Sie sind auf Seite 1von 11

IEEE TRANSACTIONS ON ENGINEERING MANAGEMENT, VOL. 64, NO.

1, FEBRUARY 2017 83

Bot-In-Time Delivery for Robotic Mobile


Fulfillment Systems
Zhe Yuan and Yeming (Yale) Gong

Abstract—This paper studies a robotic mobile fulfillment sys- information wirelessly (e.g., with WiFi) [9] and use a camera to
tem (RMFS) featured by robots transporting movables shelves to read navigational barcode stickers on the warehouse floor [10].
order pickers. The RMFS can increase productivity, reduce costs, Robots then navigate the warehouse to find the appropriate
increase order picking accuracy, and improve operational flexibil-
ity. We build queue network models to describe the RMFS with two storage shelves containing the requested items, lift the storage
protocols in sharing robots for pickers, propose the corresponding shelves up by sliding under them, and bring them to picker
algorithms, conduct numerical analyses, and evaluate the perfor- workstations [11]. Pickers then select items and place them in
mance of the RMFS by calculating the throughput time. We then cartons. Completed orders in cartons are moved to a storage zone
calculate the optimal number and the velocity of robots, and pro- and, then, to the docks for shipping when trucks arrive. Forbes
vide the effective design rules for the RMFS.
has called this business operation “bot-in-time” delivery, where
Index Terms—Distribution center, facility design, mobile fulfill- “bot” means a robot [12]. In a typical conveyor-based operation,
ment system, order picking, queue networks, robotic warehouse, a picker can pick 200–400 items per hour, whereas warehousing
speedy delivery, warehouse management.
robots can present a new item to a picker every 6 s, leading to a
I. INTRODUCTION basic rate of 600 items per hour [6]. Amazon’s latest data show
that the company currently has deployed 13 distribution centers
NLINE retailers face the problem of designing oper-
O ational processes to achieve speedy and cost-effective
delivery [1]. Driven by leading online retailers such as Amazon,
with warehousing robots [13]. The RMFS can reduce order
fulfillment time in warehouses, allowing them to meet customer
expectations for speedy delivery. Admittedly, the RMFS can be
an increasing number of consumers expect same-day delivery. used only in reducing warehousing time, and online retailers
Facing this challenge, many competitors are improving the ve- need other solutions to reduce the transportation time.
locity and efficiency of delivery with new distribution facilities Automation and computer technology is rapidly developing,
[2], and warehouse automation has become a key solution [3]. wherein flexible manufacturing and intelligent logistics systems
Robotic mobile fulfillment systems (RMFSs) apply econom- are among the important solutions for current online retailers.
ically feasible robots [4], simple shelves, basic conveyers for Intelligent robot control is a major research topic with high
picker workstations, and centralized control softwares [5]. In value, and warehousing robots play a practical role as a branch
most of the traditional conveyor-based warehouses, items are of robotics. This paper is timely based on the challenging
moved using conveyors, chutes, and carousels, or by human background of warehousing robots. Our research findings and
operators with tow racks or forklifts [6]. A typical 10 000-m2 methods can be used in the operation management of robotic
warehouse employs 75 pickers on 8-h shifts to fulfill 200 000 warehouses, with practical management significance. The main
order picks per day [7]. In an RMFS, products are stored on research problem of this paper is as follows: What are the
light, simple, and portable storage shelves in a parallel-aisle optimal number and velocity of robots to minimize the total
storage zone. These storage shelves are not fixed to the floor and throughput time in an RMFS?
can be moved by robots replacing conveyor belts and humans The rest of this paper is organized as follows: In Section II,
[8]. When an order is released, a robot can receive the order we discuss the relevant literature. In Section III, based on open
queue networks, we develop models with both dedicated robots
Manuscript received March 9, 2016; revised July 27, 2016 and November 3, and pooled robots to evaluate the throughput time. Section IV
2016; accepted November 29, 2016. Date of publication January 18, 2017; date verifies the analytical models via simulation, determines the
of current version January 18, 2017. This work was supported by the China
Scholarship Council, by the National Natural Science Foundation of China optimal number and velocity of robots, and considers congestion
under Grant 71620107002, by the National Social Science Foundation under of the system. We make concluding remarks in Section V.
Grant 16ZDA013, and by the Chutian Scholarship. Review of this manuscript
was arranged by Department Editor T. Ravichandran. (Corresponding author:
Yeming (Yale) Gong.)
Z. Yuan is with the Laboratory of Engineering Industry, Centrale- II. LITERATURE REVIEW
Supélec, Université Paris Saclay, Châtenay-Malabry, 92290, France (e-mail:
zhe.yuan@centralesupelec.fr). Mobile order fulfillment has been applied in warehouse op-
Y. (Y.) Gong is with the EMLYON Business School, Écully 69134, France erations [14]. An RMFS employs robots and a parts-to-picker
(e-mail: gong@em-lyon.com). system with mobile communication and control. Researchers
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. have studied different optimal operation and design problems
Digital Object Identifier 10.1109/TEM.2016.2634540 of parts-to-picker systems, including travel-time models [15],

0018-9391 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
84 IEEE TRANSACTIONS ON ENGINEERING MANAGEMENT, VOL. 64, NO. 1, FEBRUARY 2017

request sequencing problems [16], space estimation models


[17], throughput optimization [18], [19], and storage assignment
problems [20]. Song et al. [21] measured the key performance
of an order fulfillment system for multiple items. The optimiza-
tion of an individual warehousing robot includes battery man-
agement and time management [22]. As pioneer researchers in
mobile fulfillment systems, Wurman et al. [23] studied the co-
ordination of hundreds of cooperative autonomous vehicles in
warehouses. Enright and Wurman [24] solved the problem of
optimization and coordinated autonomy in a mobile fulfillment
system. Chen and Dong [25] formulated two order fulfillment
models considering customer priority in order to obtain reserv-
ing production capacity and reserving components for each cus-
tomer class. In order to achieve speedy delivery, we need to
consider the throughput time of the RMFS. There have been
literature reviews on travel time and waiting time in order ful-
fillment systems. Higa et al. [26] considered order waiting time Fig. 1. Layout of an RMFS.
in a continuous review base-stock system with independent and
identically distributed replenishment lead times for each stock W width of a robotic warehouse;
unit. Wensing and Kuhn [27] analyzed expected waiting times L length of a robotic warehouse;
to evaluate the performance of an inventory system. Pan et al. T r throughput time;
[28] formulated an order picking model using queue networks λr arrival rate of robots;
considering waiting time, traveling time, and picking time in or- μr service rate of robots;
der to minimize average order fulfillment time in a multipicker μP service rate of pickers.
system. Gong and De Koster [29] studied a mobile order fulfill- The RMFS is abstracted as a rasterized image, with each
ment system to achieve a time-based competitive advantage for square block corresponding to the grid position coordinates
online retailers. (X, Y ). Triangles represent pickers. Circles without squares
One of the research topics relevant to the RMFS is routing represent walking warehousing robots, which are not carrying
design [30], and the objective of routing research is to minimize shelves. Circles with solid-line squares are warehousing robots,
the throughput time of a warehouse. Lee et al. [31] proposed the which have mounted a shelf. Dotted-line squares are empty
heuristic algorithm of sequential path generation and generated storage locations, and solid-line squares are with shelves. In or-
the shortest collision-free path for multiple agents in semiauto- der to facilitate the calculation, we simplify the initial location
mated warehouses. Automated guided vehicles (AGVs) in the coordinates to (0, 0). We consider two protocols: 1) R robots
automotive industry and the manufacturing industry are more are served by one order picker; and 2) R robots are served by
widely applied in long-distance path planning [32]. Qiu et al. multiple-order pickers. We assume that velocities of all robots
[33] focused on efficient algorithms for scheduling and rout- are the same. All moving robots in the RMFS are in good work-
ing problems of AGV systems. They classified algorithms into ing condition, without failures.
algorithms for general path topologies, path layout optimiza-
tion, and specific path topology. The performance of the RMFS A. Dedicated Robots
is outstanding in fixed routing. Some literature simultaneously
addresses various design problems of AGV systems by using In this section, we consider a situation where R robots are
simulation [34] and developing analytical models, which can served by one order picker. The system is a centralized system
be applied for AGV systems [35]. Hu and Egbelu [36] studied and the order arrival process to robots follows a Poisson process.
different scheduling strategies for existing idle AGVs when ser- We assume that the system is continuous and the storage position
vice requests arrive. Few researchers have studied the design follows uniform distribution. After receiving an order request, a
and application of the RMFS, and the focus of this paper is on robot originally staying in the order picker workstation follows
the minimal throughput time with mathematical analysis and the protocol.
simulation for the RMFS. 1) Leave the order picker station.
2) Travel to the storage grid to pick a shelf which includes
III. MODEL AND ANALYSIS requested item(s).
We consider the following layout shown in Fig. 1, which is 3) Return to a picker station which is serving.
based on the layout of Zappos’s robotic order fulfillment systems 4) Return the shelf to its original storage position after the
[37]. We use the following notations: worker picks the item.
X abscissa of a robot’s position; 5) Return to the picker station.
Y ordinate of a robot’s position; We have the following lemma.
R number of robots; Lemma 1: If robot r is served by the leftmost or rightmost
N number of pickers; picker, the RMFS with a single picker can be described as an
v velocity of robots; M/G/1 queue.
YUAN AND GONG: BOT-IN-TIME DELIVERY FOR ROBOTIC MOBILE FULFILLMENT SYSTEMS 85

Proof: The arrival process follows a Poisson process. We Equation (6) can be used to design the velocity of robots, and
mainly need to show that the interservice time follows uniform we have the following design rule.
distribution. We assumed that the beginning position of robot r Design rule 1: There exists a minimal acceptable velocity of
is the leftmost picker’s position. Note that a robot receives the robots such that expected waiting time at robots is less than a
order information wirelessly, uses a camera to read navigational ωr .
barcode stickers on the warehouse floor, and travels in straight Proof: Since EWr (v) is a decreasing function of v, let the
lines. We suppose that a robot receives an order with a storage upper bound of EWr , a measure for the service level, at robots
position (Xr , Yr ). Let X ∼ U [0, W ], and the density function be ωr . The minimal acceptable velocity of robots v is given by
of X is f (x)
 1 7λr (W + L)2 W +L
0≤x≤W λr (W +L )
+ = ωr . (7)
W , 12v 2 (1 − ) v
f (x) = (1) v
0, otherwise.

The probability function of X is F (x) We calculate the squared coefficient of variation of an inter-
⎧ x arrival time of robots c2aR = 1, ρ = λr (Wv +L ) , the squared co-
⎨W, 0≤x≤W
⎪ efficient of variation of an interservice time of robots c2sR = 13 ,
F (x) = 1, x>W (2) μ = W v+L , and EWr from (6). Then, based on the equation


0, x < 0. c2dR = c2aR + 2ρ2 c2sR − 2ρ(1 − ρ)μEW from [39, eq. (37), p.
2799], we obtain the squared coefficient of variation of an in-
Let Y ∼ U [0, L], and the density function of Y is g(y) terdeparture time

1
, 0≤y≤L  
g(y) = L 2 2λ2r (W + L)2 λr (W + L)
(3) cdR = 1 + − 2λr 1 −
0, otherwise. 3v 2 v

The probability function of Y is G(y) 7λr (W + L)2 W +L


⎧y λ (W +L )
+ . (8)
12v 2 (1 − r
) v
⎨ L, 0 ≤ y ≤ L
⎪ v

G(y) = 1, y > L (4) 2) Superposition of departure streams from robots: The de-


0, y < 0. parture processes are merged into one process, which is also
the process of arrival to the picker. We use methods from [39,
Let Z = X + Y , and its density function is h(z), as shown in pp. 2794–2797] and have
Fig. 12 in Appendix A 
λaP = ΣR
1 λr . (9)
⎧ z

⎪ WL, 0≤z≤L With the asymptotic method, the squared coefficient of vari-


⎨ 1, L<z<W ation of an interarrival time of the picker is
h(z) = W (5)


W +L −z
W ≤z ≤W +L 2λ2 (W + L)2
⎪ WL , λr

⎩ 2
caP = Σr =1 R
R
1+ r
0, otherwise. Σr =1 λr 3v 2
 
So, EZ = W 2+L is the expectation of Z. We denote ES as the λr (W + L)
− 2λr 1 −
expected service time of robots, and ES = 2Ev Z = W v+L . v
R robots will carry shelves to the picker. R departure pro- ⎡ ⎤
2
cesses are merged into an arrival process to the picker. The ⎣ 7λ (W + L) W + L ⎦ .
r + (10)
2 λr (W +L ) v
interservice time of the picker follows exponential distribution. 12v 1 − v
As such, the picker service system can be described as a G/M/1
system. We describe the system with Fig. 13 in Appendix B. 3) Analysis of pickers: The expected delay time in a picker
1) Analysis of robots: We use ED to denote the expected is given by Whitt [39, p. 2802]:
delay time and ES to denote the expected service time. The
expected waiting time EW = ED + ES. From [38, eq. (154), ρ(c2aP + c2sP )
(S 2 ) λE (S 2 )
EDP = g(ρ, c2aP , c2sP ). (11)
p. 280], we have EW = λE 1
2(1−ρ) + μ , where ED = 2(1−ρ) is
2μ(1 − ρ)
the delay time and μ1 is the service time. Considering robot r 2(1−ρ)(1−c 2a P ) 2 )
7(W +L ) 2
Here, g(ρ, c2aP , c2sP ) is defined as exp[− 3ρ(c 2a P +c 2s P )
]
2
as a server in an M/G/1 queue, we have E(S ) = ,
λr (W +L ) 2
6v 2 when c2aP < 1, and 1 when c2aP ≥ 1. We 2
substitute csP = 1
μ = W +L , and ρ =
v
. Substituting E(S ), μ, and ρ into and λaP = ΣR
v 1 λr into EDP and obtain
EW, we obtain the expected waiting time at the robot r
2
1 λr (caP + 1)
ΣR
EWr = 7λr (W +L ) 2
+ W +L EDP = g(ρ, c2aP , c2sP ). (12)
v . (6) 2μP (μP − ΣR1 λr )
λ (W + L )
12v 2 (1− r v )
86 IEEE TRANSACTIONS ON ENGINEERING MANAGEMENT, VOL. 64, NO. 1, FEBRUARY 2017

We consider the mean number of jobs in queue Lq : Lemma 3: EDP (R) is a nonlinear function of R. When R ≥
3λ(W +L )
(ΣR 2 2 EDP (R) is an increasing function of R.
,
1 λr ) (caP + 1) 2v
Lq = g(ρ, c2aP , c2sP ). (13) Proof: The expected delay time at picker workstations is
2μP (μP − ΣR
1 λr )
This is also the number of robots that are carrying shelves λ(c2aP + 1)
and waiting before the picker. In this system, there is an upper EDP (R) = g(ρ, c2aP , c2sP ). (18)
2μP (μP − λ)
bound ι on Lq . We show that we can find a design rule for the
velocity of robots.
Design rule 2: There exists a minimal acceptable velocity of The monotony of EDP (R) is only related to c2aP (R). We
2
robots such that the number of robots who are carrying shelves r =1 λr = λ into caP (R) and obtain
substitute ΣR
and waiting before the picker Lq is less than ι from space con-
straints. 3λ2 (W + L)2 2λ(W + L)
c2aP (R) = 1 + −
Proof: c2aP is a decreasing function of v. g(ρ, c2aR , c2sR ) is a 2R2 v 2 Rv
decreasing function of v when c2aR < 1, and 1 when c2aR ≥ 1.  2
1 3 λ(W + L) 2
So, Lq is a decreasing function of v. By solving the following = + − . (19)
3 2 Rv 3
equation:
2 2
1 λr ) (caP (v 2 ) + 1)
(ΣR As R increases, the function EDP (R) is nonlinear. It first de-
Lq (v 2 ) = g(ρ, c2aP (v 2 ), c2sP )
2μP (μP − ΣR 1 rλ ) creases and then increases after reaching a critical point. When
=ι (14) 0 < R ≤ 3λ(W2v+L ) , the function EDP (R) decreases. When
R ≥ 3λ(W2v+L ) , EDP (R) increases. From (16), we derive the
we can get the minimal acceptable velocity of robots, v 2 . optimal robot number. 
 3λ(W +L )
Lemma 4: When 0 < R ≤ 2v , T r(R) is a decreasing
4) Analysis of system stability: For robots, λr < μr =
function, and we have an R limited by integer constraints. When
W +L , r = 1, . . . , R. For picker workstations, Σ1 λr < μP .
v R
R ≥ 3λ(W2v+L ) , T r(R) is a convex function.
5) Throughput time: We estimate the average customer
throughput time of orders, which is a sum of components: Proof: Let ∂ T∂rR(R ) = 0. We can obtain R0 from the following
the delay time EDr that an order spends in the queue wait- equation:
ing to be carried by robots, the average service time μ1r of
 2
a robot, the delay time EDP that an order carried by robots 6λ2 λ(W + L)
R02 = 2 2
g(ρ, caR , csR ) R0 −
spends in the queue waiting to be handled at picker worksta- μP (μP − λ) v
tions, and the expected picking time μ1P by order pickers. From  
v 1
T r = EDr + μ1r + EDP + μ1P , we have − . (20)
2λ(W + L) R0
7λr (W + L)2 W +L 1
Tr = + +
12v 2 (1 − λr (W +L )
v ) v μr If R > 3λ(W2v+L ) , the optimal number of robots R∗ = R0 . If
ΣR 2 R ≤ 3λ(W2v+L ) , the optimal number of robots R∗ = 3λ(W2v+L ) .
1 λr (caP + 1) 1
+ g(ρ, c2aP , c2sP ) + . (15)
2μP (μP − Σ1 λr )
R μP
6) The optimal number of robots: We assume that the work B. Pooled Robots
is equally allocated to R robots, the arrival rate of the Rth In this section, we consider a situation in which robots
robot is λr = λ/R, λr are equal, and ΣR
1 λr = λ. We obtain the can be served by multiple-order pickers. After receiving
function of T r(R) an order request, a robot follows the protocol proposed in
7 Rλ (W + L)2 1 Section III-A.
T r(R) = λ(W +L )
+ The system is centralized, and the order arrival process to
12v 2 (1 − Rv ) μr
robots follows a Poisson process with a rate of λ. Orders are
λ(c2aP + 1) 1 assigned to available robots following the rule of F CF S. Af-
+ g(ρ, c2aP , c2sP ) + . (16) ter orders arrive, each robot, originally staying with the or-
2μP (μP − λ) μP
der picker, leaves the picker and travels to storage positions to
Lemma 2: EDr (R) is a decreasing function of R. pick one shelf, which can provide items to that order. Then,
Proof: We can observe it from the robot returns to one of the pickers. The service rate of the
7 Rλ (W + L)2 picker processing an order is μp , and the service time follows
EDr (R) = λ(W +L ) a general distribution. The picker handles robots with the rule
12v 2 (1 − Rv )
of F CF S.
7λ(W + L)2 Open queue networks consist of an M/G/R queue and an
=  . (17) M/G/N queue. Two queuing systems are connected in the
12v 2 R − λ(Wv+L )
queue networks.
YUAN AND GONG: BOT-IN-TIME DELIVERY FOR ROBOTIC MOBILE FULFILLMENT SYSTEMS 87

1) Analysis of robots: The expected delay time in a robot is system, and Ss /Sr  is the number of robots that can be allowed
to stay in front of each picker station.
EDr = φ(ρ, c2aR , c2sR , R)
Considering the feasible space, we have Lq (β) ≤ Ss /Sr .
 
c2aR + c2sR From (25), we know that R and N are independent variables of
EDr (M/M/R). (21) Lq (c2aP includes R, and EDP (M/M/N ) includes N ), and the
2
optimization system is
φ(ρ, c2aR , c2sR , R) is given by Whitt [40, eq. (2.25)].
min Lq (β)
EDr (M/M/R) = μ R R R1(1−ρ) [ R(R ρ) R
!(1−ρ) ]ζ, and ζ is given by
Whitt [40, eq. (2.4)]. s.t. Lq (β) ≤ Ss /Sr . (26)
The departure process is also the arrival process to the picker.
We use methods from [39, p. 2799], and the squared coefficient We derive an optimal ratio of robots and pickers as follows:
of variation of an interarrival time of the picker is β1 = arg min{min Lq (β)|Lq (β) ≤ Ss /Sr }. (27)
2
ρ
c2aP = c2dR = 1 + (1 − ρ2 )(c2aR − 1) + √ (c2sR − 1) Considering the service level of the system, we can build the
R following optimization system:
ρ2 min T r(β)
= 1 + √ (c2sR − 1). (22)
R
s.t. Lq (β) ≤ ι. (28)
2) Analysis of pickers: The expected delay time in a
picker is We derive an optimal ratio of robots and pickers as follows:
2  β2 = arg min{min T r(β)|Lq (β) ≤ ι}.
2 2 caP + c2sP (29)
EDP = φ(ρ, caP , csP , N ) EDP (M/M/N ) (23)
2 When the number of pickers is the same, β is relevant to the
where c2aP
is given by (22), c2sP = 1, ρ = λ/(N μP ), investment cost of the robots. We can choose the minimal β
to ensure the minimum investment cost based on the satisfied
EDP (M/M/N ) = μ P N N1 (1−ρ) [ N(N!(1−ρ)
ρ) N
]ζ, and ζ is given by
service level. Considering the tradeoff between the investment
Whitt [40, eq. (2.4)].
cost of the robots and the service level in the feasible set, we
3) Analysis of system stability: For robots, λ < Rμr . For
can obtain an optimal ratio of robots and pickers
picker workstations, λ < N μP .
4) Throughput time: As in Section III-B, we have T r = β ∗ = min[β1 , β2 ]. (30)
EDr + μ1r + EDP + μ1P , where EDr is given by (21) and EDP
is given by (23) IV. NUMERICAL EXPERIMENTS
 2 
2 2 caR + c2sR Section IV-A calculates the basic results and shows the
T r = φ(ρ, caR , csR , R)
2 change of EDr , EDP , and T r with R. Section IV-B veri-
    fies the analytical models via simulation and compares sim-
1 (Rρ)R 1
ζ + ulative and analytic results. Section IV-C jointly determines
μR R R(1 − ρ) R!(1 − ρ) μr R the number and velocity of robots without considering con-
 
 2 2
 c2aP + c2sP gestion. Section IV-D considers system congestion and com-
+ φ ρ, caP , csP , N pares the systems of pooled robots and dedicated robots with
2
    congestion.
1 (N ρ)N 1
ζ + . (24)
μP N N (1 − ρ) N !(1 − ρ) μP A. Basic Calculation Results
5) Analysis of system congestion: We consider the mean We use MATLAB R2014a to verify the analytic RMFS model
number of jobs in queue Lq in Section III-A, denoting μ as the capacity of a picker. We con-
  sider a 50-m-long and 40-m-wide warehouse with an arrival rate
 2 2
 c2aP + c2sP
Lq = λφ ρ, caP , csP , N EDP (M/M/N). (25) of 0.4/min, and the velocity of robots is 10 m/min. The change
2 of robots’ expected queuing delay time EDr (R), expected pick-
This is also the number of robots that are carrying shelves ing delay time EDP (R), and throughput time T r(R) is shown
and waiting before the pickers. In this system, the congestion in Fig. 2.
near picker workstations can be serious, since the space in front Fig. 2 shows that EDr (R) is a decreasing function of R
of each picker station is limited. Therefore, there is an upper and EDP (R) is a nonlinear function of R. Since the number of
bound ι on Lq . robots R is an integer, when v = 10 m/min, 0 < R ≤ 5, T r(R) is
We define Ss as the area in front of a picker station, and Sr a decreasing function. When R ≥ 5, T r(R) is a convex function.
as the occupied area of a robot. Considering the congestion, The optimal number of robots is seven, and the result is in line
we optimize the ratio of robots and pickers as β = R/N . The with the proved conclusion R∗ = 3λ(W2v+L ) . These results are
throughput time T r is used to measure the service level of the consistent with Lemmas 2–4.
88 IEEE TRANSACTIONS ON ENGINEERING MANAGEMENT, VOL. 64, NO. 1, FEBRUARY 2017

Fig. 2. Change of ED r , ED P , and T r with R.

Fig. 4. Comparison of simulative and analytic results for the pooled model.

the throughput time obtained by the simulation, j is the index of


a simulation run, and M1 is the total number of simulation runs.
The largest and average bias between analytical and simulation
values are 7.41% and 5.54% based on 400 groups of experi-
ments. j is the scenario index featured by difference values of
λ and R.
Fig. 4 presents analytical and simulative results of the pooled
robots model Awith various values of v. We defined the bias as
1
M |T r j 2 −T r jS 2 |
M2 j =1 S2 × 100%, where T rjA 2 is the throughput
T rj
time obtained by the analytical models, T rjS 2 is the through-
put time obtained by the simulation, j is the scenario index,
featured by difference values of λ, R, and N , and M2 is
the total number of simulation runs. The largest and average
bias between analytical and simulation values are 6.67% and
Fig. 3. Comparison of simulative and analytic results for the dedicated model. 4.95% based on 1600 groups of experiments with different
parameters.
In two results, while simulative results are closer to reality,
the analytic results are based on the approximation methods
B. Verification of Analytic Results Via Simulation from [39] and [40]. This is a reason leading to the bias.
Arena 14.0 is used to simulate the single-service and multi-
service queuing systems. The purpose of the simulation experi- C. Optimal Decision
ment is to examine the matching degree between the analytical
model and the simulative system. We build simulation models An important problem for the RMFS design is to determine
in Arena 14.0 to verify the analysis model (see Appendix E for the optimal number and velocity of robots. This section shows
simulation processes of the RMFS with dedicated robots). We how to apply mobile picking systems in practice, especially to
simulate the RMFS with dedicated robots and pooled robots optimize the number and velocity of robots.
based on assumptions presented in Sections III-A and III-B for 1) Determine the Number of Robots: We consider a ware-
1000 warehousing operational hours. The average simulation house with λ = 0.4, width W = 40, and aisle length L = 50.
running time is 17.64 s per run in a computer with an Intel Core When v = 10 m/min, by solving (16), we obtain the optimal
i7-4790S CPU at 3.20 GHz, and 8-GB RAM. number of robots R∗ = 2λ(Wv +L ) = 7. Similarly, by solving
Fig. 3 presents the analytical and simulative results of the (16) and deleting values that do not satisfy the integer con-
dedicated robots model with different values of v. We defined straint, we obtain the optimal number of robots R∗ = 4. When
1
M |T r jA 1 −T r jS 1 | v = 30 m/min, the optimal number of robots R∗ = 2. When
the bias as M 1 j =1 × 100%, where T rjA 1 is the
v = 40 m/min, the optimal number of robots R∗ = 1. When
S1
T rj
throughput time obtained using the analytical models, T rjS 1 is v = 50 m/min, the optimal number of robots R∗ = 1.
YUAN AND GONG: BOT-IN-TIME DELIVERY FOR ROBOTIC MOBILE FULFILLMENT SYSTEMS 89

Fig. 7. Throughput time with the change of v and the change of R.


Fig. 5. Throughput time with the change of R.

Fig. 8. Throughput time with the change of N and the change of R.


Fig. 6. Throughput time with the change of v.

capacity, aisle width, and other constraints, the velocity of robots


Fig. 5 shows that with an increased number of robots, the has an upper bound.
throughput time increases nonlinearly. Since R is an integer, 3) Jointly Determine the Velocity of Robots: Fig. 7 shows a
when v = 10 m/min, the number of robots is less than seven. 3-D figure of the model of dedicated robots and shows that the
The throughput time decreases and then begins to increase. throughput time changes with v and R. The minimal value of
When v = 20 m/min, the number of robots is less than four. The the throughput time is 11.9 min when R = 2 and v = 50 m/min.
throughput time decreases first and then increases. When v = From Fig. 7, for the influence on throughput time, the velocity
30 m/min, the number of robots is less than two. The throughput of robots is greater than the number of robots.
time decreases and then increases. When v = 40 m/min, the Fig. 8 shows a 3-D figure of the model of pooled robots and
number of robots is less than one. The throughput time decreases shows that the throughput time changes with N and the change
and then increases. When v = 50 m/min, the number of robots is of R. When the number of robots and pickers increases, the
less than one. The throughput time decreases and then increases. throughput time decreases.
When the number of robots is more than 30, the throughput time
does not change much.
D. Considering System Congestion
2) Determine the Velocity of Robots: Without considering
aisle congestion, Fig. 6 shows that as the velocity of robots This section takes system congestion into consideration
increases, the throughput time decreases. Because of technical and shows how to optimize the ratio of robots and pickers.
90 IEEE TRANSACTIONS ON ENGINEERING MANAGEMENT, VOL. 64, NO. 1, FEBRUARY 2017

Fig. 9. L q with the change of β considering the feasible area.

Fig. 10. Comparison of the two models with congestion.

Section IV-D2 compares pooled robots and dedicated robots


with congestion.
1) Considering the Feasible Set: The size of a robot is ap-
proximately 2 ft by 2.5 ft [5] and the area in front of each picker
station is approximately 29 ft2 . We set Ss = 29 ft2 , Sr = 5 ft2 ,
so Ss /Sr  = 5. We use MATLAB R2014a to calculate the two
analytic models with warehouse parameters W = 50, L = 40,
λ = 0.4, and v = 20 m/min. Fig. 9 shows that with congestion of
the system, considering the maximal number of robots that can
be allowed to stay in front of picker station, the optimal ratio of
robots and pickers is β ∗ = 6.
2) Comparison of Pooled Robots and Dedicated Robots With
Congestion: We compare the pooled robot model with the ded-
icated robot model considering the congestion in front of each
picker station, as presented in Sections III-A and III-B.
Fig. 10 shows the trend of the throughput time of the model
with congestion in front of each picker station, as the ratio of
robots and pickers β increases. The throughput time T r(β) of
Fig. 11. Sensitivity analysis of v.
dedicated robots is a decreasing function, as the ratio of robots
and pickers β increases. The throughput time of pooled robots
T r(β) first decreases and then increases as the ratio of robots and
model changing v from 15 to 35 at a step size of 5 and β from
pickers β increases considering the congestion in front of each
2 to 18 at a step size of 4.
picker station. For example, if N = 3 and R = 3, μP = 0.5; this
Only considering the velocity of robots v, Fig. 11 shows that
satisfies the stability condition of picker workstations λ < N μP .
with the increase of v, the throughput time decreases, but the
But μr = W v+L = 0.11 does not satisfy the stability condition of
decline rate of the throughput time decreases. Only considering
robots λ < Rμr . When 1 ≤ β ≤ 11, the throughput time of the
the ratio of robots and pickers, we can see that the throughput
RMFS model with pooled robots is less than the throughput time
time first decreases. When β is around 6, we get the minimal
of the RMFS model with dedicated robots. When 11 ≤ β ≤ 20,
throughput time. After that, the throughput time increases.
the throughput time of the RMFS model with pooled robots
is more than the throughput time of the RMFS model with
V. CONCLUDING REMARKS
“dedicated robots” because of the congestion.
3) Sensitivity Analysis of the Velocity of Robots: Based on In this paper, we establish the model of dedicated robots
the congestion of the RMFS model with pooled robots, we with single-service open queuing networks and the model of
conduct a sensitivity analysis to observe the influence of the pooled robots with multiple-service open queuing networks for
velocity of robots v on the throughput time of the pooled robot the RMFS. The optimal number and velocity of warehouse
YUAN AND GONG: BOT-IN-TIME DELIVERY FOR ROBOTIC MOBILE FULFILLMENT SYSTEMS 91

Fig. 12. Density function of the travel distance Z .

robots are calculated in order to achieve effective warehous-


ing operations in terms of the total throughput time. We verify
the analytical results of the RMFS and simulate the open queue
networks systems. Numerical experiments for optimal decisions Fig. 13. Queue network model for an RMFS with “dedicated robots.”
are shown first with dedicated robots and pooled robots without
considering paths blocking. Then, we obtain the optimal ratio
of robots and pickers considering the feasible area and the con-
gestion of systems. Our research provides design support for
bot-in-time delivery and fast response for online retailers. Our
research findings and methods can be used in the facility design
and in the planning of robotic warehouses, with managerial rel-
evance particularly to online retailers handling the “last mile”
problem.
Our research has two differences with real systems. First,
our analytic model is based on Whitt approximations, which
will lead to relative errors. Second, we have not considered bat-
tery management in the RMFS. Future research might evaluate
battery charging and swapping strategies in the RMFS. Fig. 14. A queue networks model for RMFS with “pooled robots”.

APPENDIX A
DENSITY FUNCTION OF THE TRAVEL DISTANCE A. The Expected Delay Time in a Robot
In this appendix, Fig. 12 presents the density function of the The expected delay time in a robot EDr = φ(ρ, c2aR , c2sR , R)
travel distance Z. c 2 +c 2
( aR 2 sR)EDr (M/M/R),
1 (R ρ) R
APPENDIX B where EDr (M/M/R) = μR (1−ρ) [ R !(1−ρ) ]ζr and ζr =
 R −1
k = 1 (R ρ)
[ R(R
k
ρ) R −1
QUEUE NETWORK MODEL FOR THE RMFS WITH “DEDICATED +
!(1−ρ) k! ] .
ROBOTS” 2 2
φ(ρ, caR , csR , R) is defined as follows:
4(c 2 −c 2 )
In this appendix, Fig. 13 presents a queue networks model for When c2aR ≥ c2sR , φ(ρ, c2aR , c2sR , R) = ( 4c 2a R−3cs2R )φ1
aR sR
the RMFS with “dedicated robots.” c2 R c 2 +c 2
(R, ρ) + ( 4c 2 s−3c 2 )ψ( a R 2 s R , R, ρ).
aR sR
c 2 −c 2
APPENDIX C When c2aR < c2sR , φ(ρ, c2aR , c2sR , R) = ( 2c 2s R +2ca R2 )φ3
aR sR
QUEUE NETWORK MODEL FOR THE RMFS WITH “POOLED c 2 +3c 2a R c 2a R +c 2s R
(R, ρ) + ( 2cs2R 2 )ψ( 2 , R, ρ) .
ROBOTS” a R +2c s R
2 2
Where caR + csR
= 1, φ1 (R, ρ) = 1 + γ(R, ρ), φ2 (R, ρ) =
In this appendix, Fig. 14 presents a queue networks model for
1 − 4γ(R, ρ), φ3 (R, ρ) = φ2 (R, ρ) exp{ −2(1−ρ)
3ρ }, φ4 (R, ρ) =
the RMFS with “pooled robots.”
φ 3 (R ,ρ)
min{1, φ1 (R, ρ) + 2 }, and γ(R, ρ) = min{0.24, (1 −
1
APPENDIX D (4+5R ) −2 2
ρ)(R − 1) 16R ρ }.
EXPECTED DELAY TIME OF MODELS WITH “POOLED ROBOTS”
We consider the situation where robots are served by multiple- B. The Expected Delay Time in a Picker
order pickers. In this model, we consider the expected delay time
in a robot and the expected delay time in a picker. We use EDr The squared coefficient of variation of an interarrival
and EDP to denote the expected delay time in a robot and the time of the picker is c2aP = c2dR = 1 + (1 − ρ2 )(c2aR − 1) +
ρ2 2
expected delay time in a picker. √ (c2 − 1) = 1 + √ρR (c2sR − 1).
R sR
92 IEEE TRANSACTIONS ON ENGINEERING MANAGEMENT, VOL. 64, NO. 1, FEBRUARY 2017

APPENDIX E SIMULATION MODEL OF THE RMFS WITH


“DEDICATED ROBOTS”
We take RMFS with “dedicated robots” for an example, and
build a simulation model of RMFS with “dedicated robots” to
simulate the queue networks.

ACKNOWLEDGMENT
The authors would like to thank Prof. M. Jin for his useful
comments.

REFERENCES
[1] H. L. Lee and S. Whang, “Winning the last mile of e-commerce,” MIT
Sloan Manage. Rev., vol. 42, no. 4, p. 54, 2001.
[2] J. Joong-Kun Cho, J. Ozment, and H. Sink, “Logistics capability, logistics
outsourcing and firm performance in an e-commerce market,” Int. J. Phys.
Distrib. Log. Manage., vol. 38, no. 5, pp. 336–359, 2008.
[3] P. Baker and Z. Halim, “An exploration of warehouse automation im-
plementations: Cost, service and flexibility issues,” Int. J. Supply Chain
Manage., vol. 12, no. 2, pp. 129–138, 2007.
[4] G. C. Rapp, “Torts 2.0: The restatement 3rd and the architecture of partic-
ipation in american tort law,” William Mitchell Law Rev., vol. 37, p. 1582,
2010.
[5] V. Kopytoff, “In warehouses, kiva robots do the heavy lifting,” MIT
Technol. Rev., [Online]. Available: https://www.technologyreview.com/
s/428436/in-warehouses-kivas-robots-do-the-heavy-lifting/. Accessed on:
Jul. 20, 2012.
[6] E. Guizzo, “Three engineers, hundreds of robots, one warehouse,” IEEE
Spectrum, vol. 45, no. 7, pp. 26–34, Jul. 2008.
[7] R. D’Andrea, “Guest editorial: A revolution in the warehouse: A retro-
spective on kiva systems and the grand challenges ahead,” IEEE Trans.
Autom. Sci. Eng., vol. 4, no. 9, pp. 638–639, Oct. 2012.
[8] R. De Koster, T. Le-Duc, and K. J. Roodbergen, “Design and control of
warehouse order picking: A literature review,” Eur. J. Oper. Res., vol. 182,
no. 2, pp. 481–501, 2007.
[9] R. L. Nussbaum, “Changing the tooth-to-tail ratio using robotics and
automation to beat sequestration,” Air Space Power J., vol. 29, no. 5,
pp. 75–84, 2015.
[10] J. Scanlon, “How kiva robots help Zappos and Walgreens,” Bus. Week,
vol. 10, p. 2010, 2009.
[11] G. Q. Huang, M. Z. Chen, and J. Pan, “Robotics in ecommerce logistics,”
HKIE Trans., vol. 22, no. 2, pp. 68–77, 2015.
[12] C. Steiner, “Bot-in-time delivery,” Forbes Mag., vol. 3, no. 16, p. 40, 2009.
[13] [Online]. Available: https://www.amazonrobotics.com. Accessed on: Jun.
17, 2016.
[14] P. Baker and M. Canessa, “Warehouse design: A structured approach,”
Eur. J. Oper. Res., vol. 193, no. 2, pp. 425–436, 2009.
[15] J. A. Tompkins, J. A. White, Y. A. Bozer, and J. M. A. Tanchoco, Facilities
Fig. 15. The simulation model of RMFS with “dedicated robots”. Planning. Hoboken, NJ, USA: Wiley, 2010.
[16] S. C. Graves, W. H. Hausman, and L. B. Schwarz, “Storage-retrieval
The expected delay time in a picker is EDP = interleaving in automatic warehousing systems,” Manage. Sci., vol. 23,
2 2
φ(ρ, c2aP , c2sP , N )( c aP +c
2
sP
)EDP (M/M/N ), no. 9, pp. 935–945, 1977.
(N ρ) N
[17] F. Eldemir, R. Graves, and C. Malmborg, “New cycle time and space
1
where EDP (M/M/N ) = μp(1−ρ) [ N !(1−ρ) ]ζp and ζP = estimation models for automated storage and retrieval system conceptual-
 N −1 ization,” Int. J. Prod. Res., vol. 42, no. 22, pp. 4767–4783, 2004.
(N ρ) k −1
[ N(N!(1−ρ)
ρ) N
+ k = 1k ! ] . [18] R. D. Meller and J. F. Klote, “A throughput model for carousel/VLM
4(c 2 −c 2 ) pods,” IIE Trans., vol. 36, no. 8, pp. 725–741, 2004.
When c2aP ≥ c2sP , φ(ρ, c2aP , c2sP , N ) = ( 4c 2a P −3cs2P )φ1 [19] N. Litvak, “Optimal picking of large orders in carousel systems,” Oper.
aP sP
c2 P c 2 +c 2 Res. Lett., vol. 34, no. 2, pp. 219–227, 2006.
(N, ρ) + ( 4c 2 s−3c 2 )ψ( a P 2 s P , N, ρ). [20] Y. A. Bozer and M. Cho, “Throughput performance of automated stor-
aP sP
c 2 −c 2 age/retrieval systems under stochastic demand,” IIE Trans., vol. 37, no. 4,
When c2aP < c2sP , φ(ρ, c2aP , c2sP , N ) = ( 2c 2s P +2ca P2 )φ3 pp. 367–378, 2005.
aP sP
c 2 +3c 2a P c 2a P +c 2s P [21] J.-S. Song, S. H. Xu, and B. Liu, “Order-fulfillment performance measures
(N, ρ) + ( 2cs2P 2 )ψ( 2 , N, ρ). in an assemble-to-order system with stochastic leadtimes,” Oper. Res.,
a P +2c s P
vol. 47, no. 1, pp. 131–149, 1999.
Where c2aP + c2sP
= 1, φ1 (N, ρ) = 1 + γ(N, ρ), φ2 (N, ρ) [22] R. D. Andrea and P. Wurman, “Future challenges of coordinating hundreds
= 1 − 4γ(N, ρ), φ3 (N, ρ) = φ2 (N, ρ) exp{ −2(1−ρ)
3ρ }, φ4 (N, of autonomous vehicles in distribution facilities,” in Proc. IEEE Int. Conf.
φ 3 (N ,ρ)
Technol. Practical Robot Appl., 2008, pp. 80–83.
ρ) = min{1, φ1 (N, ρ) + 2 }, and γ(N, ρ) = min{0.24, [23] P. R. Wurman, R. D’Andrea, and M. Mick, “Coordinating hundreds of
1 cooperative, autonomous vehicles in warehouses,” AI Mag., vol. 29, no. 1,
1) (4+5N ) −2 2
(1 − ρ)(N − 16N ρ }. pp. 9–19, 2008.
YUAN AND GONG: BOT-IN-TIME DELIVERY FOR ROBOTIC MOBILE FULFILLMENT SYSTEMS 93

[24] J. J. Enright and P. R. Wurman, “Optimization and coordinated autonomy [39] W. Whitt, “The queueing network analyzer,” Bell Syst. Tech. J., vol. 62,
in mobile fulfillment systems,” in Proc. 9th AAAI Conf. Autom. Action no. 9, pp. 2779–2815, 1983.
Planning Auton. Mobile Robots, 2011, pp. 33–38. [40] W. Whitt, “Approximations for the GI/G/m queue,” Prod. Oper. Manage.,
[25] J. Chen and M. Dong, “Available-to-promise-based flexible order alloca- vol. 2, no. 2, pp. 114–161, 1993.
tion in ATO supply chains,” Int. J. Prod. Res., vol. 52, no. 22, pp. 6717–
6738, 2014.
[26] I. Higa, A. M. Feyerherm, and A. L. Machado, “Waiting time in an (S- 1,
S) inventory system,” Oper. Res., vol. 23, no. 4, pp. 674–680, 1975.
[27] T. Wensing and H. Kuhn, “Analysis of production and inventory systems Zhe Yuan received the master’s degree in engineer-
when orders may cross over,” Ann. Oper. Res., vol. 231, no. 1, pp. 265–281, ing from the Huazhong University of Science and
2015. Technology, Wuhan, China, in 2015. She is currently
[28] J. C.-H. Pan, P.-H. Shih, and M.-H. Wu, “Storage assignment problem working toward the Ph.D. degree in industrial en-
with travel distance and blocking considerations for a picker-to-part order gineering at CentraleSupélec, University of Paris-
picking system,” Comput. Ind. Eng., vol. 62, no. 2, pp. 527–535, 2012. Saclay, Paris, France.
[29] Y. Gong and R. De Koster, “A polling-based dynamic order picking system She has authored or co-authored articles in jour-
for online retailers,” IIE Trans., vol. 40, no. 11, pp. 1070–1082, 2008. nals such as the International Journal of Produc-
[30] P. Toth and D. Vigo, Vehicle Routing: Problems, Methods, and Applica- tion Research and Chinese Management Science.
tions. Philadelphia, PA, USA: SIAM, 2014. Her research interests include operations manage-
[31] J. H. Lee, B. H. Lee, and M. H. Choi, “A real-time traffic control scheme of ment, warehouse management, and flexibility in
multiple AGV systems for collision free minimum time motion: A routing supply chain management.
table approach,” IEEE Trans. Syst., Man, Cybern. A, Syst. Humans, vol. 28,
no. 3, pp. 347–358, May 1998.
[32] I. F. Vis, “Survey of research in the design and control of automated guided
vehicle systems,” Eur. J. Oper. Res., vol. 170, no. 3, pp. 677–709, 2006.
[33] L. Qiu, W.-J. Hsu, S.-Y. Huang, and H. Wang, “Scheduling and routing
algorithms for AGVs: A survey,” Int. J. Prod. Res., vol. 40, no. 3, pp. 745– Yeming (Yale) Gong received the Ph.D. degree in
760, 2002. management from the Rotterdam School of Manage-
[34] L. Deroussi, M. Gourgand, and N. Tchernev, “A simple metaheuristic ap- ment, Erasmus University Rotterdam, Rotterdam, the
proach to the simultaneous scheduling of machines and automated guided Netherlands, in 2009.
vehicles,” Int. J. Prod. Res., vol. 46, no. 8, pp. 2143–2164, 2008. He is an Associate Professor with the EMLYON
[35] P. Udhayakumar and S. Kumanan, “Task scheduling of AGV in FMS using Business School, Lyon, France. He studies the inter-
non-traditional optimization techniques,” Int. J. Simul. Model., vol. 9, face between operations management and IT, logis-
no. 1, pp. 28–39, 2010. tics, and supply chain management. He has authored
[36] C.-H. Hu and P. J. Egbelu, “A framework for the selection of idle vehicle or co-authored two books with Erasmus and Springer,
home locations in an automated guided vehicle system,” Int. J. Prod. Res., and articles in journals such as Production and Op-
vol. 38, no. 3, pp. 543–562, 2000. erations Management, the IEEE TRANSACTIONS ON
[37] D. Hoyt, H. Lee, and M. Marks, “Zappos.com: Developing a supply chain ENGINEERING MANAGEMENT, IIE Transactions, the European Journal of Oper-
to deliver wow,” Stanford Graduate School of Business (Cases), Case ational Research, the International Journal of Production Economics, and the
no. GS65, 2009, pp. 1–27. International Journal of Production Research.
[38] R. W. Wolff, Stochastic Modeling and the Theory of Queues. Upper Saddle Dr. Gong received the Best Paper Award in design and manufacturing from
River, NJ, USA: Pearson, 1989. the Institute of Industrial Engineers in 2010.

Das könnte Ihnen auch gefallen