuan
Saidi-Mehrabad, and Balakrishnan 2015). The studies and Gong (2017) design a speed delivery in an RMFS for
4 Y. GONG ET AL.

online retailers. They build single-service open queue- 3. Model analysis


ing networks for dedicated robots and multiple-service
This section focuses on RMFS with two customer
open queuing networks for pooled robots. They also
classes, including high-priority expedited orders and
consider the congestion in the RMFS. Zou et al. (2017)
low-priority standard orders. We consider two protocols:
study an assignment problem of workstations to robots
R robots serve for either one order picker, or serve for
in an RMFS. They propose new assignment rules and
multiple order pickers. We consider the layout of RMFS
build semi-open queueing networks to find an opti-
shown in Figure 1.
mal assignment rule. Roy et al. (2019) model multi-
Table 1 shows the symbols we use in this paper. To
class closed queueing networks to evaluate the per-
improve the performance of RMFS, we examine three
formance of single and multiple storage zones in an
main decision variables in the optimal design (Lambal-
RMFS, dedicated and pooled robot assignment, with
lais, Roy, and De Koster 2017b; Yuan and Gong 2017).
analysing order picking process and replenishment pro-
(1) We first need to determine the capacity of robots,
cess. Xie et al. (2019) present efficient order picking
featured by velocity. While a slower robot will lower the
methods an integrated model in by minimising capac-
throughput, a faster robot will create bottlenecks at picker
ity of pickers. Feng et al. (2018) study the RMFS’s
workstations and increase incidents such as the item fall-
optimisation problem of location of pickers. Lambal-
off. (2) Another variable is the optimal number of robots
lais, Roy, and De Koster (2020) construct a continu-
to balance with the capacity of pickers for improving the
ous time Markov chain to solve the optimisation inven-
performance. At Amazon, more robots are added into the
tory problem to minimise the throughput time in RMFS
floor during the holiday season. During off-peak seasons,
models. Lamballais, Roy, and De Koster (2017a) anal-
a large number of robots are put into the ‘sleep’ mode.
yse the performance by maximising the throughput
(3) The number of pickers is another decision variable to
of RMFS. They design storage areas and number of
trade off the capacities of picker stations and robots.
robots by estimating the order throughput and the aver-
age order cycle time. Queueing networks are widely
applied for RMFS models, and we analyse the RMFS
model with a picker by finite source networks. Gen- 3.1. RMFS model with customer classes and a picker
eral queueing networks cannot capture more opera- We describe the research problem by high-dimension
tional details. In our paper, we focus on multiple classes, Markov models. For the system with a single customer
we propose a specified Markov model which is based class, a researcher may not apply complex Markov mod-
on state variables, and can capture some operational els since finite source queuing models are enough for
details.

Figure 1. The layout of RMFS.


INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 5

Table 1. Notation. standard orders, excluding the ones waiting for the picker.
Symbol Meaning P is the number of robots that are in front of the picker,
E The number of expedited shipping requests in the system including the waiting ones and the one being processed
R The number of robots by the picker. These P robots may carry either expedited
n The number of pickers
S The number of robots that are processing standard orders
or standard orders. Each robot is either picking up a
P The number of robots that are in front of the picker standard order, in the robot queue before the picker, in
λs The shipping requests arrival rate the process with the picker, or picking up an expedited
λe The expedited shipping requests arrival rate
μ The distribution rate of a picker order. When P = 0 or 1, there is no robot queue before
TH The throughput of the picker station and system the picker. When robots are currently in the way of get-
ting expedited shipping orders, they are not in the robot
queue. In other words, expedited requests are waiting
analysis (see Appendix A.1). We assume that the RMFS for being processed (i.e. waiting for an available robot).
uses a random storage policy, which is common for some For example, (2, R − 1, 0) means the picker is idle, one
online retailers. In this section, we consider R robots and robot is working on one expedited shipping order, R−1
one picker. The case will be extended to multiple pickers robots are working on standard shipping order, and one
in Section 3.2. The throughput of the RMFS is denoted expedited shipping request is waiting for the next avail-
by TH. A typical online retailer uses a hybrid opera- able robot. For another example, (4, R − 5, 3) means the
tions mode, classifying customer orders into classes by picker is busy, two robots are waiting for the service from
their priorities. We use Markov models to describe RMFS the picker, R−5 robot are working on standard shipping
with two classes, including ‘standard shipping’ and ‘expe- orders, two robots are working on expedited shipping
dited shipping’. Furthermore, we consider the following orders, and two expedited shipping requests are wait-
system: ing for the next available robot. The state space is {S =
0, . . . , R; P = 0, . . . , R − S; E = R − S − P, . . . , 0}.
(1) The RMFS is applied to handle both expedited and We show the transition diagram in Appendix A.2. We
standard shipping orders. A set of robots are ded- denote the equilibrium probability for the state of (E, S, P)
icated to a picker. Each picker and her associated as Pr(E, S, P) and can calculate the throughput of RMFS
robots are independent from other pickers. by Proposition 3.1 and the lead time of an expedited order
(2) The expedited shipping requests arrive in the system by Proposition 3.2.
following Poisson Process with the rate of λe and will
be immediately processed when a robot is available. Proposition 3.1: The throughput of the picker station and
(3) Plenty of standard requests are waiting for being pro- system TH is
cessed because we are interested in the throughput.  

R ∞

(4) The system has R robots, each of which can han-
μ 1− Pr(E, S, 0) . (1)
dle both expedited and standard shipping orders.
S=0 E=R−S
It takes exponentially distributed time for a robot
from leaving the picker without an order to coming It is the actual service rate of the picker. The picker
back with the shelf with an order (either expedited is idle when no robots are with the picker. When S
or standard). The mean of the time is 1/λs . robots are dealing with standard orders, the remaining
(5) It takes the picker an exponentially distributed time, R−S robots are dealing with expedited orders and other
with the mean of 1/μ, to unload an order from a expedited shipping requests are in the queue.
robot.
(6) A robot, once unloaded by the picker, will pro- Proposition 3.2: The lead time of an expedited order is
cess an expedited request if any expedited request
waits. Otherwise, the robot will immediately process Le 1 1 LR
+ + + . (2)
a standard request. λe μ λS μ(1 − Prpi )

We define the state of the Markov Chain by a triplet of The proof of Proposition 3.2 is shown in Appendix
(E, S, P). E is the number of expedited shipping requests A.3. It is not easy to directly compute the equilibrium
in the system, including both the ones in the expedite probability of the complex Markov system. We present
waiting line and those are processed by robots. How- methods and an algorithm to specify parameter matrix
ever, E does not include the expedited requests that are of these high-dimension Markov models, and make
currently in the queue waiting for or in the process by the complex calculation feasible (see Appendices A.4
the picker. S is the number of robots that are processing and A.5).
6 Y. GONG ET AL.

The objective function of RMFS model with customer We define the state of the Markov Chain by (E, S, P1 ,
classes and a picker is . . . , Pn ). E is the number of expedited shipping requests
 ∞
 in the system, including the ones in the expedite waiting

R 
max TH(μ, R, λs ) = μ 1 − Pr(E, S, 0) . line and the ones that are processed by robots. However,
S=0 E=R−S
E does not include the expedited request that is currently
(3) in the queue waiting for or processed by pickers. S is the
number of robots that are processing standard orders,
excluding the ones waiting for pickers. Pi is the num-
3.2. RMFS model with multiple pickers
ber of robots that are in front of picker i, including the
In this section, we further consider the application of waiting ones and ones being processed by to-share or
RMFS to online retailers with multiple picker worksta- not-to-share model. Pi includes robots which carry either
tions and two classes, including classes ‘Standard Ship- expedited or standard orders.
ping’ and ‘Expedited Shipping’. When the number of Each robot is either picking up a standard order, in the
pickers n ≥ 1, we need to handle the problem of shar- robot queue before pickers, in the process with pickers, or
ing pickers. The throughput
 of  not-to-share model is picking up an expedited order. When Pi = 0 or 1, it exists
given by TH = nμ(1 − RS=0 ∞ E=R−S Pr(E, S, 0)) with no robot queue n before picker i. When S + n
i=1 Pi <
the assumptions in Section 3.1. R, R − S − i=1 Pi robots are currently in the way of
For the to-share model, we consider the following getting expedited
n shipping orders. In other words, E −
system: (R − S − i=1 Pi ) expedited requests are waiting for
being processed (i.e. waiting for an available robot).
(1) An RMFS is applied to handle both expedited and The state space is {S = 0, . . . , R; P1 = 0,  . . . , R − S; P2 =
standard shipping orders. The RMFS has n pickers. 0, . . . , R− S − P1 ; Pn =  0, . . . , R − S − n−1i=1 Pi ; E = R
n n
(2) An RMFS has R robots, each of which can han- − S − i=1 Pi , R − S − i=1 Pi + 1, . . . , ∞}. Appendix
dle both expedited shipping and regular shipping. It A.6 lists possible transitions if the current state is
takes exponentially distributed time for a robot from (E, S, P1 , . . . , Pn ).
leaving a picker without an order to coming back To calculate the throughput of RMFS system TH with
with the shelf with an order (either expedite or stan- n pickers, we apply the following proposition:
dard to the same or different picker). The mean of
this time is 1/λs . The exponential distribution could Proposition 3.3: The throughput of RMFS system TH
be justified by the large value of R. with n pickers is
(3) The expedited shipping requests arrive in the sys-
⎛ n−1
tem following a Poisson Process with the rate of λe 
n 
R 
R−S R−S−
i=1 Pi
and will be immediately processed when a robot is TH = μi ⎝1 − ···
available. i=0 S=0 P1 =0 Pn =0
(4) Plenty of standard shipping requests are waiting ⎞


for being processed because we are estimating the
× Pr(E, S, P1 , . . . , Pi = 0, . . . , Pn )⎠.
capacity of the system for expedited orders. n
E=R−S− i=1 Pi
(5) It takes the picker i, i = 1, . . . , n, an exponentially
distributed time, with the mean of 1/μi , to pick an (4)
order from a robot. In practice, the service rates
across pickers can vary. For a pick workstation i, its throughput is
(6) A robot that is once unloaded by a picker will process ⎛ n−1
R−S− Pi
an expedited request if there are expedited requests 
R 
R−S i=1
waiting. Otherwise, the robot will immediately pro- μi ⎝1 − ···
cess a standard request. S=0 P1 =0 Pn =0

(7) The RMFS has n + 1 queues in the system: (1) ∞

expedited shipping requests waiting for available × Pr(E, S, P1 , . . . , Pi = 0, . . . , Pn )⎠ .
robots, and (2) robots with picked orders waiting for n
E=R−S− i=1 Pi
picker i.
(8) A robot is assigned to the picker who has the min- It is the actual service rate of the picker i. The picker i
imum number of robots in front of her, including is idle when Pi = 0 (no robots are with the picker). The
the ones in the queue and the ones under service. Pi throughput of RMFS system with n pickers is the sum of
robots stay in front of picker i. actual service rates of n pickers.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 7

Therefore, the objective function of to-share RMFS


model with customer classes and multiple pickers is

max TH(μi , R, λs , n)
⎛ 
R−S− n−1
n R 
R−S i=1 Pi
= μi ⎝1 − ···
i=0 S=0 P1 =0 Pn =0



× Pr(E, S, P1 , . . . , Pi = 0, . . . , Pn )⎠ .
n
E=R−S− i=1 Pi
(5)

The objective function of not-to-share RMFS model


with customer classes and multiple pickers is
 ∞

 R  Figure 2. Simulation of RMFS models with two classes and one
max TH(μ, R, λs , n) = nμ 1 − Pr(E, S, 0) .
picker.
S=0 E=R−S
(6)

4. Numerical experiments
In this section, we present numerical experiments to
verify RMFS with customer classes and multiple pick-
ers by simulation and to evaluate the performance of
RMFS, including comparing different operational pro-
tocols in terms of throughput. We also solve optimal
decision problems to determine the number of robots,
their capacities, and the number of pickers.

4.1. Verification of analytic results by simulation


We use Arena 14.0 to simulate the RMFS with two classes
and one picker(see Figure 2) and multiple pickers (see
Figure 3). For an RMFS with ‘expedite orders’ and ‘stan-
dard orders’, we simulate it based on assumptions pre-
Figure 3. Comparison of the analysis and simulation for RMFS
sented in Section 3.1 for 1000 warehouse operational
models with multiple pickers.
hours. The average simulation running time is 18.49 sec-
onds per run on a computer with an ‘Intel Core i7-4600U’
CPU@2.1GHz and 16 GB RAM. We calculate the RMFS runs. The average bias between analytical and simulation
model specified by Table A1 and Proposition 3.1, and use is 1.11% based on 360 groups with different parameters
Matlab to calculate results by the analytic Markov model λs , λe and μ, with values R = 1:30, μ for 3 times, λs and
(for future details in algorithms and implementation, see λe for twice respectively. g is the scenario index, featured
Appendices A.4 and A.5). by different values of μ and R. Figure 2 presents analyti-
We define the bias as cal and simulates results when μ = 5, μ = 7 and μ = 9.
The units of parameters λs , λe , and μ are units/second.
1  |THg − THg |
M A S
× 100%, RMFS models were specified by Table A1 and Propo-
M THgS sition 3.3. Figure 3 is the simulation of RMFS system with
g=1
multiple pickers. The average bias between analytical and
where THgA is the throughput value obtained using the simulation is 3.79% based on 810 groups with different
analytical Markov models and THgS is the throughput parameters in n and μ, with values R = 1:30, n = 2:10,
value obtained by the simulation, g is the index of sim- μ for 3 times. Figure 3 presents analytical and simulative
ulation run, and M is the total number of simulation results with parameters of μ = 4, n = 2.
8 Y. GONG ET AL.

Figure 4. The impact of ‘expedited shipping’ requests on the Figure 5. Comparison of the throughput of RMFS model with
throughput. single class and two classes.

4.2. Performance evaluation two classes increases until the number of robots reaches
4.2.1. The impact of ‘expedited shipping’ requests on nine, where the throughput is 4.99, equals to μ, because
the throughput Prob(P = 0) ≈ 0. If R is infinite, the throughput TH of
Figure 4 shows the influences of expedited order on the this system is μ. If R is less than arg TH(μ), discussing
throughput of RMFS.We change the arrival rates of expe- the difference of total throughput between single class
dited orders λe and the number of robots R to observe the and two classes is interesting. We define the increasing
change of throughput with the λe . We interestingly find rate as |THR − THR−1 |/THR−1 × 100%, where THR is
that when the number of robots R is lower than a cer- the throughput when there are R robots in the system.
tain number, we can observe the influence of λe . Given The highest increasing rate of two classes is 84.91% from
a λs = 2 in this group of experiments, when the num- R = 1 to R = 2. The highest increasing rate of single class
ber of robots R is less than 25, the throughput increases is 55.56%, also from R = 1 to R = 2. When R < 6, the
with λe . When the number of robots R is large enough increasing rate of two classes is higher than the increas-
(after 25 robots), the expedited order has no significant ing rate of single class. When R = 6, the increasing rate
effect on the throughput. When the system remains sta- of two classes is 4.46%, which is lower than the increas-
ble, the arrival rate of the expedited order can affect the ing rate of single class 5.88%. When R > 6, the increasing
throughput under the condition that the number and the rate of single class is higher than the increasing rate of
capacity of robots cannot handle. But when the number two classes. The largest throughput gap between single
and the capacity of robots are large enough to handle, the class and two classes is 1.33 and the highest increasing
throughput will be mainly influenced by the capacity of rate is 37.67%, when R = 6. The average throughput gap
the picker, μ. between single class and two classes is 0.71 for 1 ≤ R ≤
30. In sum, the function TH(R) of two classes increases
4.2.2. Comparison of two protocols: single class or faster and it also has a faster convergence than the func-
two classes tion TH(R) of single class. When R = 6, the performance
We analyse the RMFS of a single class presented in gap between two classes and single class reaches the
Appendix A.1 using finite source models, and Markov maximum.
models to analyse the RMFS of two classes described in
Section 3.1. We use Matlab to calculate two analytic mod- 4.2.3. Comparison of one-picker and multiple-pickers
els with the parameters of μ = 5, λs = 2 for single class models
and μ = 5, λs = 2, λe = 1 for two classes. The analysis of one-picker model and multiple-pickers
Figure 5 shows that TH(R) is an increasing function with to-share model are respectively presented in
with the number of robots R, when 1 ≤ R ≤ 30. The Section 3.1 and Section 3.2. As the number of robots
throughput of RMFS model with two classes is higher R increases as shown in Figure 6, the throughput
than the throughput of RMFS model with single class. As gap between the number of pickers n = 3 and n = 1
the number of robots R increases, the function TH(R) of increases until R = 7. After R = 7, the increasing rate
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 9

Figure 7. Determine the number of robots.


Figure 6. Comparison of one-picker and multiple-pickers mod-
els.

of n = 3 and n = 1 are both close to zero. The maxi-


mum and average throughput gap between n = 3 and
n = 1 is 9 and 6.21 respectively. We define the gap rate as
|TH3 − TH1 |/TH1 × 100%, where TH3 is the through-
put with 3 pickers, TH1 is the throughput with a picker.
The maximum and the average gap rate between n = 3
and n = 1 is 472.65% and 244.55% respectively.

4.3. Optimal decision


4.3.1. Determine the optimal number of robots
Figure 7 shows TH(R) is an increasing function in R and
it will converge into a value TH¯ = nμ. After TH(R) con-
verges, the system cannot significantly benefit from the
increase of R. Considering the cost of robots, we can set
the following condition,|(TH(R) − TH)/ ¯ TH|¯ ≤ ξr such
that the relative difference between TH(R) and converged
value TH ¯ is less than a very small value ξr . ξr is deter-
Figure 8. Determine the capacity of robots.
mined by the tradeoff between the increase of TH and
the cost of robots.
We can get the optimal number of robots by
4.3.2. Determine the capacity of robots


¯
TH(R) − TH Figure 8 shows TH(λs ) is an increasing function of λs

R = min R :
¯
TH ≤ ξr , (7) and it will converge into a value TH ¯ = nμ. The sys-
tem cannot significantly benefit from the increase of λs
which is the smallest value of R to satisfy the condition. when it is high. We can set the following condition,
We can use Figure 7 to determine the number of ¯ TH|
|(TH(λs ) − TH)/ ¯ ≤ ξλs such that the relative differ-
robots. For example, when we set ξr = 0.5% and have ence between TH(λs ) and the converged value TH ¯ is less
μ = 9, the optimal value is (11, 8.96), where the value than a very small value ξλs .
of 11 denotes the optimal number of robots, the value of We can get the optimal capacity of robots by
8.96 denotes the corresponding throughput. When μ =
8, the optimal value is (11, 7.98). When μ = 7, the opti-

mal value is (10, 6.98). When μ = 6, the optimal value is ¯
TH(λs ) − TH
λ∗s = min λs : ≤ ξλ ,
(8)
(9, 5.98). When μ = 5, the optimal value is (8, 4.98). TH¯ s
10 Y. GONG ET AL.

Figure 10. Contour lines with considering the capacity of robots


Figure 9. Jointly determine the capacity λs and the number of
λs and the number of robots R.
robots R given a target TH.

which is the smallest value λs to satisfy the condition.


Since the increase of λs and the capacity of robots will
lead to cost.
We can use Figure 8 to determine the capacity of
robots with the increasing parameters of λs . For exam-
ple, if we set ξλs = 0.5%, when μ = 9 the optimal value
is (3.5, 8.97), where 3.5 denotes the optical capacity of
robots and 8.97 denotes the throughput with the opti-
mal capacity of robot. When μ = 8, the optimal value is
(3, 7.96).When μ = 7 the optimal value is (3, 6.99).

4.3.3. Jointly determine the capacity of robots and


the number of robots
We use Matlab to calculate the analytic RMFS model,
denoting λs as the capacity of robots. We can use Figure 9
to jointly determine the capacity of robots considering Figure 11. Jointly determine the capacity of pickers μ and the
number of robots R.
the throughput and the contour line of R and λs when
the throughput reaches optimal value. We define contour
lines of R and λs as follows. Given a target TH 0 , a range 4.3.4. Jointly determine the capacity of a picker and
1 ≤ R ≤ R̄ for R, a range λs ≤ λs ≤ λ̄s , a contour line is a the number of robots
set of points (R, λs ) ∈ {(R, λs ) | TH(R, λs ) = TH 0 , ∀1 ≤ We use Matlab to calculate the analytic RMFS model,
R ≤ R̄, λs ≤ λs ≤ λ̄s }. denoting μ as the capacity of a picker. We can
Figure 9 shows that TH(R, λs ) with the number of use Figure 11 to jointly determine the capacity of a
robots R increases and the capacity of robots λs . The picker considering the throughput and the contour
maximal throughput is 9 when μ = 9 and λe = 1. line TH(R, μ) given a target throughput TH 0 . We
Figure 10 shows contour lines TH(R, λs ) = TH 0 with define contour lines of R and μ as follows. Given a
considering the capacity of robots. Contour lines respec- target TH 0 , a range 1 ≤ R ≤ R̄ for R, a range μ ≤
tively represent TH = 9 to TH = 1. For the same μ ≤ μ̄, a contour line is a set of points (R, μ) ∈
throughput TH 0 , as the capacity of robots increases, the {(R, μ) | TH(R, μ) = TH 0 , ∀1 ≤ R ≤ R̄, μ ≤ μ ≤ μ̄}.
needed number of robots declines. A warehouse manager Figure 11, a three-dimensional figure, shows how the
can use Figure 9 in her optimal planning. For example, if throughput changes with the number of robots R in a
she sets TH = 8 as the target for a picker and wants to range 1 ≤ R ≤ 20 and the capacity of pickers μ in a
put six robots in an RMFS for a picker, the λs can be set range 1 ≤ μ ≤ 9. The maximal throughput is 9 when
as two. μ = 9, 13 ≤ R ≤ 20. In sum, when R ≤ 12 the change
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 11

Figure 12. Contour lines with the capacity of pickers. Figure 13. Jointly determine the number of pickers and the num-
ber of robots.

of throughput depends more on the number of robots.


When R > 12, the change of throughput depends more
on the capacity of pickers.
Figure 12 shows contour lines with considering the
capacity of pickers. Contour lines respectively represent
TH(R, μ) = 9 to TH(R, μ) = 1. For the same through-
put TH 0 , as the capacity of pickers increases, the needed
number of robots declines. We find the following obser-
vations:

Observation 4.1: When μ is relatively small for a TH 0 ,


TH is not sensitive to R. It is unnecessary for the RMFS
manager to increase the number of robots R.

Observation 4.2: When R is relatively small for a TH 0 , Figure 14. Contour lines with the number of pickers.
TH is not sensitive to μ. It is unnecessary for the RMFS
manager to increase the capacity of pickers μ.
the throughput reaches optimal value. We define contour
Figure 12 can explain Observation 1 if the capacity of a lines of R and n as follows. Given a target TH 0 , a range
picker μ < 3, the vatical line for contour line shows that 1 ≤ R ≤ R̄ for R, and a range n ≤ n ≤ n̄, a contour line is
TH is not sensitive to R. Figure 12 can explain Observa- a set of points (R, n) ∈ {(R, n) | TH(R, n) = TH 0 , ∀ 1 ≤
tion 2. For example, if the number of robots R ≤ 2, the R ≤ R̄, n ≤ n ≤ n̄}.
horizontal lines for contour line from TH(R, μ) = 4 to The three-dimensional Figure 13 shows how the
TH(R, μ) = 9 show that TH is not sensitive to μ. In sum, throughput changes with the number of robots and the
the capacity of pickers should be greater than or equal number of pickers. The maximal throughput is 15.99,
to 3. When the number of robots is large, increasing the when n = 6 and R = 15. When n ≤ 3, the throughput
capacity of pickers is an effective method to increase the increases with the number of robots and the number
throughput. of pickers increase. For n ≥ 4, the throughput decreases
when R ≤ 2 and then increases with the number of
4.3.5. Jointly determine the number of pickers and robots. When R > 6, the change of throughput depends
the number of robots more on the number of pickers.
We use Matlab to calculate the analytic multiple-pickers Figure 14 shows contour lines with considering the
RMFS model. We can use Figure 13 to jointly deter- number of pickers. Contour lines respectively repre-
mine the number of pickers and the number of robots sent TH(R, n) = 16 to TH(R, n) = 5. We discuss contour
considering the throughput and the contour line when lines in two cases: (1) When TH(R, n) = 7, 8 . . . , 16, as
12 Y. GONG ET AL.

the number of pickers increases, the needed number National Social Science Foundation of China (16ZDA013),
of robots is with a curve shape of ‘’. This is relative and Erasmus+ [grant number 2019-1-FR01-KA203-063063].
to which one (pickers vs. robots) is the bottleneck for Yeming GONG issupported by Business Intelligence Center of
EMLYON.
the performance. Taking TH(R, n) = 8 as an example,
when the number of pickers is less than 2, the bottle-
neck is at the pickers. Increasing the number of robots ORCID
does not increase throughput too much. When the num-
Mingzhou Jin http://orcid.org/0000-0002-2387-8129
ber of robots is less than 4, the bottleneck is at the
Zhe Yuan http://orcid.org/0000-0002-3656-8840
robots. Increasing the number of pickers does not signif-
icantly increase throughput. (2) When TH(R, n) = 5, 6,
the needed number of robots increases with the number
of pickers. Generally speaking, only when the number References
of robots is large, increasing the number of pickers is Aiello, G., M. Enea, and G. Galante. 2002. “An Integrated
an effective method to increase the throughput. A man- Approach to the Facilities and Material Handling System
ager can use contour lines to optimise the planning of an Design.” International Journal of Production Research 40
RMFS. For example, given a target TH 0 = 16 and n = 5, (15): 4007–4017.
Azadeh, Kaveh, René De Koster, and Debjit Roy. 2019.
she can set an optimal number of robots R = 9.
“Robotized and Automated Warehouse Systems: Review
and Recent Developments.” Transportation Science 53 (4):
917–945.
5. Concluding remarks
Baker, Peter, and Marco Canessa. 2009. “Warehouse Design:
This paper studies a new order fulfilment system featured A Structured Approach.” European Journal of Operational
by a number of robots transporting movables shelves to Research 193 (2): 425–436.
Bechtsis, Dimitrios, Naoum Tsolakis, Dimitrios Vlachos, and
order pickers. In our model, we consider order classes, Eleftherios Iakovou. 2017. “Sustainable Supply Chain Man-
expedited and standard, to represent the needs of online agement in the Digitalisation Era: The Impact of Auto-
retailers to provide the responsive service to their cus- mated Guided Vehicles.” Journal of Cleaner Production 142:
tomers. We use throughput to measure the performance 3970–3984.
of the system along with the lead time of expedited Bell, Jennie. 2019. “Zappos Explores Fresh Customer Service
Tactics With New Kentucky Fulfillment Center & Store.”
orders. We develop Markov models to build the relation-
https://footwearnews.com/2019/business/retail/zappos-
ship between those performance metrics and system con- fulfillment-center -retail-outlet-louisville-kentucky-120277
figuration, featured by the number of robots, the capacity 0182/.
of robots, and the number of pickers. Numerical results Bozer, Yavuz A., and Myeonsig Cho. 2005. “Throughput Per-
provide design rules to balance the three design variables. formance of Automated Storage/retrieval Systems Under
Warehouse managers can use those rules to determine Stochastic Demand.” IIE Transactions 37 (4): 367–378.
Calzavara, Martina, Fabio Sgarbossa, and Alessandro Persona.
the optimal configuration of the system. 2019. “Vertical Lift Modules for Small Items Order Picking:
The limitation of this study is that we mainly consider An Economic Evaluation.” International Journal of Produc-
stable stochastic systems, and have not considered insta- tion Economics 210: 199–210.
ble stochastic systems. We use approximate methods to De Koster, René, Tho Le-Duc, and Kees Jan Roodbergen. 2007.
construct and estimate the Markov models. It is interest- “Design and Control of Warehouse Order Picking: A Litera-
ture Review.” European Journal of Operational Research 182
ing to consider non-Poisson process for order arrivals.
(2): 481–501.
This paper could be further extended to study different Dukic, Goran, Tihomir Opetuk, and Tone Lerher. 2015. “A
the dwell point policies for the returned shelves. Other Throughput Model for a Dual-tray Vertical Lift Module with
rules of placing shelves may further help to improve the a Human Order-picker.” International Journal of Production
productivity of the RMFS but require new models and Economics 170: 874–881.
computational methods. The robots charging issue could Eldemir, Fahrettin, R. J. Graves, and C. J. Malmborg. 2004.
“New Cycle Time and Space Estimation Models for
also be an interesting future topic. Automated Storage and Retrieval System Conceptualiza-
tion.” International Journal of Production Research 42 (22):
Disclosure statement 4767–4783.
Enright, John J., and Peter R. Wurman. 2011. “Optimization
No potential conflict of interest was reported by the author(s). and Coordinated Autonomy in Mobile Fulfillment Systems.”
Association for the Advancement of Artificial Intelligence
Workshop, WS-11-09, 33–38.
Funding
Farahvash, Pooya, and Thomas O. Boucher. 2004. “A Multi-
This research was supported in part by the National Nat- agent Architecture for Control of AGV Systems.” Robotics
ural Science Foundation of China (No. 71620107002), the and Computer-Integrated Manufacturing 20 (6): 473–483.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 13

Fazlollahtabar, Hamed, and Mohammad Saidi-Mehrabad. Lamballais, Tim, Debjit Roy, and M. B. M. De Koster. 2017b.
2015. “Methodologies to Optimize Automated Guided Vehi- “Estimating Performance in a Robotic Mobile Fulfillment
cle Scheduling and Routing Problems: A Review Study.” System.” European Journal of Operational Research 256 (3):
Journal of Intelligent & Robotic Systems 77 (3-4): 525–545. 976–990.
Fazlollahtabar, Hamed, Mohammad Saidi-Mehrabad, and Lamballais, Tim, Debjit Roy, and René B. M. De Koster. 2020.
Jaydeep Balakrishnan. 2015. “Mathematical Optimization “Inventory Allocation in Robotic Mobile Fulfillment Sys-
for Earliness/tardiness Minimization in a Multiple Auto- tems.” IISE Transactions 52 (1): 1–17.
mated Guided Vehicle Manufacturing System via Integrated Lee, Moon-Kyu, and Hark Hwang. 1988. “An Approach in the
Heuristic Algorithms.” Robotics and Autonomous Systems Design of a Unit-load Automated Carousel Storage System.”
72: 131–138. Engineering Optimization 13 (3): 197–210.
Feng, Lijuan, Mingyao Qi, Shijia Hua, and Qingte Zhou. 2018. Lee, C. K. M, Bingbing Lin, K. K. H. Ng, Yaqiong Lv, and
“Picking Station Location in Traditional and Flying-V Aisle W. C. Tai. 2019. “Smart Robotic Mobile Fulfillment System
Warehouses for Robotic Mobile Fulfillment System.” In 2018 with Dynamic Conflict-free Strategies Considering Cyber-
IEEE International Conference on Industrial Engineering and physical Integration.” Advanced Engineering Informatics 42:
Engineering Management (IEEM), 1436–1440. IEEE. 100998.
Gademann, Arnoldus Johannes Reinerus Maria, and Steef L. Li, Jianbin, Rihuan Huang, and James B. Dai. 2017. “Joint
van de Velde. 2000. “Positioning Automated Guided Vehicles Optimisation of Order Batching and Picker Routing in the
in a Loop Layout.” European Journal of Operational Research Online Retailer’s Warehouse in China.” International Journal
127 (3): 565–573. of Production Research55 (2): 447–461.
Gong, Yeming, and Rene De Koster. 2008. “A Polling-based Litvak, Nelly. 2006. “Optimal Picking of Large Orders in
Dynamic Order Picking System for Online Retailers.” IIE Carousel Systems.” Operations Research Letters34 (2): 219–227.
Transactions 40 (11): 1070–1082. Lu, Shaoping, Chen Xu, Ray Y. Zhong, and Lihui Wang. 2017.
González, Ramón, Mirko Fiacchini, José Luis Guzmán, “A RFID-enabled Positioning System in Automated Guided
Teodoro Álamo, and Francisco Rodríguez. 2011. “Robust Vehicle for Smart Factories.” Journal of Manufacturing Sys-
Tube-based Predictive Control for Mobile Robots in Off- tems 44: 179–190.
road Conditions.” Robotics and Autonomous Systems 59 (10): Luna, Ryan, and Kostas E. Bekris. 2011. “Efficient and Complete
711–726. Centralized Multi-Robot Path Planning.” In 2011 IEEE/RSJ
Gu, Jinxiang, Marc Goetschalckx, and Leon F. McGinnis. 2010. International Conference on Intelligent Robots and Systems,
“Research on Warehouse Design and Performance Evalua- 3268–3275. IEEE.
tion: A Comprehensive Review.” European Journal of Oper- Malsch, Thomas, and Knuth Dohse. 1993. Breaking From Tay-
ational Research 203 (3): 539–549. lorism: Changing Forms of Work in the Automobile Industry.
Guizzo, Erico. 2008. “Three Engineers, Hundreds of Robots, Cambridge, United Kingdom: Cambridge University Press.
One Warehouse.” IEEE Spectrum 45 (7): 26–34. McHaney, R. 1995. “Modelling Battery Constraints in Dis-
Hassini, Elkafi. 2008. “Carousel Storage Systems.” In Facil- crete Event Automated Guided Vehicle Simulations.” Inter-
ity Logistics: Approaches and Solutions to Next Generation national Journal of Production Research 33 (11): 3023–3040.
Challenges, 199–234, USA: CRC Press. Meller, Russell D., and John F. Klote. 2004. “A Throughput
Hitchcock, Riley James, Sarah A. McElroy, Darlene A. Veloso, Model for Carousel/VLM Pods.” IIE Transactions 36 (8):
Edmund James Golby Spencer, Corrine C. Barchanowicz, 725–741.
Aram Marabyan, Erik Davin Kokkonen, Joshua Lynton Merschformann, Marius. 2018. “Active Repositioning of Stor-
Lowy, and Brian Porter. 2018. “Robotic Systems and Meth- age Units in Robotic Mobile Fulfillment Systems.” In Opera-
ods in Prediction and Presentation of Resource Availability.” tions Research Proceedings 2017, 379–385. Springer.
US Patent App. 15/859,225, May 3. Merschformann, Marius, Tim Lamballais, René de Koster, and
Hoyt, David, Hau Lee, and Michael Marks. 2009. “Zappos.com: Leena Suhl. 2018. “Decision Rules for Robotic Mobile Ful-
Developing a Supply Chain to Deliver Wow.” fillment Systems.” Preprint. arXiv:1801.06703.
Hulten, Charles R., Edwin R. Dean, and Michael Harper. 2007. Merschformann, Marius, Lin Xie, and Daniel Erdmann. 2017.
New Developments in Productivity Analysis. Vol. 63. Chicago, “Path Planning for Robotic Mobile Fulfillment Systems.”
USA: University of Chicago Press. Preprint. arXiv:1706.09347.
Hwang, Hark, Young-keun Song, and Kap-hwan Kim. 2004. Merschformann, Marius, Lin Xie, and Hanyi Li. 2017.
“The Impacts of Acceleration/deceleration on Travel Time “RAWSim-O: A Simulation Framework for Robotic Mobile
Models for Carousel Systems.” Computers & Industrial Engi- Fulfillment Systems.” Preprint. arXiv:1710.04726.
neering 46 (2): 253–265. Milutinovi, Dejan, and Pedro Lima. 2006. “Modeling and
Kiran, Ali S., and Barbaros C. Tansel. 1989. “Optimal Pickup Optimal Centralized Control of a Large-size Robotic Pop-
Point Location on Material Handling Networks.” Interna- ulation.” IEEE Transactions on Robotics 22 (6): 1280–
tional Journal of Production Research 27 (9): 1475–1486. 1285.
Knust, Frederik, and Lin Xie. 2019. “Simulated Anneal- Nicolas, Lenoble, Frein Yannick, and Hammami Ramzi. 2018.
ing Approach to Nurse Rostering Benchmark and Real- “Order Batching in An Automated Warehouse with Several
world Instances.” Annals of Operations Research 272 (1-2): Vertical Lift Modules: Optimization and Experiments with
187–216. Real Data.” European Journal of Operational Research 267
Lamballais, Tim, Debjit Roy, and M. B. M. De Koster. 2017a. (3): 958–976.
“Estimating Performance in a Robotic Mobile Fulfillment Pazour, Jennifer A., and Russell D. Meller. 2013. “The Impact of
System.” European Journal of Operational Research 256 (3): Batch Retrievals on Throughput Performance of a Carousel
976–990. System Serviced by a Storage and Retrieval Machine.”
14 Y. GONG ET AL.

International Journal of Production Economics 142 (2): Vincent, J. 2018. “Welcome to the Automated Warehouse of the
332–342. Future.” The Verge.
Razumov, Sergey. 2005. “Robotic Retail Facility.” US Patent Vis, Iris F. A.. 2006. “Survey of Research in the Design and
App. 10/832,383, October 27. Control of Automated Guided Vehicle Systems.” European
Roodbergen, Kees Jan, and Iris F. A. Vis. 2009. “A Survey of Journal of Operational Research 170 (3): 677–709.
Literature on Automated Storage and Retrieval Systems.” Welch, Adrienne. 2015. “A Cost-Benefit Analysis of Amazon
European Journal of Operational Research 194 (2): 343–362. Prime Air.”
Rosenblatt, Meir J., Yaakov Roll, and D. Vered Zyser. 1993. Wurman, Peter R., Raffaello D’Andrea, and Mounte Mick. 2008.
“A Combined Optimization and Simulation Approach for “Coordinating Hundreds of Cooperative, Autonomous
Designing Automated Storage/retrieval Systems.” IIE Trans- Vehicles in Warehouses.” AI Magazine 29 (1): 9–19.
actions 25 (1): 40–50. Xie, Lin, Nils Thieme, Ruslan Krenzler, and Hanyi Li. 2019.
Roshan, K., A. Shojaie, and M. Javadi. 2019. “Advanced Allo- “Efficient Order Picking Methods in Robotic Mobile Fulfill-
cation Policy in Class-based Storage to Improve AS/RS Effi- ment Systems.” Preprint. arXiv:1902.03092.
ciency Toward Green Manufacturing.” International Journal Xu, Fuyi, Hendrik Van Brussel, Marnix Nuttin, and Ronny
of Environmental Science and Technology 16 (10): 5695–5706. Moreas. 2003. “Concepts for Dynamic Obstacle Avoidance
Roy, Debjit, Shobhit Nigam, René de Koster, Ivo Adan, and and Their Extended Application in Underground Naviga-
Jacques Resing. 2019. “Robot-storage Zone Assignment tion.” Robotics and Autonomous Systems 42 (1): 1–15.
Strategies in Mobile Fulfillment Systems.” Transportation Xu, Xianhao, Bipan Zou, Guwen Shen, and Yeming Gong.
Research Part E: Logistics and Transportation Review 122: 2016. “Travel-time Models and Fill-grade Factor Analysis
119–142. for Double-deep Multi-aisle AS/RSs.” International Journal
Sabattini, Lorenzo, Mika Aikio, Patric Beinschob, Markus of Production Research 54 (14): 4126–4144.
Boehning, Elena Cardarelli, Valerio Digani, and Annette Yang, Peng, Yongfei Peng, Bin Ye, and Lixin Miao. 2017. “Inte-
Krengel, et al. 2017. “The Pan-robots Project: Advanced grated Optimization of Location Assignment and Sequenc-
Automated Guided Vehicle Systems for Industrial Logistics.” ing in Multi-shuttle Automated Storage and Retrieval Sys-
IEEE Robotics & Automation Magazine 25 (1): 55–64. tems Under Modified 2 N-command Cycle Pattern.” Engi-
Scanlon, Jessie. 2009. “How Kiva Robots Help Zappos and neering Optimization 49 (9): 1604–1620.
Walgreens.” Retrieved April 10: 2010. Yu, Hu, and Yugang Yu. 2019. “Optimising Two Dwell Point
Schenone, M., G. Mangano, S. Grimaldi, and A. Cagliano. 2019. Policies for AS/RSs with Input and Output Point At Oppo-
“Estimating Travel Times in Dual Shuttle AS/RSs.: A Revised site Ends of the Aisle.” International Journal of Production
Approach.” International Journal of Industrial Engineering Research 57(21): 6615–6633.
Computations 10 (3): 405–420. Yuan, Zhe, and Yeming Yale Gong. 2017. “Bot-in-time Delivery
Shabanov, Vladimir, and Donat Ivanov. 2019. “Organization for Robotic Mobile Fulfillment Systems.” IEEE Transactions
of Information Exchange in Coalitions of Intelligent Mobile on Engineering Management 64 (1): 83–93.
Robots.” In 2019 International Conference on Industrial Engi- Zou, Bipan, Yeming Gong, Xianhao Xu, and Zhe Yuan. 2017.
neering, Applications and Manufacturing (ICIEAM), 1–5. “Assignment Rules in Robotic Mobile Fulfilment Systems
IEEE. for Online Retailers.” International Journal of Production
Singh, S. P., and M. K. Tiwari. 2002. “Intelligent Agent Frame- Research 55 (20): 6175–6192.
work to Determine the Optimal Conflict-free Path for An
Automated Guided Vehicles System.” International Journal
of Production Research 40 (16): 4195–4223. Appendix
Soukhal, Ameur, Ammar Oulamara, and Patrick Martineau. A.1 RMFS model with single customer class
2005. “Complexity of Flow Shop Scheduling Problems with
Transportation Constraints.” European Journal of Opera- In this section, we consider RMFS models with single customer
tional Research 161 (1): 32–41. class:
Tompkins, James A., John A. White, Yavuz A. Bozer, and
Jose Mario Azaña Tanchoco. 2010. Facilities Planning. USA: A.1.1 Single-picker system
Wiley. We use the method of finite source models to describe RMFS
Turban, Efraim, Jon Outland, David King, Jae Kyu Lee, Ting- model with single class. Let k denote the number of customers
Peng Liang, and Deborrah C. Turban. 2018. “Order Fulfill- in the system (queue plus service in picker’s workstation), λ
ment Along the Supply Chain in e-Commerce.” In Electronic denote the arrival rate in the system, μ denote the service rate
Commerce 2018, 501–534. Springer, Cham. in the system. We have the limiting probability pk ,
Umar, Umar Ali, M. K. A. Ariffin, N. Ismail, and S. H. k
Tang. 2015. “Hybrid Multiobjective Genetic Algorithms λ R!
pk = p0 , 0 < k ≤ R; 0, k > R, (A1)
for Integrated Dynamic Scheduling and Routing of Jobs μ (R − k)!
and Automated-guided Vehicle (AGV) in Flexible Manu-
facturing Systems (FMS) Environment.” The International
Journal of Advanced Manufacturing Technology 81 (9-12):
2123–2141.
Ventura, José A., and Chulung Lee. 2001. “A Study of the Tan-
dem Loop with Multiple Vehicles Configuration for Auto-
mated Guided Vehicle Systems.” Journal of Manufacturing
Systems 20 (3): 153–165. Figure A1. RMFS model with single class.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 15

Figure A2. RMFS model with multiple servers.

where (3) The rate from a state of (E, S, P) to a state of (E − 1, S, P +


 R k −1 1) is (R − S − P) λs . The flows from bottom to top repre-
 λ R!
p0 = . (A2) sent arrivals of a robot carrying an expedited order at the
μ (R − k)! picker (see black and solid line). This event will happen
k=0
only when E ≥ 1.
A.1.2 Multiple-pickers system (4) The flows from right to left represent the completion of
We use Markovian models to describe an RMFS model with one picker job with the rate of μ (see blue and dashed line).
single class and m pickers. Let k denote the number of cus- This includes two situations. If E + S + P > R, the event
tomers in the system (queueing plus service in picker’s work- will happen like event (4.2). It is impossible to see state
station). E + S + P < R. If E + S + P < R, then there are idle robots.
This leads to the following set of birth-death coefficients. However, idle robots will always take a standard order so

that E + S + P < R causes contradictory.
(R − k)λ, 0 ≤ k ≤ R − 1 (4.1) The picker finishes the process of a robot and no
λk =
0, otherwise expedited request is waiting for robots. This event

will happen when E + S + P = R and P ≥ 1. The
kμ, 0 ≤ k ≤ m
μk = state will become (E, S + 1, P − 1).
mμ, k ≥ m
(4.2) The picker finishes a pickup and expedited requests
For the region 0 ≤ k ≤ m − 1, we have are waiting for robots. This event will happen only
k when E + S + P > R and P > 0.
 λ(R − i)
k−1
λ
pk = p0 = p0 CRk , 0 ≤ k ≤ m − 1.
i=0
(i + 1)μ μ
A.3 Proof of Proposition 3.2
(A3)
For the region m ≤ k ≤ R, we have Proof: The first term is the expected time for an expedited
order to wait for an available robot. By the Little’s law, the wait-

m−1
λ(R − i)  λ(R − i)
k−1
ing time equals Le /λe , where Le is the average waiting line
pk = p0
(i + 1)μ i=m mμ length for expedited orders for being processed and is
i=0
k 
R 
R−S ∞

λ k!
= p0 CRk mm−k , m ≤ k ≤ R. (A4) Le = (E − (R − S − P))Pr(E, S, P). (A5)
μ m! S=0 P=1 E=R−S−P+1

The second term is the expected time for the expedited order
A.2 The transition diagram for RMFS with two to be processed by the picker.
classes and a picker The third term is the time when a robot travels from
We show the transition diagram in Figure A3 and Table A1, the picker station to the expedited order and then trav-
els back and carries the order to the picker station. Prpi =
which has the following four types of transitions. R ∞
S=0 E=R−S Pr(E, S, 0) is the probability when the pick is
(1) The rate from a state of (E, S, P) to a state of (E + 1, S, P) is idle. The average line length of robots waiting for the picker
λe . The flows from top to bottom represent arrivals of an can be obtained from
expedited shipping request (see red and dashed line). 
R 
R ∞

(2) The rate from a state of (E, S, P) to a state of (E, S − 1, P + LR = P Pr(E, S, P), (A6)
1) is Sλs . The flows from left to right represent arrivals of P=0 S=0 E=R−S−P
a robot carrying a standard order at the picker (see black R ∞
and solid line). This event will happen only when S ≥ 1. and P(P = p) = S=0 E=R−S−P Pr(E, S, P). 

Table A1. The transition rates and diagram for RMFS model with two classes.
Event Next state Transition rate Comment
(1) One expedited request arrives (E + 1, S, P) λe
(2) One robot carrying a regular order comes back to the picker (E, S − 1, P + 1) Sλs When S ≥ 1
(3) One robot carrying an expedited order comes back to the picker (E − 1, S, P + 1) (R − S − P)λs Only when E ≥ 1
(4.1) The picker finishes the process of a robot and there is no (E, S + 1, P − 1) μ Only When E+S+ P = R and P ≥ 1
expedited request waiting for robots
(4.2) The picker finishes a pickup and expedited requests waiting (E, S, P − 1) μ When E+S+ P > R and P > 0
for robots
16 Y. GONG ET AL.

Figure A3. A Markov chain model for RMFS with two classes.

A.4 Calculate equilibrium probability OF(e,s,p) = (θ(e+1,s,p) c(e,s,p) λe + θ(e,s−1,p+1) c(e,s,p) SλS
The state space is  = {S = 0, . . . , R; P = 0, . . . , R − S; E = + θ(e−1,s,p+1) c(e,s,p) (R − S − P)λS
R − S − P, . . . , ∞}. The number of states is N. Let C be the state
space which satisfies conditions in comments of Table A1. Let + θ( e, s + 1, p − 1)c(e,s,p) μ


+ θ(e,s,p−1) c(e,s,p) μ)π(e,s,p) . (A7)
1, (e, s, p) 1, (e, s, p)C
θ(e,s,p) = , c(e,s,p) = .
0, otherwise 0, otherwise

We define transition rate from state (e, s, p) to (e , s , p ) as


Proof: For the state (e, s, p), its outgoing flows may go to states
q((e,s,p),(e ,s ,p )) .
in outgoing destination set

O(e,s,p) = {(e + 1, s, p), (e, s − 1, p + 1), (e − 1, s, p + 1),


Proposition A.1: The value OF(e,s,p) of outgoing flows from a
state (e, s, p) is (e, s + 1, p − 1), (e, s, p1 )},
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 17


but the state may be not in state space . The value of outgoing for any state (e ,s ,p ) ∈ O(e,s,p) ,
flows is if θ( e , s , p ) = 1 and c( e , s , p ) = 1,

q((e,s,p),(e ,s ,p ); for (e ,s ,p )
=(e,s,p)) π(e,s,p) . (A8) A(x, y) = A(x, y) − q(e,s,p),(e ,s ,p ) ,
x = Findex(E, S, P, R)y = Findex(E, S, P, R)
s p e
end
Then we apply Table A1 to specify OF(e,s,p) and get the 2.2 Specify the incoming flow to state (e,s,p)

proposition.  for any state (e ,s ,p ) ∈ I(e,s,p) ,
if θ( e , s , p ) = 1c( e , s , p ) = 1,
Proposition A.2: The value IF(e,s,p) of incoming flows to a state A(x, y) = A(x, y) + q(e ,s ,p ),(e,s,p) ,
(e, s, p) is i = Findex(E, S, P, R), j = Findex(E , S , P , R)
end
IF(e,s,p) = θ(e−1,s,p) c(e−1,s,p) λe π(e−1,s,p)
end
+ θ(e,s+1,p−1) c(e,s+1,p−1) (S + 1)λS π (e, s + 1, p − 1) end
end
+ θ(e+1,s,p−1) c(e+1,s,p−1) (R − S − P + 1)λS
3. Compute equilibrium probability
+ θ(e,s−1,p+1) c(e,s−1,p+1) μ A
π= .
b
+ θ(e,s,p+1) c(e,s,p+1) μ (A9)
In Algorithm 1, we need to specify the total number of states.
Proof: For states (e, s, p), its incoming flows may be from Let the upper bound of E be U, we have Proposition A.3.
an incoming source set I(e,s,p) = {(e − 1, s, p), (e, s − 1, p +
1), (e + 1, s, p − 1), (e, s − 1, p + 1), (e, s, p + 1)}, but some Proposition A.3: The total number of states is N = 16 (1 +
states may be not in the status space . The value of incoming R)(2 + R)(3U − R + 3).
flows is
  
q((e ,s ,p ),(e,s,p); for (e ,s ,p )
=(e,s,p)) π(e ,s ,p ) . (A10) Proof: By state space, we have N = RS=0 R−S P=0 (U − R + S +
s p e
P + 1). By Mathematica, its value is 16 (1 + R)(2 + R)(3U −
R + 3). For example, if the upper bound of E is R, the num-
Then we apply Table A1 to specify IF(e,s,p) and get the  
ber of status is N = RS=0 R−S 1
P=0 (S + P + 1) = 6 (1 + R)(2 +
proposition.  R)(3 + 2R). 
The equilibrium equation of state (e, s, p) is IF(e,s,p) − In Algorithm 1, a critical step is to specify positions and
OF(e,s,p) = 0. Let π be an N-dimension vector of equilibrium indexes for a state (e, s, p) in matrix A. This is important for
probability, Q be N × N matrix, and O be an N-dimension calculating throughput. In our program, we use a function
vector of 0. The nth row of Q consists of coefficients of the equi- Findex(e, s, p, R) to compute the position in matrix A by Propo-
librium equation for corresponding state (e, s, p), and the nth sition A.4.
column of Q consists of coefficients of state (e, s, p) in different
equilibrium equations. We have πQ = 0. Proposition A.4: The corresponding row index of state (e, s, p)
Let 1 be an N-dimension
 vector
 of 1. The additional nor- in matrix A is
malisation condition is RS=0 R−S P=0

E=R−S−P π(E,S,P) = 1,
x = Findex(e, s, p, R)
or π = 1. Let A = (Q, 1), b = (0, 1), we have πA = b.
1
1
= 1 + e + 2p + (p − 1)p − R − pR + s + ps + pU
2
A.5 An algorithm to calculate equilibrium
1
probability of complex Markov systems − s(−7 + 3R2 − 3Rs + s2 − 9U − 6RU + 3SU)
6
In this appendix, we present the algorithm used to calculate (A11)
equilibrium probability of Markov systems.
Proof: Based on Algorithm 1, the corresponding row index of
Algorithm A.1: state (e, s, p) in matrix A is:

1. Initiatation x = Findex(e, s, p, R)
1.1 Compute the number of states N; p−1

s−1 
R−S 
1.2 Initiate the space of A by A = zeros(N + 1,N); = (U − R + S + P + 1) + (U − R + s + P + 1)
A(N + 1,:) = ones(1,N); S=0 P=0 P=0
1.3 Initiate the space of b by b = zeros(N + 1,1);
b(N + 1) = 1; + (e − (R − s − p) + 1) (A12)
1.4 Initiate the space D = zeros(N + 1,1); We can calculate the value of i by Mathematica and get
1.5 Input the value of parameters. Proposition A.4. If U = R, we have:
2. Specifying matrix A
for S = 0:R x = Findex(e, s, p, R)
for P = 0:(R − S) 
s−1 
R−S
for E = (R − S − P):U = (U − R + S + P + 1)
2.1 Specify the outgoing flow from state (e,s,p) S=0 P=0
18 Y. GONG ET AL.

Table A1. The transition rates and diagram for RMFS model with multiple pickers.
Event Next state Transition rate Comment
(1) One expedited request arrives (E + 1, S, P1 , . . . , Pn ) λe
(2) One robot carrying a regular order comes back to (E, S − 1, P1 , . . . , Pl + 1, . . . , Pn ) Sλs When S > 0 and l = arg mini Pi
the picker n 
(3) One robot carrying an expedited order comes back (E − 1, S, P1 , . . . , Pl + 1, . . . , Pn ) (R − S − i=1 Pi )λs Only when R − S − ni=1 Pi > 0 and
to the picker l = arg mini Pi 
(4.1) Pickers finish the process of a robot and there is (E, S + 1, P1 , . . . , Pi − 1, . . . , Pn ) μi Only When R − S − ni=1 Pi = E and
no expedited request waiting for robots Pi > 0 
(4.2) The picker finishes a pickup and there are (E, S, P1 , . . . , Pi − 1, . . . , Pn ) μi When R − S − ni=1 Pi < E and Pi > 0
expedited requests waiting for robots

p−1
 (2) The rate from a state of (E, S, P1 , . . . , Pn ) to a state of
+ (U − R + s + P + 1) + (e − (R − s − p) + 1) (E, S − 1, P1 , . . . , Pl + 1, . . . , Pn ) is Sλs . This event will
P=0 happen only when S > 0.
1 2 (3) The rate from a state of (E, S, P1 , . . . , Pn ) to a state of (E −
= [3p + p(3 + 6s) + s(7 + 9R + 3R2 − s2 )] 1, S, P1 , . . . , Pl + 1, . . . , Pn ) is R − S − ni=1 Pi λs , repre-
6
senting the arrival of a robot carrying an expedited order
+ (e − (R − s − p) + 1) (A13) at picker. This event will happen only when R − S −
the
n
We arrange the column associated with the state (e, s, p) with i=1 i > 0.
P
the same sequence in rows, and the state (e, s, p) is in yth column (4) The completion of job by the picker with a rate of μi . This
of matrix A and y = Findex(e, s, p, R).  includes the following two classes:
(4.1) The picker finishes the process of a robot and there
is no expedited request waiting for  robots. This event
A.6 The transition diagram for RMFS with two will happen only when R − S − ni=1 Pi = E and
classes and multiple pickers Pi > 0.
(4.2) The picker finishes a pickup and there are expedited
We explain the transition rate in Table A1 as follows, requests waiting for robots. This event will happen
only when R − S − ni=1 Pi < E and Pi > 0.
(1) The rate from a state of (E, S, P1 , . . . , Pn ) to a state of (E +
1, S, P1 , . . . , Pn ) is λe , caused by an expected order arrived.

View publication stats

You might also like