Optimization of a Modular Drone Delivery System

Jaihyun Lee
Department of Mechanical Engineering
University of Michigan
Ann Arbor, MI, United States
[email protected]

Abstract— Drones have recently become a promising solution avoid obstacles play significant role in the stability and the
for rapid parcel delivery due to advances in battery technology robustness of a drone system [9, 10].
and navigation systems. Drones have inherited limitations in
battery capacity and payload, which make their efficient operation There are few studies related to increasing the operational
and management a critical problem for a successful delivery efficiency of a drone delivery system. Drones have limitations
system. Adopting modularity in the drone design can provide in the delivery range and payload, which makes the efficient
operational benefits to increase overall fleet readiness and reduce operation and management more critical compared to ground
overall fleet size. This paper discusses the potential value of delivery systems. Previously, Traveling Salesman Problem
introducing modular design to a drone delivery system. We (TSP) is suggested to optimize the routes of the drones that show
propose an optimization method for the operation management of substantial savings when drones and trucks are operated together
a fleet of modular delivery drones. This paper presents simulation [11]. The suggested TSP lacks a fleet perspective by focusing on
results that compare the proposed method with existing operation a single drone operation. In another study, a TSP problem is
management methods. The results show that a simple operation formulated for combined truck-drone delivery network systems
management strategy can make a drone delivery system unstable to optimize truck routes, and a clustering method is proposed to
with increasing demand on certain types of modules in the fleet. determine the optimal number of drone launch sites and
The results comparing modular and non-modular drone operation locations [12]. Another study proposes a mixed-integer linear
also prove that the proposed operation management method with programming formulation for a drone delivery fleet to minimize
modular drones can save delivery time and energy consumption
operation costs by optimizing the delivery routes [13]. The focus
during a delivery operation over non-modular drones.
in these studies is on routing rather than a complete management
Keywords—drone delivery system; modularity; optimization; of the entire fleet including scheduling, inventory management,
parcel delivery; dynamic programming; operations management; maintenance and repair.
Introducing modular architectures to the drone design
provides several operational advantages in a drone delivery
I. INTRODUCTION system [14]. A major drawback in the operation of drones is the
Unmanned aerial vehicles (UAVs), or drones, have been limited battery capacity. With lithium-ion batteries used in most
successfully applied to disaster management [1], 3D mapping [2] commercial drones, flight time is approximately 30 minutes, and
and precision agriculture [3]. Recently, drones are considered as non-modular drone is unavailable for delivery for more than one
a promising solution to rapid parcel delivery. Parcel delivery hour when recharging the battery [15]. A modular design with
using drones has many advantages over existing ground vehicle swappable batteries is expected to perform better than drones
delivery using trucks. Delivery time can be reduced since drones with integrated batteries in terms of readiness. Furthermore,
are not interrupted by the established infrastructure such as since a set of drone types with distinct capabilities is necessary
traffic lights, and the volume of traffic. It is also possible to to accomplish various types of delivery tasks, modularity can
reduce the cost of maintenance in a drone delivery system since reduce the overall fleet size thanks to module sharing.
drones are less expensive than ground vehicles and easier to In this paper, we propose an operation management strategy
repair. Various major companies expressed their interest in based on the optimization of modular drone delivery systems.
drone delivery including Amazon with their service “Prime Air” For a given set of modular designs, this strategy determines the
[4], DHL with “Parcelcopter” [5] and Google with “Project schedule and the assignment of modular combinations to meet
Wing” [6]. the delivery demand. One study introduces fuzzy controller [16,
Several research studies on drones have contributed to 17] and develop a reinforcement fuzzy learning scheme for
making a drone delivery system feasible by increasing flight robots playing a differential game operation [18]. Another study
time and improving navigation capabilities. Studies on lithium- is stochastic neighborhood approach is used to optimize shift
ion batteries with high energy density have shown schedule management for tank trucks [19]. We suggest an
improvements on flight time compared to nickel-cadmium and optimization algorithm with a forward-looking approach and
nickel-metal hydride batteries [7]. In addition, a wide range of compare with existing methods such as dynamic programming
different augmentation systems to assist GPS by providing [20] under a randomly generated delivery scenario.
accuracy and integrity have improved the positioning and We apply the proposed operation management strategy to a
navigation of drones [8]. Moreover, technologies to detect and modular and a non-modular drone system to identify potential

978-1-5090-4623-2/17/$31.00 ©2017 IEEE

benefits of modularity in rapid parcel delivery. The results show four independent variables. is given by the demand, and
that modular drone operation with the suggested optimization is determined by both the demands and the drone design.
algorithm can save delivery time and energy consumption,
compared with, not only, modular drone operation based on The four independent variables can be attributed to motor,
existing management algorithms, but also non-modular rotor, carrier, and battery, respectively. We define these
operation based on either management algorithm. components as modules, and use these variables to create
different variants of these modules. The definition of modules
determines the modular architecture of drones while variants
create distinct modular drones with different capabilities. Table
II. MODULARITY IN DRONE DELIVERY I summarizes the module definitions and the associated variables.
In this section, we describe the process used to modularize
a fleet of delivery drones and map the predefined set of modular TABLE I. MODULE DEFINITIONS
drones to the delivery demand. A module in this study is Modules
characterized by the functionality of its components that defines Module 1 Module 2 Module 3 Module 4
its type and a set of variables that differentiates variants of a Carrier Battery Propeller Motor
module type. The delivery demand is represented by mass, Number of Variants N (= 2) N (= 3) N (= 3) N (= 3)
volume, delivery distance and time of the order. These values Variables Volume Capacity Diameter Power
are randomly generated following probability distributions to
define a reasonable scenario. B. Modular Drone Assignment
Mapping a delivery demand to a modular drone can be
A. Modular Architecture
formulated as an optimization problem. Given a modular
We heuristically define a modular architecture for the drone architecture with module types and a set of module variants
fleet based on the independent variables in Equations (1) – (5) for each module type as in Table I, the number of possible
governing the drone motion. While some studies define the modular combinations is ∏ . Physical constrains exist in
modular system architecture using functional or physical practice, which make some of these combinations infeasible. In
decomposition [21, 22], the focus of this paper is not on modular this study, we focus on a case with four module types. We define
design but on the feasibility of modularity in the context of drone the variables describing different module variants as
delivery operation, in this paper. The complete system [ , , , ] = [ , , , ]. For given delivery demand, the
optimization combining design and operation is left to a future constraints that a modular drone must satisfy can be formulated
study. Equation (1) defines an approximate relationship based as:
on Newton’s second law of motion for the horizontal drone
velocity in terms of the total mass , the gravitational ≤ , (6)
acceleration , the total motor thrust , the air density , the
/
drag coefficient , and the effective area of drone . The total ≤ , (7)
mass includes the mass of the drone defined by the sum of the
weights of its modules and the load/parcel as shown in
Eq. (2). is the total number of modules representing one drone. 0.5(2 ) / ≤ , (8)
Equation (3) relates motor thrust to the rotor diameter , and the
motor power , using static thrust calculation and fluid ≤ , (9)
mechanics. Equation (4) assumes that the effective drone area is
related to the carrier volume , which is the space where where is the mass of module , and L is the distance to the
load/parcel are placed. The energy capacity of the battery delivery location specified by a demand, and is the parcel
must be larger than the energy required to complete delivery in
amount of time as shown in Eq. (5). volume. Equation (6) defines the volume constraint for the
parcel to fit into the modular drone. Equations (7) – (8) are the
requirements for the drone to be able to lift the load and hover
= 2 1−( / ) /( ), (1) in the air. Equation (9) is the condition for the battery to provide
enough energy throughout the delivery. To optimize the drone
=∑ + , (2) assignment to the demand, we use the following cost function
in Equation (10) which is the weighted sum of the average power
/
= , (3) consumptions given by the first term and the delivery time given
by the second term. in the cost function is the function that
=
/
, consists of as represented in Equation (1).
(4)
≤ , (5) =∑
( , , )
+
( , , )
(10)
, , , and are the four independent variables related to
The values of the weighting coefficients and define
the physical drone design. Note that the drone mass is not an
the relative importance of energy consumption and delivery time
independent variable since it can be modeled as a function of the
in the optimization. is minimized by selecting a module
combination that satisfies the conditions in (6) – (9). If the A. Simulation Overview
number of feasible modular combinations is small, it is possible This section describes the high-level process flow in the
to find the modular drone that minimizes for each separate simulation of drone delivery. We simulate two separate
delivery demand using enumeration. The physical attributes of scenarios with modular and non-modular drones to identify
each module, , depends on the drone assignment. This drone potential operational benefits of modularity for parcel delivery.
assignment considers the availability of the required modules in The simulation models for two scenarios have some differences
the inventory as the constraints. In this paper, we use [ , ] = since the operation of modular drones requires additional steps
[20, 1] such as assembly and disassembly that are not applicable to non-
modular drones. Another important difference is that non-
TABLE II. ORTHOGONAL ARRAY SET OF NON-MODULAR DRONES modular drones have an integrated battery system while the
Modules battery in modular drones is swappable.
Non-modular
Module 1 Module 2 Module 3 Module 4
Drone
Carrier Battery Propeller Motor
1 2 1 1 3
2 2 1 2 2
3 2 1 3 1
4 2 2 1 2
5 2 2 2 1
6 2 2 3 3
7 2 3 1 1
8 2 3 2 3
9 2 3 3 2

C. Non-modular Drone Assignment

The performance of drone delivery operation is affected by Fig. 1. Operation of non-modular drones
the capabilities of drones in the fleet in non-modular drone. To
observe the effect of number of different types of non-modular
drones on the fleet performance and compare it with that of
modular operation, we define two cases with different number
of types of non-modular drones and compare each operation; an
operation with one type of non-modular drone that can meet all
delivery requirement, and an operation consisting of nine types
of non-modular drones selected by an orthogonal array based on
Taguchi method [23].
In this paper, we assume a non-modular drone consists of the
same “modules” used in modular drones. Unlike modular drones,
Fig. 2. Operation of modular drones
the modules of non-modular drones are fixed and irreplaceable.
The drone for the operation with one type of non-modular Figure. 1 and 2 depict the process flow followed in the two
consists of the second variant of module 1, the third variant of different scenarios. The process starts with randomly generated
module 2, the first variant of module 3, and the first variant of orders defined by mass, volume, travel distance and the time of
module 4. This set of modules are chosen based on the Eqs. (1) the order. A drone assignment is performed following the
– (5) to fulfill the maximum delivery distance requirement process described in Section II-B and the availability of the
explained in ‘Randomized Order’ in Section III-B. required drones/modules is checked. Section III-C elaborates on
Table 2 shows the orthogonal array set used in the operation how this process can be modified by integrating feedback from
with nine types of non-modular drones. Note that 18 non- the inventory. Modular drones undergo an assembly process.
modular drones can be created with (2 3 ) orthogonal When drones are ready, they are dispatched for the delivery
array in the generated case. We assume that carrier is operation. After the delivery, modular drones are disassembled.
standardized to be second variant of module 1, for simplicity, For modular drones, only the battery modules are recharged after
thereby modifying to (3 ) orthogonal array set consisting of disassembly while for non-modular drones, the entire drone is
9 types of non-modular drones in delivery operation. unavailable while its battery is recharged because battery is
integrated.
B. Simulation Assumptions
III. OPERATION OF DRONE DELIVERY The assumptions involved in each step of the simulation
In this section, we describe the details of the simulation process described in the previous section are summarized rest.
model for the drone delivery operation, and suggest an algorithm
to manage the system operation considering the system stability  Randomized Order—We assume that delivery orders do not
and the operation efficiency. exceed the capabilities of the available drones. Amazon
reports that 86% of all Amazon packages weigh 2.3 kg or
below, and the targeted distance is less than 14 km from the
drone station [24]. For the simulation scenario, we use a
 normal distribution with a mean of 1.4 kg and a standard Algorithm 1 Static Drone Assignment
deviation of 0.9 kg for the weight of one parcel and set the Input:
range between 0 kg and 2.3 kg. A normal distribution with a —Time when a drone is requested
mean of 6 km and a standard deviation of 7.5 km is used for — The matrix containing order information
the delivery distance of one delivery order and limit the range
— The vector representing the number of variants for
between 0 km to 13.5 km. This paper considers the volume
each module type
and the order time of one parcel that follows a uniform
distribution with the range between a volume of zero cubic — The vector containing predefined module/drone
meter and a volume of one cubic meter. The total number of assignments
randomly generated orders for each day is 100. — The vector containing the inventory levels, with
elements for the variant of the module
 Drone Assignment—We assume that all modules are
compatible with each other but the assigned drones must Output:
satisfy the constraints in Eqs. (6) – (10). We assume a drone
— The time when a drone is requested
fleet with four module types as shown in Table I with
variants = 2, = = = 3. This paper defines the — A vector containing drones/modules send for
modules as independent variables in Eqs. (1) – (5). Details operation
are explained in Section II-A. — The matrix containing order information

 Availability Check—For a drone to be considered available, // Initialize the set of modules

all corresponding modules must be available in the inventory. =
 Assembly—We assume that the time to assemble a complete // Set to the column of
modular drone is always less than 5 minutes. = ( ,:)
// Iterate over the number of module types
 Operation—Drones can be deployed only during daytime. for ∈ {1, … length( )} do
We assume operation time as a half of one day. At night, // Iterate over the number of variants
drones are not sent for operation but orders are accumulated.
We do not model damage or maintenance needs in this for ∈ {1, … } do
simulation. We also do not model the delivery routing // Check predefined assignment corresponding to
problem. the order
if = then
 Disassembly—All modular drones are disassembled after // Check inventory level is empty
return, i.e., the inventory only contains modules in a modular if = 0 then
drone operation scenario. This assumption is necessary in ←0
practice since modules might need service after operation
// Substitute the ( + 1) column of
and assembly/disassembly time is relatively short compared
to the delivery time. All modules besides the battery are sent ( + 1, : ) =
back to the inventory after disassembly, while battery return ( , , )
modules and regular drones go to the recharging.
 Battery Recharge—Non-modular drones and battery We refer to the static drone assignment as Algorithm 1. This
modules are unavailable during the recharging process. algorithm chooses the optimal drone that minimizes the cost
Battery recharge time is assumed to be one hour in this in Eq. (10) for a given demand. When the required modules are
simulation. not available, the only possible decision is to wait until they
become available with returning drones from delivery operation.
 Inventory—Initial number of non-modular drones and This might not be a preferred approach since it might lead to
modules are predefined. The inventory levels are recorded unnecessary delays in delivery. It is generally possible to have
and used for the drone delivery. alternative drones that can satisfy the constraints in Eqs. (6) –
C. Operation Management (9), and a better management strategy can be developed
considering these alternatives.
This section elaborates on the three methods for operation
management problem that are a preprocessed static drone Dynamic programming is a commonly used method in
assignment, dynamic programming, and an assignment with the operations research for management problems [25]. We can
proposed customized algorithm. All methods can be used in formulate the problem using the same cost function in Eq. (10)
operations with more than two non-modular drones. Since the with following constraint defining module availability,
operation with one type of non-modular drone is a special case
where there is only one selection to be made when assigning ≤ , (11)
drones, only Algorithm 1 is applicable. The drone assignment in
operation receives the feedback from the inventory regarding the in addition to Eqs. (6) – (9). Here, is the inventory level and
availability of the assigned modules. is the requested number of variant of the module .
 Algorithm 3 Custom Drone Assignment the entire time horizon to be known. Therefore, it is not possible
Input: to formulate a general dynamic programming problem over a
— Time when a drone is requested long period. In this paper, we formulate the problem for one day.
— The matrix containing order information We refer to this approach as Algorithm 2. Second, the decision
to assign a drone at a time affects not only the decision and
— The vector representing the number of variants for
inventory levels at the next time, but also that at the time when
each module type
the drones come back. Thus, we do not guarantee the global
— The vector containing predefined module/drone optimality.
assignments
— The vector containing the inventory levels, with To account for the impact of the time delay in the
elements for the variant of the module management decision, we propose a custom optimization
— A matrix containing the returning modules with algorithm. We refer to this method as Algorithm 3. We define
the following cost function,
elements for the variant of the module
= ( + + ) (12)
Output: ( , , ) ( , , )
— The time when a drone is requested where is the difference of the time between when drones are
— A vector containing drones/modules send for requested and when they are actually deployed. Compared with
operation in Eq. (10), an additional term is added here to the objective
— The matrix containing order information function used in Algorithm 2 to penalize the delay in the delivery.
The weight coefficients , , and determine the relative
for ∈ { , + 1} do importance of each term. In this paper, we use [ , , ] =
if = then [20, 1, 1]
COL ← COLUMN( ( )) //assign the k column
We assume that all the drones dispatched at a certain time
← ( , : ) //assign the inventory level eventually return after the delivery without any loss. Therefore,
else the number of returning drones can be predicted at the time they
COL ← COLUMN ( ) + COLUMN( ( + 1)) are sent. Algorithm 3 finds the optimal drones at the time when
//Add the returned modules to the inventory drones are requested ( ). Then, it solves the optimization
← ( ,:) + ( ,:) problem again at the next time ( + ∆t) by assuming all requests
end if are delayed. Since the number of modules in the inventory is
//Solve Dynamic Equation different from that of modules at the time before ( ) because of
for ∈ {1, … COL} do returning drones, the optimal drone to operate changes under the
for ∈ {1, … , (1)} do cost function in Eq. (12). After comparing the values of cost
//Check whether inventory is empty function, we determine whether it is beneficial to deploy drones
if , > 0 then or delay the order to wait for returning drones.
for ∈ {1, … (2)} do In Algorithm 3, not only the given time but also the next time
if , > 0 then is considered. In case of the optimization at the given time,
for ∈ {1, … (2)}do Algorithm 2 is used to solve optimization problem. Since the
inventory level has changed due to returning drones at the next
if , > 0 then time, the constraints have changed with time. Similarly, the
for ∈ {1, … (3)} do optimization problem at the next time can be solved with the
if , > 0 then different constraints using Algorithm 2. Algorithm 3 calculates
//Calculate cost function in (14) and compares the minimum value of the cost function at both
( )← ( , , , , , , ) time, thereby leading to better decisions than dynamic
//Find the modular combination to minimize the sum of programming. This method does not guarantee global optimality
cost function and information of orders but we propose it as a computationally efficient way to account
[VAL( ), IND( ), R( )] ← ( ( ), , ) for time delay.
end for
if VAL( ) > VAL( + 1) then
← IND( + 1) ∧ ← ( + 1) IV. RESULTS
else In this section, we show the simulation results corresponding
← IND( + 1) ∧ ← ( ) to the three different algorithms discussed in this paper. We
return ( , , ) compare the results to justify the need for a custom algorithm in
the operation of modular drone delivery. We run a 5-day
simulation for each algorithm for the modular and non-modular
There are several issues when implementing dynamic drone fleet under the same conditions. The trend of accumulated
programming for this problem. First, dynamic programming is orders after one-day simulation is shown to demonstrate that
computationally expensive when the number of state variables
is large. Also, dynamic programming requires the demand over
modular and non-modular drone fleet operation are both stable nine non-modular drones has a similar tendency with the
with Algorithm 2 and 3. modular drone operation in Figure 6.

Fig. 3. Battery modules in the inventory with Algorithm 1. Fig. 4. Battery modules in the inventory with Algorithm 2.

Modular drone fleet simulation starts with an initial

inventory of ten modules for each variant of module 2, 3, and 4,
and fifteen modules for each variant of module 1 to have thirty
modules for each type initially. The non-modular simulation
with one drone has an initial inventory of thirty non-modular
drones, and the non-modular simulation with orthogonal array
has that of three drones for each non-modular drone, thereby
meeting the same initial total number of modules as modular
drone simulation has.
Figures 3–5 present the inventory levels corresponding to the
battery modules for 5 days with three different algorithms. The
results for battery modules are the most critical since it takes
more time for the battery modules to become available after the
delivery operation unlike other modules. As seen in all three
figures, the inventory levels for some battery module variants
reach zero several times throughout the time horizon. It shows
that same orders are delayed during these intervals. As seen in
Figure 3, the first variant is out of stock and other variants have
barely been used, implying that the orders requiring the first
Fig. 5. Battery modules in the inventory with Algorithm 3.
battery module variant are delayed and accumulated. Other
variants are actively used in Figures 4 and 5, showing that drone
delivery operation with either Algorithm 2 or Algorithm 3
outperforms Algorithm 1 in terms of timing. Figure 8 indicates that the drone delivery operation with one
Figure 6 shows the trend of initial orders after one-day of type of non-modular drone is stable when operated by Algorithm
modular drone operation, suggesting that the system is unstable 1, on contrary to the operation with nine types of non-modular
when managed with Algorithm 1. On the other hand, Algorithm drones. The main reason is that different types of drones are not
2 and Algorithm 3 achieve system stability by keeping these allowed to replace one another in Algorithm 1, causing delays in
orders at manageable levels of ‘Night-time’ represents the orders delivery orders when nine types of drones are used.
accumulated throughout the night when drones do not operate. Figure 9 compares the average energy consumption of non-
This result indicates that modular drone operation has finished modular and modular drones when each algorithm is applied.
the all orders during the day-time when operated by Algorithm Modular drone operation with Algorithm 3 takes approximately
2 or 3. As seen in Figure 7, the drone delivery operation with 8.6 % less energy than one with Algorithm 2. In case of the non-
modular drone operation with nine types, the energy Comparing modular and non-modular drone operations, the
consumption does not have significant difference between with average energy consumption of modular drone operation is
Algorithm 2 and 3. In addition, the energy consumption of the lower by 4 percent, indicating no significant improvement in
non-modular operation with nine types has no improvement terms of energy consumption.
compared that of the operation with one type of non-modular
drone.

Fig. 6. Initial orders for each day after modular drone fleet operation.

Fig. 9. Averaged energy consumption percentage of drones, normalized based

on single non-modular drone operation.

Fig. 7. Initial orders for each day after non-modular drone fleet operation with
nine types of drones.

Fig. 10. Average delivery time percentage of drones, normalized based on

single non-modular drone operation.

We compare the average delivery time in Figure 10.

Comparing non-modular drone operations, the operation with
nine types takes less time to deliver the packages than the
operation with one type, indicating that the proper management
of drone delivery operation with nine non-modular drone can
reduce delivery time.
When Algorithm 3 is applied to modular drone operation, the
Fig. 8. Initial orders for each day after non-modular drone fleet operation with package can be delivered 23.5 % faster than non-modular drone
one type of drone. operation with one variance, 14.5 % faster than non-modular
drone operation with nine variances, and 4.4 % faster than [7] S. Moore and P. Schneider, "A Review of Cell Equalization Methods for
modular drone operation where Algorithm 2 is implemented. Lithium Ion and Lithium Polymer Battery Systems," SAE Technical
Paper, doi:10.4271, 2001.
This result shows that the proposed algorithm outperforms
[8] National Coordination Office for Space-Based Positioning, Navigation,
existing methods due to the forward-looking nature. Modular and Timing. (2015, Dec.) GPS.gov: augmentation systems. [Online].
drone operation takes less time to deliver packages regardless of Available: http://www.gps.gov/systems/augmentations/
the operation algorithm, showing that modularity in drone [9] Xin-Zhong Peng, Huei-Yung Lin and Jyun-Min Dai, "Path planning and
design has the potential to improve delivery time. obstacle avoidance for vision guided quadrotor UAV
navigation", Control and Automation (ICCA) 2016 12th IEEE
International Conference on, pp. 984-989, 2016.
[10] F. Andert, F. Adolf, L. Goormann and J. Dittrich, "Mapping and path
V. CONCLUSION planning in complex environments: An obstacle avoidance approach for
This paper discusses the potential value of introducing an unmanned helicopter", Robotics and Automation (ICRA) 2011 IEEE
International Conference on, pp. 745-750, 2011.
modularity to a drone delivery system. We propose a forward-
[11] N. Agatz, P. Bouman, and M. Schmidt, “Optimization Approaches for the
looking fleet operation management strategy to improve the Traveling Salesman Problem with Drone”, ERIM Report Series
performance of drone delivery. We compare the proposed Reference No. ERS-2015-011-LIS, June 27, 2016.
method with a stable management strategy and Dynamic [12] S. M. Ferrandez, T. Harbison, T. Weber, R. Sturges. “Optimization of a
Programming. We apply all three management strategies to both truck-drone in tandem delivery network using k-means and genetic
modular and non-modular drone delivery operations to identify algorithm”, Journal of Industrial Engineering and Management, vol. 9, no
the potential benefits of modular design for drone delivery. 2, pp. 374-388, 2016.
[13] K. Dorling, J. Heinrichs, G. G. Messier, and S. Magierowski, “Vehicle
The results in this paper show that the drone delivery system Routing Problem for Drone Delivery”, IEEE Transactions on System Man
is unstable with the static management strategy for both modular and Cybernetic: Systems, 2016.
and non-modular drone operations. Although the drone delivery [14] D. Campagnolo, and A. Camuffo, “The Concept of Modularity in
system is stable with both dynamic programming and the Management Studies: A Literature Review”, International Journal of
proposed method, the latter can save a delivery time and energy. Management Reviews, vol. 12, pp. 259–283, 2010.
Also, the results show that modular drone delivery system has [15] A. Abdilla, A. Richards, and S. Burrow, “Power and endurance modelling
of battery-powered rotorcraft”, In: Intelligent Robots and Systems (IROS),
potential to improve delivery time compared to a non-modular 2015 IEEE/RSJ International Conference on. IEEE, p. 675-680, 2015.
system. [16] Igor Škrjanc, Sašo Blažič, and Osvaldo Agamennoni, “Identification of
In this paper, we compare a fully modular system with a non- dynamical systems with a robust interval fuzzy mode”, Automatica, vol.
41, chap. 2, pp. 327-332.
modular drone system. The results indicate that a semi-modular
[17] Precup, Radu-Emil, Stefan Preitl, and Pter Korondi, “Fuzzy controllers
system with a replaceable battery is an interesting case to with maximum sensitivity for servosystems.”, IEEE Transactions on
analyze for future work. We also define the modules for our Industrial Electronics, vol. 54, chap. 3, pp. 1298-1310.
analysis heuristically. In a future study, we plan to combine the [18] S.N. Givigi, H.M. Schwartz, and Xiaosong Lu, “A Reinforcement
design decisions on the module definitions with the operation Learning Adaptive Fuzzy Controller for Differential Games”, Journal of
management strategy under an integrated framework. Intelligent & Robotics Systems, 59.1, pp 3-30, 2010.
Performing a long-term analysis with such an integrated [19] I.P. Solos, I.X. Tassopoulos, and G.N. Beligiannis, “Optimizing Shift
framework can be used to investigate the profitability of a Scheduling for Tank Trucks Using an Effective Stochastic Variable
Neibourhood Approach”, vol. 14, no. 1, pp. 1-26, 2016.
modular approach in drone delivery with a business perspective.
[20] D. P. Bertsekas, “Dynamic programming and optimal control”, Belmont,
MA: Athena Scientific, vol. 1, no. 2, 1995.
[21] A. E. Bayrak, A. X. Collopy, B. I. Epureanu and P. Y. Papalambros, "A
REFERENCES computational concept generation method for a modular vehicle fleet
design", 2016 Annual IEEE Systems Conference (SysCon), Orlando, FL,
[1] S. M. Adams, and C. J. Friedland. “A survey of unmanned aerial vehicle
pp. 1-8, 2016.
(UAV) usage for imagery collection in disaster research and management.”
In Proc. of 9th Int. Workshop on Remote Sensing for Disaster Response. [22] Chun-Che Huang and A. Kusiak, "Modularity in design of products and
2011. systems," in IEEE Transactions on Systems, Man, and Cybernetics - Part
A: Systems and Humans, vol. 28, no. 1, pp. 66-77, 1998.
[2] F. Nex, and F. Remondino. "UAV for 3D mapping applications: a review"
Applied Geomatics, vol. 6, no. 1 pp. 1-15, 2014. [23] W.H. Yang and Y.S. Tarng, “Design optimization of cutting parameters
for tuning operations based on the Taguchi method”, Journal of materials
[3] C. Zhang, and J.M. Kovac, “The application of small unmanned aerial
processing technoology, cahpter 84.1, pp. 122-129.
systems for precision agriculture: a review”, Precision Agriculture (2012),
vol. 13, Issue 6 , pp 693-712, 2012. [24] D. Bamburry. "Drones: Designed for Product Delivery", Design
Management Review, vol. 26, no. 1, pp. 40-48, 2015.
[4] Amazon.com, Inc. Amazon Prime Air. [Online]. Available:
https://www.amazon.com/primeair [25] Facci, Andrea Luigi, Luca Andreassi, and Stefano Ubertini,
“Optimization of CHCP (combined heat power and cooling) systems
[5] Deutsche Post DHL Group. DHL parcelcopter 3.0. [Online]. Available:
operation strategy using dynamic programming.”, Energy, vol. 66, pp.
http://www.dpdhl.com/en/media_relations/specials/parcelcopter.html
387-400, 2014.
[6] Google Inc. Project Wing. [Online]. Available:
https://www.solveforx.com/projects/wing/