Drone-Based Material Transfer System in A Robotic Mobile Fulfillment Center

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 17, NO.

2, APRIL 2020 957

Drone-Based Material Transfer System in a Robotic


Mobile Fulfillment Center
Andy Ham

Abstract— Deploying unmanned aerial vehicles, also known as


drones, for final-mile delivery in logistics operations has inspired
this research. One conceivable scenario is to use the drone to
transfer jobs in a robotic mobile fulfillment center. Each job
might be characterized by origin, destination, time window,
and precedence relation. In particular, a take-away conveyor is
continuously pushing forward shipping cartons, wherein drones Fig. 1. Transbots for pickup and dropoff tasks: Amazon [40] (left),
must pick up jobs and drop them off to the cartons within Berkshire Grey [4] (middle), and Ocado [35] (right).
the time window since the cartons would be closed and labeled
at the end of the conveyor. Furthermore, each job contains
two subtransfers for drone: flight from drone’s current location
to pickup location and flight to dropoff location. Two exact
approaches are proposed—a mixed integer programming and a
constraint programming—and tested for a real-time perspective.

Note to Practitioners—Inspired by Amazon’s significant expan-


sion of its army of warehouse robots and the adaptation of
drones for final-mile small parcel delivery, this article has mixed
those two emerging technologies and proposed a drone-powered
material transfer system in a warehouse, wherein jobs would
be picked up by drone, transferred, and dropped off directly
to shipping cartons. The most encouraging finding is that the
proposed method impressively proved the optimality of 144 jobs
with 10 drones that requires 28.8 transfers per drone on average,
within 1 s.
Index Terms— Material transfer, robotic mobile fulfillment
system (RMFS), transbot, unmanned aerial vehicle (UAV), ware-
house.
I. I NTRODUCTION

A ROBOTIC mobile fulfillment system (RMFS) is a new


type of automated storage and part-to-picker order pick-
ing system. It is particularly suited for e-commerce distribution
Fig. 2. Robot-based configuration [12].

inventory shelves from the storage area and bring them to


centers that handle strong demand fluctuations and large
the stations where, guided by software and laser pointers that
assortments of small products [30]. Fig. 1 depicts different
identify the proper inventory on the shelving unit, the operators
mobile robots used in RMFS.
pick the job and put it into the outgoing shipping cartons. Emp-
In particular, Amazon utilized 45 000 mobile robots across
tied bins are filled at replenishment stations and decremented
20 RMFS centers [6]. Amazon’s innovative approach to ware-
at pick stations [12].
house automation uses hundreds of custom-built mobile robots
With a parallel effort, Amazon and several other compa-
(transbots hereafter) that carry small shelving units and deliver
nies announced their consideration of rapid delivery using
products (jobs hereafter) to human operators, as depicted
unmanned aerial vehicles (UAVs), also known as drones, for
in Fig. 2. The operators stand at work stations along the
final-mile small parcel delivery. It seems that a similar idea
perimeter of a storage area that is filled with thousands of
can be applied to RMFS operation wherein jobs could be
storage shelves (pods). These shelving units are a collection
picked up by drone, transferred, and dropped off directly to
of storage locations, called bins. The transbots fetch specific
shipping cartons, assuming the drone has a robot arm to pick
Manuscript received September 6, 2019; accepted November 6, 2019. Date up a job. The benefits of employing drones as transbots in a
of publication December 3, 2019; date of current version April 7, 2020. This warehouse include speed and cost. Amazon’s transbot operates
article was recommended for publication by Associate Editor M. Dotoli and
Editor S. A. Reveliotis upon evaluation of the reviewers’ comments. at a max speed of 5 mph (Machine Design [41]), whereas
The author is with the School of Engineering, Liberty University, Lynch- Amazon’s delivery drone runs at a max speed of 50 mph [14].
burg, VA 24501 USA (e-mail: mham@liberty.edu). Currently, the transbot must carry shelving units back to
Color versions of one or more of the figures in this article are available
online at http://ieeexplore.ieee.org. storage area after being picked by operators, but this task can
Digital Object Identifier 10.1109/TASE.2019.2952523 be removed since the drone does not transfer the shelves, but
1545-5955 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded on February 10,2021 at 05:44:28 UTC from IEEE Xplore. Restrictions apply.
958 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 17, NO. 2, APRIL 2020

Fig. 3. Foldable robotic arm enables drones to perform tasks that could not be
performed otherwise. A foldable arm could be folded into a compact volume
and deployed on command. The arm equipped with an adequate end-effector
could grasp an object or inspect inside a narrow deep space [25].

a specific item on shelves. Furthermore, since one drone can


potentially complete more transfer requests than several low-
profile moving transbots, it will lead to a cost reduction. Now, Fig. 4. Drone-based configuration (proposed in this article).
the question is “Does such a drone exist?”
A team of South Korean researchers has brought folding
arms and drone technology together to create a device that The rest of this article is organized as follows: a literature
can collect objects (Fig. 3). The proposed foldable arm helps review is presented in Section II and the proposed mixed-
drones overcome access issues by allowing them to reach into integer programming (MIP) and constraint programming (CP)
confined spaces. The origami-inspired design [22] enables a models are developed in Section III. Computational results
foldable structure to be lightweight, compact, and scalable are reported in Section IV, and finally, Section V covers the
while maintaining its kinematic behavior [25]. conclusion.
Although the arm has several limitations (the length, the
stiffness, and the degrees of freedom) that need to be yet II. L ITERATURE R EVIEW
overcome, drones equipped with the proposed foldable arm Lamballais et al. [30] addressed the design problem of
could replace transbots in warehouses in a near future as the RMFS related to throughput, cycle time, number of robots,
relevant research is on-going [3]. Such a proposed system and different layouts, but they could not find literature related
is depicted in Fig. 4, wherein drones pick up, transfer, and to warehouse transbot scheduling problem other than the one
drop off jobs directly to shipping cartons. To maximize a by Enright and Wurman [12]. Instead, the traditional pickup
space footage of storage, the proposed system employs the and delivery problem with time windows (PDPTWs) or dial-a-
inventory storage tower. Each floor of the tower is filled with ride problem (DARP), which considers a set of point-to-point
bins. Each bin that holds the same type of jobs would not transportation jobs defined by customer demands that must be
be layered on top of other bins in order to facilitate drone’s collected from pickup locations and delivered to paired dropoff
approach. A drone navigates through the tower, fetches a locations, can be applied to the proposed material transfer
specific job from the bin, and drops off the job directly to drone scheduling problem (MTDSP). However, each job has
a shipping carton without a human intervention. to be transported directly from its pickup location to its paired
Enright and Wurman [12] listed the five subproblems dropoff location in our problem. In addition, the decision must
related to its system: inventory pod selection problem, pod be made in a real-time (less than a couple of seconds) in a
storage allocation, order allocation problem, replenishment high-volume RMFS operation.
allocation problem, and robot allocation problem. This article Gerkey and Matarić [15] categorize the features and com-
concentrates on the robot allocation problem, simultaneously plexity of multirobot task allocation (MRTA) problems with
finding optimal sequences and assignments of drones to jobs three axes. Then, the study was extended by Korsah et al. [26].
in a setting where drones are identical and in which it is The MTDSP falls into the category of ST–SR–TA since each
desirable to minimize the maximum travel completion time of robot can execute only one task at a time (ST), each task
all jobs. Each job is characterized by origin, destination, and requires exactly one robot to achieve it (SR), and each robot
time window. A take-away conveyor is continuously pushing is allocated several tasks that must be executed over time (TA).
forward shipping cartons, wherein drones must drop off jobs This article adopted an MIP formulation by Cordeau [7] and
to the cartons within time window since the cartons would be modified it to reflect a unit capacity of a drone. Ropke and
closed and labeled at the end of the conveyor. Furthermore, Cordeau [36] and Cordeau and Laporte [8] propose the well-
each job contains two subtransfers for a drone: flight from known benchmark instances on the PDPTW problem. This
the drone’s current location to pickup location and flight to article adopted the same instances to estimate the performance
dropoff location. of the proposed models.

Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded on February 10,2021 at 05:44:28 UTC from IEEE Xplore. Restrictions apply.
HAM: DRONE-BASED MATERIAL TRANSFER SYSTEM IN A ROBOTIC MOBILE FULFILLMENT CENTER 959

Maciejewski et al. [33] found that online management of 3) All jobs are assumed to be light-weighted (<5 pounds).
taxis has become standard, and nowadays, the focus is on For heavy-weighted jobs, specialized drone could be
operating a taxi fleet both at the microscopic level of detail built and modeled as an unrelated parallel machine
and in real time. At present, because of recent information and scheduling problem.
communications technology (ICT) advances, the dispatching 4) Recharging time of the drone can be neglected, e.g.,
decision more often is based on the taxis’ exact geographical by swapping batteries [23].
locations, in order to send the nearest one. This strategy 5) The drone has a capability of collision avoidance [32],
seems myopic and far from optimal because dispatch decisions [39] so this article does not consider a dynamic route
are optimized only locally by considering a single request decision.
or vehicle. To achieve higher efficiency dispatching, they 6) The travel time of drone between nodes is known
suggested to take a more global look at its process. Therefore, a priori. The traffic coordination challenges that arise
this article focuses on real-time scheduling of transbots by in the context of many automated unit-load material
considering entire jobs and transbots together. handling systems [9] are not considered.
To the best of our knowledge, there is no work on the
MTDSP. The contributions of this article are threefold. First, A. MIP Approach
the MTDSP, which we will encounter in the near future, The notation used in this article is summarized in the
is addressed for the first time. Second, an MIP formulation following.
is proposed. Third, a novel CP formulation is introduced for Sets:
real-time decision (<1 s).
J jobs ( j ).
V drones (v).
III. M ATERIAL T RANSFER D RONE S CHEDULING P ROBLEM
Parameters:
The problem may be formally defined as follows. Let J
represent the set of jobs {1, . . . , n}, where n denotes the ej EST of job j.
number of jobs (items to deliver). Then let V represent the lj LST of job j.
set of drones {1, . . . , k}, where k denotes the number of ti j travel time of drone between jobs i and j.
parallel drones. The MTDSP can be m vi j large positive number (max{0,l j + ti j − e j }).
 defined
 on a directed
graph G = (N, A), where N = P D {2n + 1, 2n + 2},
Decision Variables:
P = {1, . . . , n}, D = {n + 1, . . . , 2n}. Subsets P and D
contain pickup and dropoff nodes, respectively, while nodes Bvi the time at which drone v begins service at node i .
2n + 1 and 2n + 2 represent the origin and destination depots. X vi j 1 if drone v travels from node i to j .
With each job j is associated an origin node i and a destination
node n +i . A is the set of arcs. For each arc (i, j) ∈ A, the time The original MIP formulation provided by Cordeau [7] is
required for the drone to travel between nodes is given by ti j . modified to represent MTDSP.
The travel time is assumed to be given with a high accuracy as MIP
the research of optimal trajectory planning for multiple robots
with collision avoidance is ongoing [13], [38]. Min Max{Bvi + (ti,2n+2 )X v,i,2n+2 v ∈ V, i ∈ D} (1)
A time window [e j , l j ] is associated with each job j , where X vii = 0 ∀v ∈ V, i ∈ N (2)
e j and l j represent the earliest start time (EST) and the latest  
X vi j = 1 ∀i ∈ P (3)
start time (LST), respectively. Furthermore, each job has to be
v∈V j ∈N
transferred directly from pickup location to its paired dropoff  
location since a drone has a unit capacity. Naturally, the pickup X vi j = X v,i+n, j ∀v ∈ V, i ∈ P (4)
must occur prior to the dropoff. The objective is to schedule j ∈N j ∈N

parallel drones to minimize the maximum completion time X v,2n+1, j = 1 ∀v ∈ V (5)
of all jobs while satisfying the time window and precedence j ∈N
relationship. The following operating conditions are assumed.  
Xv ji = X vi j ∀v ∈ V, i ∈ P ∪ D (6)
1) The drone has a unit capacity. Therefore, the drone must j ∈N j ∈N
navigate to pair shipping carton for dropoff, after picking 
up a job. Namely, the drone always starts with pickup X v,i,2n+2 = 1 ∀v ∈ V (7)
i∈N
and immediately follows by dropoff. 
2) Since the take-away conveyor is continuously pushing X v,i, j = 1 ∀i ∈ P (8)
forward shipping cartons and the cartons would be v∈V ,
j =i+n
closed and labeled at the end of the conveyor, the travel
distance is calculated for the worst-case scenario as Bv j ≥ Bvi +ti, j −m vi j (1− X v,i, j ) ∀v ∈ V , (i, j ) ∈ N (9)
follows: when the drone flies to the shipping carton, e j ≤ Bv j ≤ l j ∀v ∈ V, j ∈ N (10)
the travel distance is calculated from the job’s location X vi j ∈ {0, 1}, Bv j ≥ 0 ∀v ∈ V, i ∈ N, j ∈ N. (11)
to the end of the conveyor line in order to ensure a
timely drop.

Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded on February 10,2021 at 05:44:28 UTC from IEEE Xplore. Restrictions apply.
960 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 17, NO. 2, APRIL 2020

Objective (1) is to minimize the maximum completion time to quickly come up with an efficient solution for scheduling
of all jobs. The returning flight distance to the depot is added problems. CP is an optimization paradigm that can handle a
into the time at which the drone begins service at the last large number of constraint types that go beyond MIP-style
dropoff node i . Constraint (2) forbids a cycling. Constraints linear constraints. While MIP focuses on the objective function
(3) and (4) collectively ensure that each job is served exactly and its optimality, CP focuses on satisfying the constraints
once and that the pickup and dropoff nodes are visited by the and it is designed to find feasible solutions faster. CP is
same drone. Constraints (5)–(7) guarantee that the route of a suitable technique for solving sequencing and schedul-
each drone starts at the origin depot and ends at the destination ing applications, feasibility problems, and highly constrained
depot. Constraint (8) ensures that visiting at the pickup node i problems [11], outperforming MIP [17], [34]. In particular,
is immediately followed by the dropoff node j for each paired Laborie [27] compared different approaches like MIP, CP,
job ( j = i + n). Constraint (9) calculates the time at which constraint integer programming (CIP), logic-based Benders
drone v begins service at node i , while (10) imposes time decompositions (LBBD), and SAT-modulo theories (SMT)
windows. for resource allocation and scheduling problems and demon-
Furthermore, the following valid inequalities (12) and (13) strated that CP model outperformed all other approaches
are added into the model as suggested by Desrochers and and close all the 335 benchmark instances created by
Laporte [10]. These constraints are valid for feasible regions of Hooker [19].
the problem studied and can lead faster solutions by tightening The search within IBM ILOG CP Optimizer is equipped
the feasible region of the linear programming (LP) relaxation with the presolve functionality, some constraint propagation
algorithms, the temporal linear relaxation used to guide the
 search, and the two search space exploration strategies that
Bvi ≥ ei + max{0,e j − ei + ti j }X v j i (12)
are used concurrently: the large neighborhood search (LNS)
j ∈N\{i}
 (for producing good-quality solutions) and failure-directed
Bvi ≤ li + max{0,li − l j + ti j }X vi j . (13) search (for proving infeasibility or optimality). LNS is a
j ∈N\{i} component of CP Optimizer automatic search for scheduling
The impact of tightening a solution space with the lower consisting of successive relaxation and reoptimization phases.
and upper bounds is also tested in experiments, referring It operates by first finding an initial feasible solution. Then,
the model without bounds as MIP and the one with bounds a number of iterations are carried out, and each iteration
MIPvi . The proposed model quickly proved the optimality comprising a relaxation step followed by a reoptimization
of a minisample problem. However, it could not handle the of the relaxed solution. This process continues until some
test instances with 60+ jobs and 6+ drones after 0.5 h of condition is satisfied, typically, when the solution can be
computational time due to a large number of binary vari- proved to be optimal or when a time limit is reached.
ables, so this article proposes another approach: CP, which In CP Optimizer, this approach is made more robust by the
also proves an optimality. CP has been applied to a wide application of machine learning techniques to portfolios of
range of combinatorial optimization problems, excelling most large neighborhoods and completion strategies in order to
notably in scheduling applications [1] and CP has established converge on the best combined method for the problem being
itself as a strong competitor to mathematical programming- solved [29]. Because of space limitations, readers are recom-
based approaches, often outperforming state-of-the-art MIP mended to review Laborie and Rogerie [28], Laborie [27], and
solvers [37]. IBM Software [20].
It is worth mentioning that CP formulation is very different
B. CP Approach from MIP formulation. Furthermore, there is no standard
Formulating a problem in an MIP has been taken granted in CP formulation. Namely, it varies to each CP package,
to be the first step to address the problem. However, Hillier unlike a similar MIP formulation [17]. Therefore, this arti-
and Lieberman [18] discuss that no presentation of the basic cle attempts to provide a generic formulation based on CP
ideas of MIP is complete these days without introducing a Optimizer.
relatively recent development—the incorporation of the tech- 1) CP1 (Generic Model for DARP With Time Window): The
niques of CP—that is promising to greatly expand our ability proposed CP model is built upon the following three decision
to formulate and solve MIP models. variables.
As a practitioner in the field, the author has visited a J j interval representing the j th job.
countless number of factories and warehouses and noticed J 2V j,v optional interval representing the j th job on
that practitioners working on scheduling projects have dif- drone v.
ficulties in developing, validating, implementing, debugging, Sv collection of interval variables ( J j ) assigned to drone v.
and maintaining complex MIP models, not to mention its slow An interval variable represents an interval of time dur-
computational speed. Therefore, practitioners often decompose ing which a task occurs. An interval variable can represent
the original problem into two subproblems: allocation and optional tasks to be scheduled. The domain of an interval
sequencing, where the allocation is determined by an MIP variablex is a subset of {Absent}∪{[s, e)|s, e∈ Z,s ≤ e}. Let
and the sequencing is generated by a set of dispatching Ñ = P D {2n+ 1,…, 2n+ 1 + 2k}, while nodes 2n+ 1
rules (or heuristic). On the other hand, CP provides a great thru 2n+ 2 + 2k represent the origin and destination depots
expressiveness feature and demonstrates an effective algorithm for each drone.

Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded on February 10,2021 at 05:44:28 UTC from IEEE Xplore. Restrictions apply.
HAM: DRONE-BASED MATERIAL TRANSFER SYSTEM IN A ROBOTIC MOBILE FULFILLMENT CENTER 961

CP1
MinMax j ∈ Ñ {endOf(J j )} (14)
alternative(J j , [ J 2V j v ]v∈V ) ∀ j ∈ Ñ (15)
e j ≤ startOf(J j ) ≤ l j ∀ j ∈ Ñ (16)
Fig. 5. Streamlined modeling of pickup and dropoff. The transfer between
endBeforeStart(Ji , J j , ti j ) ∀(i, j ) ∈ Ñ :i = j − n (17) different pairs is modeled as a transfer time, while the transfer within a pair
is modeled as a size of interval.
before(Sv , J 2V iv , J 2V j v ) ∀v ∈ V :i = j − n (18)
presence(J 2V iv ) = presence(J 2V j v ) ∀v ∈ V :i = j − n
(19) the end of a given interval variable a equals the start of a
given interval variable b, and the function prev(p, a, b) states
presenceOf(J 2V j v ) = 1 ∀v ∈ V (20)
that if intervals a and b are resent, a appears just before b
first(Sv , J 2V j v ) ∀v ∈ V, j ∈ Ñ : j = v + 2n (21) in sequence p; no other intervals can be ordered between a
presenceOf(J 2V j v ) = 1 ∀v ∈ V, j ∈ Ñ : j = v +2n (22) and b in sequence p. Finally, constraint (25) that checks the
last(Sv , J 2V j v ) ∀v ∈ V, j ∈ Ñ : j = k + v + 2n (23) capacity of drone is dropped because this restriction is already
satisfied by constraints (26) and (27). All other constraints
noOverlap(Sv , t) ∀v ∈ V, j ∈ Ñ : j = k + v + 2n (24)
 remain in CP2.
stepAtStart(J 2V j v , 1) CP2
∀ j ∈P
 endAtStart(Ji , J j , ti j ) ∀(i, j ) ∈ Ñ :i = j − n (26)
+ stepAtStart(J 2V j v , −1) ≤ 1 ∀v ∈ V. (25) prev(Sv , J 2V iv , J 2V j v ) ∀v ∈ V :i = j − n. (27)
∀ j ∈D

Objective (14) minimizes the time required to complete all 3) CP3 (Streamlined Model for MTDSP): The previous
jobs. The expression endOf (i ) represents the end of interval two CP models defined pickup and dropoff as separate jobs.
variable i whenever the interval variable is present (otherwise, Here, we present an alternative CP model that merges those
its value is 0 by default). Constraint (15) assigns each job two transfers into one as depicted in Fig. 5. Since each job
to exactly one of drones. Both intervals of J j and J 2V j,v is known a priori, both pickup and dropoff can be integrated
start and end together with this chosen one. Constraint (16) into a single transfer job, reducing the number of decision
imposes time windows. Constraints (17) and (18) ensure the variables into a half. Ham [16] also proposed a similar idea
precedence relationship between the pickup and paired dropoff of merging two consecutive transfers since a pickup must be
jobs. Constraint (19) ensures that the pickup and dropoff nodes immediately followed by paired dropoff since a drone has unit
are visited by the same drone. Constraints (20) and (21) force capacity in a parcel-delivery-drone scheduling problem.
the drone to start from the depot and Constraints (22) and (23) The ...MTDSP can ... be  redefined on a directed graph
force to return to the depot at the end of the tour. Constraint G = ( N , A), where N = P {n +1, . . . , n +1+2k}. Note the
(24) prevents intervals in a sequence from overlapping and dropoff nodes are not included here. For each arc (i, j ) ∈ A,
enforces a minimal distance between consecutive intervals. let t˜i, j represent the time required for the drone to travel
The expression noOverlap(s,t) defines a chain of nonoverlap- between nodes. Suppose a drone is navigating to pickup node
ping intervals, and any interval in the chain is constrained j after completing a dropoff at i . Then, the transfer time can
to end before the start of the next interval in the chain. If a be determined as td(i), p( j ), where d(i ) and p( j ) refer to the
transition distance matrix t is specified, it defines the minimal dropoff and pickup nodes, respectively. The reduced number of
distance that must separate two consecutive intervals in the decision variables would shorten the computational run time.
sequence. Constraint (25) forbids the drone to carry more CP3
than one job. The function stepAtStart(a,h) is an elementary
Min Max j ∈N {endOf(J j )} (28)
cumulative function expression representing the contribution ...
beginning at the start of an interval. The start of interval a sizeOf(J j ) = t p( j ),d( j ) ∀ j ∈ N (29)
...
is the start of production or consumption. The height of the alternative(J j , [ J 2V j v ]v∈V ) ∀ j ∈ N (30)
function is represented by h. In our model, each assignment ...
e p( j ) ≤ star t O f (J j ) ≤ l p( j ) ∀ j ∈ N (31)
of pickup (dropoff) job to drone increases (decreases) the ...
ed( j ) ≤ end O f (J j ) ≤ ld( j ) ∀ j ∈ N (32)
cumulative function at the start of job. ...
2) CP2 (Customized Model for MTDSP): The CP1 pro- presenceOf(J 2V j v ) = 1 ∀v ∈ V, j ∈ N : j = v +2n (33)
...
vides a generic model for DARP with time window. However, first(Sv , J 2V j v ) ∀v ∈ V, j ∈ N : j = v + 2n (34)
in the MTDSP, a drone can carry only one job at a time, which ...
presenceOf(J 2V j v ) = 1 ∀v ∈ V, j ∈ N : j = k + v + 2n
forces the drone to visit paired dropoff node immediately after
picking up a job. Therefore, the CP1 model can be customized. (35)
...
Constraints (17) and (18) are replaced with constraints (26) last(Sv , J 2V j v ) ∀v ∈ V, j ∈ N : j = k + v + 2n (36)
and (27) to ensure the immediate tours from pickup to its noOverlap(Sv , t) ∀v ∈ V. (37)
paired dropoff node. The function endAtStart(a,b) states that

Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded on February 10,2021 at 05:44:28 UTC from IEEE Xplore. Restrictions apply.
962 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 17, NO. 2, APRIL 2020

Constraint (29) guarantees the transfer time between paired TABLE I


jobs. Constraint (30) assigns each job exactly to one of the MTDSP: MIP AND CP R ESULTS ON I NSTANCES OF ROPKE
AND C ORDEAU [36]
drones. Constraints (31) and (32) impose time windows at
pickup and dropoff nodes, respectively. All remaining con-
straints are similar to the ones in CP1.
The performance of the three different CP models can
be improved by using variable orderings heuristics (VOH).
In the CP search strategy, the order of the search phases in
the array is important. The variables are instantiated phase
by phase starting by the first phase of the array. There are
some models where giving such an order can have a dramatic
impact on the solution time. Beck et al. [2] discussed that
the “likelihood of finding a solution” could be treated prob-
abilistically. For a given decision, there is some probability
over all possible subsequent decisions that the choice will
lead to a solution. For the purposes of our investigation, this
article implements instantiation orderings over variable groups
defined as: V := {I 2V j,v , Sv }, similar to the method proposed
by Booth et al. [5].

IV. C OMPUTATIONAL E XPERIMENTS


In this section, the effectiveness of the proposed model is
examined. The MIP, CP, and flow control models are all coded
in IBM OPL 12.8.0 on a personal computer with an Intel Core
i7-4770 CPU with 16 GB of RAM.

A. Problem Instances
The proposed models were tested using the problem
instances proposed by Ropke and Cordeau [36]. There are
four classes of instances (AA, BB, CC, and DD) with
different time windows with width W . Classes AA and BB
have W = 60, and classes CC and DD have W = 120. The
capacity of each drone and load was set to 1 to reflect a unit
capacity of a drone. Each class consists of ten instances and
the number of pairs of pickup and dropoff nodes of these
instances is multiples of five. The smallest instance of a
class has 30 pairs and the largest instance has 75 pairs. The
class name and the number of pairs of pickup and dropoff
nodes determine the instance name. For example, AA30 is
the instance in class AA with |P| = |D| = 30. The instances
consider homogeneous fleet and a single depot located at the
middle of a [0, 50] ×[0, 50] square.
The proposed models were also tested using the problem
instances proposed by Cordeau and Laporte [8] since they
have a large instance. There is a data set of 20 randomly
generated instances, containing between 24 and 144 pairs.
The load was set to 1 for pickup and −1 for dropoff.
The time windows were set to the length of the planning
horizon [0, 1440]. The data set was split into two groups.
In group (a), narrow time windows were set, while in
group (b), wider time windows were created. Due to the
unit capacity of the drones in our experiment, all instances
turned out to be infeasible. Note the original instances have
the vehicle capacity of six. Therefore, the latest start (l j )
is ignored to ensure the feasibility, while the earliest start

Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded on February 10,2021 at 05:44:28 UTC from IEEE Xplore. Restrictions apply.
HAM: DRONE-BASED MATERIAL TRANSFER SYSTEM IN A ROBOTIC MOBILE FULFILLMENT CENTER 963

TABLE II CP3 (streamlined model for MTDSP) proved the optimality


MTDSP: CP R ESULTS ON I NSTANCES OF C ORDEAU of all instances in 1.23 and 0.26 s on average, respectively.
AND L APORTE [8]
Table II gives the results of the experiments using the
proposed models on instances by Cordeau and Laporte [8].
Column 1 identifies the name of the instance, column 2 con-
tains the number of drones (k), column 3 contains the number
of pairs (n), and column 4 calculates the required number
of transfers per drone on average. The results from MIP are
not included because both MIP models failed to find feasible
solutions for most instances except (R1a, R2a, R7a, and R1b).
Column 5 (6 and 7) records the elapsed time given in seconds
for CP1 (CP2 and CP3), respectively. When an instance could
not find any feasible solution in 10 s, then the column has the
symbol –. Column 10 contains the optimal objective values.
The results of CP models are impressive. The CP1 quickly
proved the optimality except R10a, R1b, R4b, R5b, R6b, and
R10b instances, spending 4.01 s on average. CP2 proved the
optimality except R1b and R5b instances. Finally, CP3 proved
the optimality of all instances in 0.12 s on average.
Finally, the CP3 is tested using the problem instances
designed by Li and Lim [31], which have more than 500 pairs.
The latest start (l j ) is ignored to ensure the feasibility, while
the earliest start (e j ) is conserved. These data are down-
loaded from https://www.sintef.no/projectweb/top/pdptw/li-
lim-benchmark/1000-customers/. CP3 proved the optimality of
all LC1 and LC2 instances with the makespan minimization
objective within 18.4 s with 100 drones, but 90.7 s with
500 drones.
V. C ONCLUSION
Inspired by Amazon’s significant expansion of its army
of warehouse robots and the adaptation of drones for final-
mile small parcel delivery, this article has mixed those two
(e j ) is conserved. MIP and CP logs (to construct Gantt- emerging technologies and proposed a drone-powered material
chart) for each instance are located at the following link: transfer system in a warehouse, wherein jobs could be picked
https://drive.google.com/open?id=1Z4qoypEhCn4kXJl53to4S up by the drone, transferred, and dropped off directly to ship-
y5LP574MoNo. ping cartons. Contrary to its current mobile robots, which carry
shelving units to picking stations (or replenishment stations)
B. Results where human operators pick up (or replenish) jobs, this article
Table I gives the results of the experiments using the proposes a direct delivery of jobs via drone, erasing all human
proposed models, all solved by the optimization software operations occurring at picking and replenishment stations. For
IBM OPL 12.8. The top row indicates the column number. this MTDSP, constrained by time window and precedence rela-
Column 1 in Table I identifies the name of the instance and tion, with the objective to minimize a maximum completion
Column 2 contains the number of drones (k). Columns 3 and time over all jobs, MIP and CP models are suggested. The
4 give the results of the MIP. Column 3 records the elapsed original MIP formulation provided by Cordeau [7] is modified
time given in seconds. The bold indicates that CPLEX found to reflect a unit capacity of drone. Then, the modified MIP
the optimal solution for the corresponding instance. When an model is improved by valid inequalities. Similarly, a generic
instance could not find any feasible solution in 1800 s, then CP model (CP1) for DARP with a time window is initially
the column has the symbol –. Column 4 reports the optimality proposed. After experiencing a slow computational run time,
gap. Similarly, columns 5 and 6 give the results of the MIPvi . two other CP models are proposed. CP2 removes the capacity
The valid inequalities do help CPLEX find solutions faster and constraint and ensures the consecutive tours from pickup to
converge to optimal in several instances (AA55, BB50, CC40, dropoff of paired jobs. CP3 streamlines pickup and dropoff
and DD40), but fail on the most of CC and DD instances. transfers by merging them, reducing the number of decision
Column 7 (8 and 9) records the elapsed time given in sec- variables into a half. The transfer between different pairs is
onds for CP1 (CP2 and CP3), respectively. Column 10 contains modeled as a transfer time, while the transfer within a pair is
the optimal objective values. The results of CP models are modeled as a size of interval.
impressive. The CP1 (generic model for DARP) quickly The computational study demonstrates the merit of CP over
proved the optimality except BB70 instance, spending 2.17 s MIP. The most encouraging finding is that CP3 impressively
on average. Both CP2 (customized model for MTDSP) and proved optimality of test instances (up to 144 pairs and

Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded on February 10,2021 at 05:44:28 UTC from IEEE Xplore. Restrictions apply.
964 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 17, NO. 2, APRIL 2020

10 drones) within a second when each drone performs 20–36 [10] M. Desrochers and G. Laporte, “Improvements and extensions to the
transfers. CP3 also proved the optimality of all LC1 and Miller-Tucker-Zemlin subtour elimination constraints,” Oper. Res. Lett.,
vol. 10, no. 1, pp. 27–36, 1991.
LC2 instances [31] in 18.4 s on average with 100 drones, [11] E. B. Edis and I. Ozkarahan, “A combined integer/constraint program-
but 90.7 s with 500 drones. ming approach to a resource-constrained parallel machine scheduling
Overall, the results indicate that CP is a promising tech- problem with machine eligibility restrictions,” Eng. Optim., vol. 43,
no. 2, pp. 135–157, 2011.
nology for the MTDSP. The CP technology would be able
[12] J. J. Enright and P. R. Wurman, “Optimization and coordinated auton-
to pave the way for a real-time scheduler (RTS), which has omy in mobile fulfillment systems,” in Proc. 9th AAAI Conf. Automated
been a dream of practitioners in manufacturing and warehouse. Action Planning Auton. Mobile Robots, Aug. 2011, pp. 33–38.
Formulating a scheduling problem in MIP is traditionally [13] R. J. Gaskins and J. A. Tanchoco, “Flow path design for automated
guided vehicle systems,” Int. J. Prod. Res., vol. 25, no. 5, pp. 667–676,
assumed to be the first step to scope the problem, mainly 1987.
because it proves optimality, although it is limited to small [14] A. Boyle, “Jeff Bezos shares big ideas, from Amazon’s drone landing
test instances. Now, CP proves optimality for much larger pads to a 10,000-year clock,” GeekWire, Oct. 23, 2016. [Online].
test instances so that researchers are recommended to model Available: https://www.geekwire.com/2016/jeff-bezos-pathfinder-drone-
clock-space/
a scheduling problem in CP as well as MIP. [15] B. P. Gerkey and M. J. Matarić, “A formal analysis and taxonomy of
As this is the first article to address the use of a fleet of task allocation in multi-robot systems,” Int. J. Robot. Res., vol. 23, no. 9,
drones in a warehouse, a couple of areas can be foreseen for pp. 939–954, 2004.
future research. [16] A. M. Ham, “Integrated scheduling of m-truck, m-drone, and m-depot
constrained by time-window, drop-pickup, and m-visit using constraint
1) Collaboration With Other Robots: In this article, programming,” Transp. Res. C, Emerg. Technol., vol. 91, pp. 1–14,
the drone is assumed to be the only constrained resource, Jun. 2018.
but sharable packaging robots, inspection robots, and [17] A. M. Ham and E. Cakici, “Flexible job shop scheduling problem with
parallel batch processing machines: MIP and CP approaches,” Comput.
picking robots along with transfer-drones can work Ind. Eng., vol. 102, pp. 160–165, Dec. 2016.
collectively in a future warehouse. Also, this arti- [18] F. S. Hillier and G. J. Lieberman, Introduction to Operations Research,
cle assumes that there are only light-weighted jobs 10th ed. Englewood Cliffs, NJ, USA: Prentice-Hall, 2015.
(<5 pounds) in a warehouse, allowing the proposed [19] J. N. Hooker, “Planning and scheduling by logic-based benders decom-
position,” Oper. Res., vol. 55, no. 3, pp. 588–602, 2007.
drone-based transfer. However, a warehouse would carry [20] IBM ILOG CPLEX Optimization Studio V12.6.3., IBM Softw., Armonk,
heavy- and light-weighted jobs. In this case, a team NY, USA, 2015.
of drones could cooperate with each other to handle [21] Intel. (Apr. 21, 2015). The Most Sophisticated Manufacturing Process
the heavy-weighted. Khamis et al. [24] discussed the in the World [Video file]. [Online]. Available: https://www.youtube.com/
watch?v=-KTKg0Y1snQ
similar problem, especially when it comes to heteroge- [22] Z. Kan, Y. Zhang, Y. Yang, Y. A. Tse, and M. Y. Wang, “An origami-
neous robots equipped with different capabilities that inspired monolithic soft gripper based on geometric design method,”
are required to perform various tasks with different in Proc. 2nd IEEE Int. Conf. Soft Robot. (RoboSoft), Apr. 2019,
pp. 470–476.
requirements and constraints in an optimal way.
[23] Q. Kang, J. Wang, M. Zhou, and A. C. Ammari, “Centralized charging
2) Simulation Study: It will be very interesting to contrast strategy and scheduling algorithm for electric vehicles under a battery
the robot-based configuration [12] and this drone-based swapping scenario,” IEEE Trans. Intell. Transp. Syst., vol. 17, no. 3,
configuration via a simulation study, in terms of cost and pp. 659–669, Mar. 2016.
[24] A. Khamis, A. Hussein, and A. Elmogy, “Multi-robot task
throughput. allocation: A review of the state-of-the-art,” Cooperative Robots and
Sensor Networks. Cham, Switzerland: Springer, 2015, pp. 31–51.
R EFERENCES [25] S.-J. Kim, D.-Y. Lee, G.-P. Jung, and K.-J. Cho, “An origami-inspired,
[1] P. Baptiste, C. Le Pape, and W. Nuijten, Constraint-Based Scheduling: self-locking robotic arm that can be folded flat,” Sci. Robot., vol. 3,
Applying Constraint Programming to Scheduling Problems, vol. 39. no. 16, 2018, Art. no. eaar2915.
New York, NY, USA: Springer, 2012. [26] G. A. Korsah, A. Stentz, and M. B. Dias, “A comprehensive taxonomy
[2] J. C. Beck, P. Prosser, and R. J. Wallace, “Variable ordering heuristics for multi-robot task allocation,” Int. J. Robot. Res., vol. 32, no. 12,
show promise,” in Proc. Int. Conf. Princ. Pract. Constraint Program. pp. 1495–1512, 2013.
Berlin, Germany: Springer, Sep. 2004, pp. 711–715. [27] P. Laborie, “An update on the comparison of MIP, CP and hybrid
[3] C. D. Bellicoso, L. R. Buonocore, V. Lippiello, and B. Siciliano, approaches for mixed resource allocation and scheduling,” in Proc. Int.
“Design, modeling and control of a 5-DoF light-weight robot arm Conf. Integr. Constraint Program., Artif. Intell., Oper. Res., Jun. 2018,
for aerial manipulation,” in Proc. 23rd Medit. Conf. Control Automat. pp. 403–411.
(MED), Jun. 2015, pp. 853–858. [28] P. Laborie and J. Rogerie, “Reasoning with conditional time-intervals,”
[4] Berkshire Grey. (2019). Ecommerce Fulfillment. [Online]. Available: in Proc. FLAIRS Conf., 2008, pp. 555–560.
https://www.berkshiregrey.com/solutions/ecommerce-fulfillment/ [29] P. Laborie, J. Rogerie, P. Shaw, and P. Vilím, “IBM ILOG CP optimizer
[5] K. E. C. Booth, T. T. Tran, G. Nejat, and J. C. Beck, “Mixed-integer for scheduling,” Constraints, vol. 23, no. 2, pp. 210–250, 2018.
and constraint programming techniques for mobile robot task planning,”
[30] T. Lamballais, D. Roy, and M. B. M. De Koster, “Estimating per-
IEEE Robot. Autom. Lett., vol. 1, no. 1, pp. 500–507, Jan. 2016.
formance in a robotic mobile fulfillment system,” Eur. J. Oper. Res.,
[6] S. Shead, “Amazon now has 45,000 robots in its warehouses,”
vol. 256, no. 3, pp. 976–990, 2017.
Business Insider, Jan. 3, 2017. [Online]. Available: http://www.
businessinsider.com/amazons-robot-army-has-grown-by-50-2017-1/ [31] H. Li and A. Lim, “A metaheuristic for the pickup and delivery
[7] J.-F. Cordeau, “A branch-and-cut algorithm for the dial-a-ride problem,” problem with time windows,” Int. J. Artif. Intell. Tools, vol. 12, no. 2,
Oper. Res., vol. 54, no. 3, pp. 573–586, 2006. pp. 173–186, 2003.
[8] J.-F. Cordeau and G. Laporte, “A tabu search heuristic for the static [32] B. Lu, M. Coombes, B. Li, and W.-H. Chen, “Improved situation
multi-vehicle dial-a-ride problem,” Transp. Res. B, Methodol., vol. 37, awareness for autonomous taxiing through self-learning,” IEEE Trans.
no. 6, pp. 579–594, 2003. Intell. Transp. Syst., vol. 17, no. 12, pp. 3553–3564, Dec. 2016.
[9] G. Daugherty, S. Reveliotis, and G. Mohler, “Optimized multiagent [33] M. Maciejewski, J. Bischoff, and K. Nagel, “An assignment-based
routing for a class of guidepath-based transport systems,” IEEE Trans. approach to efficient real-time city-scale taxi dispatching,” IEEE Intell.
Autom. Sci. Eng., vol. 16, no. 1, pp. 1–19, Jan. 2019. Syst., vol. 31, no. 1, pp. 68–77, Jan./Feb. 2016.

Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded on February 10,2021 at 05:44:28 UTC from IEEE Xplore. Restrictions apply.
HAM: DRONE-BASED MATERIAL TRANSFER SYSTEM IN A ROBOTIC MOBILE FULFILLMENT CENTER 965

[34] C. Maleck, G. Nieke, K. Bock, D. Pabst, and M. Stehli, “A comparison Andy Ham received M.S. degree in operations
of an CP and MIP approach for scheduling jobs in production areas with research/industrial engineering (OR/IE) from The
time constraints and uncertainties,” in Proc. Winter Simul. Conf. (WSC), University of Texas at Austin, Austin, TX, USA, in
Dec. 2018, pp. 3526–3537. 2000, and the Ph.D. degree in industrial engineering
[35] Ocado. (May 9, 2018). Inside a Warehouse Where Thousands of Robots from Arizona State University, Tempe, AZ, USA,
Pack Groceries [Video File]. [Online]. Available: htps://www.youtube. in 2009.
com/watch?v=4DKrcpa8Z_E He worked for Samsung Electronics, Kiheung,
[36] S. Ropke and J.-F. Cordeau, “Branch and cut and price for the pickup South Korea, Samsung Austin Semiconductor,
and delivery problem with time windows,” Transp. Sci., vol. 43, no. 3, Austin, GlobalFoundries, Malta, NY, USA, AMD,
pp. 267–286, 2009. Austin, and IBM/ILOG, San Jose, CA, USA, in the
[37] F. Rossi, P. van Beek, and T. Walsh, Handbook of Constraint Program-
areas of modeling, real-time dispatching, real-time
ming. Amsterdam, The Netherlands: Elsevier, 2006.
[38] M. Turpin, K. Mohta, N. Michael, and V. Kumar, “Goal assignment and scheduling, supply chain management, and decision analysis. He is currently
trajectory planning for large teams of interchangeable robots,” Auton. working as a Full Professor in industrial and systems engineering with
Robots, vol. 37, no. 4, pp. 401–415, 2014. Liberty University, Lynchburg, VA, USA, while he serves as a Technical
[39] K. Zhang, D. Zhang, A. de La Fortelle, X. Wu, and J. Grégoire, Consultant for Berkshire Grey, Boston, MA, USA: artificial intelligence
“State-driven priority scheduling mechanisms for driverless vehicles (AI) and Robotics-based omnichannel fulfillment company. His research is
approaching intersections,” IEEE Trans. Intell. Transp. Syst., vol. 16, currently focusing on real-time scheduling of drones and robots in smart
no. 5, pp. 2487–2500, Oct. 2015. factories, smart warehouses, and logistics industry. His research has been
[40] Amazon. (2014). Meet the Robots Making Amazon Even Faster published in peer-reviewed journals such as the IEEE T RANSACTIONS
[Video file]. [Online]. Available: https://www.youtube.com/watch?v= ON AUTOMATION S CIENCE AND E NGINEERING, the IEEE T RANSAC -
UtBa9yVZBJM TIONS ON S EMICONDUCTOR M ANUFACTURING , Transportation Research
[41] Machine Design. (2017). Changing the Future of Warehouses With Part C: Emerging Technologies, Applied Mathematical Modelling, Interna-
Amazon Robots. [Online]. Available: https://www.machinedesign.com/ tional Journal of Production Research, Computers and Industrial Engineering,
motion-control/changing-future-warehouses-amazon-robots and so on.

Authorized licensed use limited to: UNIVERSITAS GADJAH MADA. Downloaded on February 10,2021 at 05:44:28 UTC from IEEE Xplore. Restrictions apply.

You might also like