Professional Documents
Culture Documents
Abstract—A new variant of the two-echelon routing problem is delivery vehicles are limited by the established infrastructure
investigated, where the truck and the drone are used to cooper- and geographical obstacles, such as roads and rivers, which
atively complete the deliveries of all parcels. The truck not only would vastly add to the distribution time and operating costs.
acts as a tool for parcel delivery but also serves as a moving depot
for the drone. The drone can carry several parcels and take off A large number of ground vehicles also bring about atmo-
from the truck, while returning to the truck after completing the spheric pollution and traffic jam in urban areas, which is
delivery. The energy consumption model for the routing process certainly against the public wish for the environmental pro-
of the drone is analyzed, when it is utilized to deliver multiple tection. These problems drive many researchers to look for
parcels. A two-stage route-based modeling approach is proposed more efficient delivery ways and thus show great interests in
to optimize both the truck’s main route and the drone’s adjoint
flying routes. A hybrid heuristic integrating nearest neighbor and drone delivery.
cost saving strategies is developed to quickly construct a feasible Delivery by unmanned aerial vehicles (UAVs), or drones,
solution. The simulated annealing algorithm is integrated with has been feasible due to the rapid development of automation
Tabu search, to improve the quality of the solution as well as and artificial intelligence technologies. In general, drones can
the search efficiency. Random instances at different scales are significantly decrease the delivery cost and time, as they can
used to test the performance of the proposed algorithm. A case
study based on the practical road network in Changsha, China, fly directly to recipients with being less blocked by obstacles
is presented, through which the sensitivity analysis is conducted and not considering ground traffic condition. Besides, drones
with respect to some critical factors. are more environment-friendly with lower power consump-
Index Terms—Heuristic, simulated annealing (SA) algorithm, tion and less air pollution. Thus, realizing these advantages,
truck and drone, two-echelon routing, vehicle routing. many companies have investigated drone delivery problems.
In addition to the first few companies applying UAVs, such
as Amazon, DHL, and Google [2], [3], JD has tried the drone
I. I NTRODUCTION delivery in several provinces in China. Alibaba Group and
VER the years, the full-blooming E-commerce has Baidu Company also utilize the parcel copter for take-out.
O incentivized the express delivery to skyrocket. It is
reported that over 600 million parcels need to be delivered only
However, depending upon the characteristics of drones, it may
not be optimal to deliver-by-drone to all customers. Restricted
in one day of the double-eleven shopping festival in China [1]. by short endurance and low load capacity, drones have limited
Facing such a big challenge to delivery so many parcels, tradi- flight time and can only carry small packages.
tional distribution based on ground vehicles can hardly satisfy Considering these practical difficulties, both trucks and
the customers for their increasing requirement on shorter drones have their own limitations and advantages, which are
delivery time, as they expect to receive their parcels in one summarized in Table I. Nevertheless, if they work together,
or two days after they complete the online orders. Ground there are several advantages due to their strong complementar-
ity. Primarily, the effective delivery range of drone is enlarged,
Manuscript received November 12, 2019; accepted January 17, 2020. Date when the ground vehicle acts as a moving depot for car-
of publication February 10, 2020; date of current version November 18, 2021. rying the drone. With longer traveling distance and larger
This work was supported in part by the National Natural Science Foundation
of China under Grant 71771215, and in part by the Natural Science load capacity, ground vehicles can serve dual roles as both
Fund for Distinguished Young Scholars of Hunan Province under Grant a mobile depot and a delivery resource. Furthermore, drones
2018JJ1035. This article was recommended by Associate Editor M. K. Tiwari. are less limited by the ground traffic and can perform better
(Corresponding author: Jianmai Shi.)
Yao Liu, Zhong Liu, and Jianmai Shi are with the Science and in some areas hard for trucks to reach, such as some places
Technology on Information Systems Engineering Laboratory, College of with traffic congestion or inconvenient transportation. As for
System Engineering, National University of Defense Technology, Changsha the drone’s short endurance, it can be reused with replacing
410073, China (e-mail: liuyao13@nudt.edu.cn; liuzhong@nudt.edu.cn;
jianmaishi@gmail.com). the battery or charging on the ground vehicles. Thus, the com-
Guohua Wu is with the School of Traffic and Transportation bination of drone and truck has promised a bright prospect.
Engineering, Central South University, Changsha 410075, China (e-mail: Some companies have applied this interesting approach.
guohuawu@csu.edu.cn).
Witold Pedrycz is with the Department of Electrical and Computer Mercedes-Benz released a conceptual car called “vision van”
Engineering, University of Alberta, Edmonton, AB T6R 2V4, Canada (e-mail: in September 2016 [4]. Equipped with a fully automated cargo
wpedrycz@ualberta.ca). loading system, this van deploys the delivery drone to get
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TSMC.2020.2968839. parcels to customers quickly. As described, it can significantly
Digital Object Identifier 10.1109/TSMC.2020.2968839 reduce delivery time, and increase efficiency. In addition, UPS
2168-2216
c 2020 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: Indian Institute Of Technology (Banaras Hindu University) Varanasi. Downloaded on February 06,2023 at 20:29:36 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: TWO-ECHELON ROUTING PROBLEM FOR PARCEL DELIVERY BY COOPERATED TRUCK AND DRONE 7451
TABLE I
C OMPARISON OF D RONE AND G ROUND V EHICLE FOR D ELIVERY
Authorized licensed use limited to: Indian Institute Of Technology (Banaras Hindu University) Varanasi. Downloaded on February 06,2023 at 20:29:36 UTC from IEEE Xplore. Restrictions apply.
7452 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 12, DECEMBER 2021
drones equipped with image sensors. Persistent intelligence, refueling unmanned ground vehicle (UGV) for successful
surveillance, and reconnaissance (PISR) routing problem are mission completion. Savuran and Karakaya [21] proposed
also considered by minimizing the time of delivering the a route optimization method for a carrier-launched UAV. In
collected data to the control station [14]. this problem, the UAV visits the targets to execute the mis-
In the works mentioned above, the endurance capacity is set sion while the carrier keeps on moving on its own route.
as a fixed value. However, based on the actual situation, the Furthermore, a similar routing problem is investigated [22]
effect of the payload weight on the energy consumption of the and the difference is that the GV travels on the road network
drone during the flight should not be ignored. There are several while its UAV visits the targets beyond the road. Although
studies mentioned the relationship between the endurance and these problems all investigate the routing problem in the com-
payload weight. Mufalli et al. [15] considered the surveillance bination of trucks and drones, the payload of the drone stays
mission in which the available flight time would be reduced the same while visiting multiple targets in one route, in which
after attaching a sensor. Dorling et al. [8] proposed two vehicle the relationship between payload and energy consumption
routing problem (VRP) variants for drone delivery. Calculated would not be involved.
with a linear approximation function, the energy consump- From the above review, we can see that a new variant of
tion varies linearly with payload and battery weight. Besides, the classical VRP has not been studied. Different from the
Song et al. [16] mainly focus on the characteristics of the traditional two-echelon VRP with only trucks involved [23],
UAV logistics system and discuss the effect of cargo weight and also different from the TSP-D with only one customer’s
on flight ability. However, in these problems, only drones are parcel delivered by the drone [2], [3], [6], this article stud-
applied for the delivery, and the truck is not included. ies the two-echelon routing problem for cooperated truck and
Since the combination of drone and truck can greatly drone, especially when multiple parcels can be delivered in
improve the delivery efficiency, we are aware of some works one drone’s flying route and the effect of varying payload
investigating the problem. To overcome the flight-range limita- on energy consumption is considered. Therefore, this new
tion of the drone, Kim and Moon [17] proposed a truck-drone problem is studied in this article. The energy consumption
system, in which a fixed drone station is established to store model of the drone during the parcel delivering process is
drones and charging devices, and a truck is applied to con- analyzed, and a mathematical model is developed to optimize
nect the drone station and the far package distribution center. both the truck and drone’s routes in the next section.
With the truck acting as the mobile station for the drone,
Murray and Chu [2] introduced the problem, flying sidekick III. P ROBLEM D ESCRIPTION AND M ODEL D EVELOPMENT
traveling salesman problem (FSTSP), which proposes a mixed A. Definitions and Notations
integer liner programming formulation and a heuristic. In
detail, the heuristic is based on the idea of truck first, drone In order to facilitate the model formulation, several defini-
second (TFDS), in which they first construct a truck-only tions are first introduced as follows.
route and iteratively replace the truck nodes by drone nodes Directed Main-Route: It is a closed path traveled by the
to reduce the objective value. Recently, this problem is also truck starting from the depot for visiting some customer nodes
solved by some variable neighborhood heuristics [18], which and returning to the depot, e.g., (0,1,2,3,0) shown in Fig. 1(a).
perform much better on the large-scale instances. There is Directed Subroute: It is path segment of a directed main-
also a slightly different problem—called traveling salesman route. For example, there are ten directed subroutes in the
problem with drone (TSP-D) [6], in which the drone has to directed main-route, (0,1,2,3,0), in Fig. 1(a), which are (0,1),
follow the same road network as the truck. This problem is (1,2), (2,3), (3,0), (0,1,2), (1,2,3), (2,3,0), (0,1,2,3), (1,2,3,0),
also modeled as a MILP formulation and solved by a heuris- and (0,1,2,3,0).
tic based on local search. Afterwards, some variants of this Adjoint Subroute: It is a directed path, including one or
problem are also explored. Ha et al. [3] introduced a time span several customer nodes that can be visited by the drone when
into the TSP-D, which the drone and the truck can wait for the truck travels a corresponding subroute. For example, in
each other with the limited waiting time. Besides, two heuris- Fig. 1(a), (1,4,5,3) is an adjoint subroute of directed subroute
tics of two concepts: 1) drone first, truck second (DFTS) and (1,2,3), which means the drone takes off from the truck at
2) TFDS are designed. Yurek and Ozmutlu [7] also designed node 1, conducts the delivery tasks of nodes 4 and 5, and
a decomposition-based iterative algorithm. After determining then returns to the truck at node 3. Also, subroutes (1,4,3)
the truck route in the first stage, a mixed-integer linear pro- and (1,5,3) are also potential adjoint subroutes of directed
gramming model is solved to optimize the drone route and fix subroute (1,2,3).
the truck routing at the same time. However, in these works, The notations used in the problem description and model
only one customer’s parcel can be arranged on the drone in development are presented in Table II.
each flying route.
Cooperation between drone and truck is also widely used B. Problem Description
for surveillance and mapping. There is a research exploit- In the 2E-RP-T&D, one truck takes one drone to corporately
ing the cooperative air and ground surveillance, in which complete the delivery of all parcels. In this situation, the capac-
the framework and algorithm are proposed for search and ity of truck on load and endurance are usually assumed to be
localization [19]. Additionally, with the restriction of the sufficient for all parcels [2], [3]. However, the payload capac-
UAV’s fuel capacity, Maini and Sujit [20] considered the ity of the drone is limited and known. The problem aims at
Authorized licensed use limited to: Indian Institute Of Technology (Banaras Hindu University) Varanasi. Downloaded on February 06,2023 at 20:29:36 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: TWO-ECHELON ROUTING PROBLEM FOR PARCEL DELIVERY BY COOPERATED TRUCK AND DRONE 7453
TABLE II
N OTATION AND I TS I MPLICATIONS
Authorized licensed use limited to: Indian Institute Of Technology (Banaras Hindu University) Varanasi. Downloaded on February 06,2023 at 20:29:36 UTC from IEEE Xplore. Restrictions apply.
7454 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 12, DECEMBER 2021
D. Model Development
When multiple parcels are delivered by the drone in a sin-
gle flight, the model development for the two-echelon rout-
ing problem becomes quite difficult. Thus, to offer some
more precisely understanding on the problem definition, we
(b) presented a two-stage route-based mathematical formulation.
Fig. 3. Illustration of the energy consumption process for the drone in
Given the network G = (N, E), we assume that all possible
a subroute with 3 parcels. (a) Changes of the battery power in the flying directed main-routes of the truck can be enumerated and let R
process. (b) Changes of the battery remaining capacity in the flying process. be the set of these main-routes. Each route r ∈ R starts from
the depot V0 = {0}, visits one or several customers in V, and
as follows: return to the departure, e.g., (0,1,2,3,4,5,0) in Fig. 2. Although
(wd + Gi )vij there is no limitation on the load capacity and the endurance
pij = +e (1) of the truck, each route r ∈ R should satisfy two constraints.
370ηγ
One is that the indegree and outdegree of any node (including
where η is the conversion efficiency of the motor, γ is the lift customer nodes and depot) in r must both equal to 1, and
ratio, and e is the energy loss of the drone battery [24]. the other is that no subloop exists in route r. For any directed
Then the flying time from customer i to customer j and the main-route r (r ∈ R), its cost can be calculated and noted as cr .
corresponding energy consumption is as shown When route r (r ∈ R) is selected for the truck, let Vr be
the set of customers that are visited by the truck, then all cus-
dij tomers in Vr must be served by the drone, which ensures that
1
tij = ds (2) all parcel would be delivered to its corresponding customer.
0 vij
dij Also, we assume that all directed subroutes in a given main-
(wd + Gi )vij 1 route r (r ∈ R) can be enumerated and let Sr be the set of all
Fij = pdt = +e ds. (3)
370ηγ 0 vij subroutes in main-route r.
For the subroute l (l ∈ Sr ) in route r, we assume that all
When the brand and type of the drone is predetermined, feasible adjoint subroutes for the drone can be enumerated and
the value of all the related parameters in (1)–(3) are known. the set is denoted as Rrl . Each adjoint subroute m (m ∈ Rrl )
Usually, the drone has a maximum power, P, due to the lim- starts from the first customer of the corresponding subroute l
itations of battery and motor. In order to obtain the highest and ends at the last customer of l after visiting one or several
delivery efficiency, we assume that the drone would maintain customers in Vr . Let Vrlm be the set of customer nodes that are
its maximum power during the flight, which means served by the drone in adjoint subroute m (m ∈ Rrl ) and Erlm
be the set of arcs in adjoint subroute m (m ∈ Rrl ). Not only the
pij = P. (4) consumed energy in subroute m should not exceed the maxi-
mum payload of the drone battery but also the total weight of
Thus, based on the above assumption, the flying speed and all corresponding parcels must be within the allowable load
time from customer i for customer j can be estimated as range of the drone
370ηγ (P − e) Fij ≤ P ∀m ∈ Rrl (8)
vij = (5) (i,j)∈Erlm
wd + Gi
dij (wd + Gi ) wi ≤ WH ∀m ∈ Rrl . (9)
tij = . (6)
370ηγ (P − e) i∈Vrlm
Authorized licensed use limited to: Indian Institute Of Technology (Banaras Hindu University) Varanasi. Downloaded on February 06,2023 at 20:29:36 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: TWO-ECHELON ROUTING PROBLEM FOR PARCEL DELIVERY BY COOPERATED TRUCK AND DRONE 7455
According to the energy consumption model, the drone’s The objective function (14) minimizes the total cost given
traveling cost of adjoint subroute m can be calculated and by two main components, which are the truck’s traveling cost
denoted as crlm . and the drone’s flying cost. zr ∗ is obtained through solving
Let xr ∈ {0, 1} be a binary variable equal to 1 if directed Model 1. Constraints (15) ensure that only one truck route
route r (r ∈ R) is selected for the truck, and 0 otherwise. can be selected. Constraint (16) define the decision variables.
Given a truck route r ∈ R and a directed subroute l ∈ Sr , For some directed main-route r (r ∈ R), there may be no
let yrlm ∈ {0, 1} be a binary variable equal to 1 if adjoint feasible solution can be obtained in Model 1, and we let zr ∗
subroute m (m ∈ Rrl ) is selected for the drone, and 0 otherwise. to be +∞ in this situation.
Furthermore, given two customers i, j (i, j ∈ V0 ∪ V) and an We can see that the 2E-RP-T&D can be solved by utilizing
adjoint subroute m (m ∈ Rrl ), let aijrl ∈ {0, 1} be a binary Models 1 and 2. First, all the possible directed main-routes of
parameter equal to 1 if arc (i, j) ∈ Erl and 0 otherwise, let the truck should be enumerated to obtain set R. For each main-
birlm ∈ {0, 1} be a binary parameter equal to 1 if i ∈ Vrlm and route r (r ∈ R), all the adjoint subroutes should be enumerated
0 otherwise. and the optimal subroutes of the drone can be obtained by
Based on the above assumptions and definitions, we develop solving Model 1. Then the optimal solution for minimizing
a two-stage route-based mathematical formulation for 2E- the overall cost of the truck and drone can be calculated by
RP-T&D. In the first stage, the model for optimizing the Model 2.
drone’s adjoint subroutes for a given main-route of the truck is
proposed, which can be used to calculate the minimum cost of E. Enumeration Algorithm Based on Branch-and-Cut
the drone for all the possible main-routes in R. In the second In order to verify the model, we construct an enumeration
stage, the overall cost for both truck and drone is optimized. algorithm based on branch-and-cut. The process of the algo-
1) Model 1 (Optimizing the Drone’s Adjoint Subroutes): rithm refers to the established model. Specifically, the idea of
For any given truck route r (r ∈ R), the sets, Er , Vr , and Vr , branch-and-cut is added to improve the enumeration efficiency.
are determined. On this basis, all possible and feasible drone Step 1 (Configure the Set R of Truck Main-Routes): Suppose
adjoint subroutes corresponding to route r can be calculated, that there are n customers (n = |V|), and the number of cus-
that is set Rrl and Sr can also be obtained. Then the minimum tomers with heavy or big parcels that must be delivered by
cost for the drone’s adjoint subroutes to visit all nodes in Vr truck is nmt (nmt ≤ n). The number of customers covered in
can be calculated through Model 1, which is formulated as the truck route should not be less than nmt . Therefore, con-
follows: sidering the direction of truck routes, the number of possible
Min zr = crlm yrlm (10) directed
truck routes can be calculated as
nmt +1 n−nmt n
l,m 0
Cn−n Anmt
+C 1
A + · · · +C A ×2
mt n mt n−n mt nmt +1 n−n mt n
s.t. aijrl yrlm ≤ 1 ∀i, j ∈ Er (11) mt
n−n
l,m = i
Cn−n Anmt +i × 2. (17)
mt nmt +i
birlm yrlm = 1 ∀i ∈ Vr (12) i=0
l,m i
Cn−n mt
means that i nodes (0 ≤ i ≤ n − nmt ) are selected
yrlm ∈ {0, 1} ∀m ∈ Rrl , l ∈ Sr . (13) among the customers that can be delivered by drone, and
+i
The objective function (10) minimizes the cost of the drone then nmt + i nodes can form Annmt mt +i
× 2 directed main-routes.
for delivering all the parcels that are not delivered by the Equation (17) is also the number of routes in R.
truck in route r. Given a truck route r(r ∈ R), the con- Step 2 (Cut the Set R for Drone Adjoint-Routes in Step 3):
straint (11) ensures that each path segment of truck route r Since the truck cost is only related to the travel distance, we
can only correspond to at most one adjoint subroute, which can start without considering the truck direction, and decide
avoid intersection among drone routes. Constraint (12) impose the direction when constructing drone routes later. Besides,
that all customer nodes in Vr must be visited exactly once. when the truck route contains all nodes in V, there is no cus-
Finally, constraint (13) define the decision variables. tomer left to be served by the drone. Thus,
themt −1
set R can be cut
2) Model 2 (Optimizing the Truck’s Directed Main-Route): into a subset Rsub , which only contains n−ni=0
i
Cn−n Anmt +i
mt nmt +i
To solve the problem, we have to enumerate all the possible truck routes.
directed main-routes for the truck, and find the optimal result Step 3 (Arrange Rest Customers for Each Truck Main-
with the minimum total cost for both the truck and the drone, Route): For a truck route r(r ∈ R) containing nmt + i nodes,
in which the drone’s optimal cost for each truck route r, zr ∗ , is there are
calculated by solving Model 1. The model for optimizing the (nmt + i + 1) + (nmt + i) + · · · + 2 + 1
truck’s directed main route is noted as Model 2 and presented = (nmt + i + 2)(nmt + i + 1)/2 (18)
as follows:
subroutes, configuring the set Sr . Then the rest n − nmt − i
Min Z = cr xr + xr z∗r (14) customers can be allocated to these subroutes, and theoretically
r r generate (nmt + i + 2)(nmt + i + 1)(n − nmt − i/2 schemes.
s.t. xr = 1 (15) However, many infeasible situations are included, caused by
r three constraints: 1) intersection of drone routes; 2) capacity
xr ∈ {0, 1} ∀r ∈ R. (16) of drone payload; and 3) capacity of drone endurance.
Authorized licensed use limited to: Indian Institute Of Technology (Banaras Hindu University) Varanasi. Downloaded on February 06,2023 at 20:29:36 UTC from IEEE Xplore. Restrictions apply.
7456 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 12, DECEMBER 2021
TABLE III
S UPPOSED PARAMETERS W HEN V ERIFYING THE M ODEL
Fig. 5. Illustration of the solution for six customer nodes. (a) Truck-only tour
solution solved by NNS. (b) Truck & drone solution by cost saving replace.
Authorized licensed use limited to: Indian Institute Of Technology (Banaras Hindu University) Varanasi. Downloaded on February 06,2023 at 20:29:36 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: TWO-ECHELON ROUTING PROBLEM FOR PARCEL DELIVERY BY COOPERATED TRUCK AND DRONE 7457
ing node of a drone’s adjoint subroute, is added at the end of the flying cost of the subroute (—1 − 2), c 2-3··· is the flying
D
this adjoint subroute. As illustrated in Fig. 6(b), Customer 1 is cost of the subroute (2 − 3—) and cD ···1-2-3··· is the flying cost
delivered by the drone while customers 2 and 3 are delivered of the subroute (—1 − 2 − 3—).
by the truck in sequence. At this case, the feasibility should be Based on the above four conditions, check all the nodes
judged whether the drone can serve Customer 2 after finishing visited by the truck, calculate the saved cost for every possible
the delivery of Customer 1. Besides the limitation of payload, node and replace the one with maximum cost saving by drone
energy consumption of the whole route needs to be recalcu- visiting. Repeat this evaluation and replacing step until no
lated according to the method in Section III-C. The same check more positive cost saving can be found, which means the total
would also be conducted for the following two conditions. If cost cannot be decreased through changing the delivery ways
the replacement in condition 2 is feasible, the saved cost can of customers.
be calculated as (cT42 + cT23 + cD ···1-2 ) − (c43 + c···1-2-3 ), where
T D 3) Main Optimization Procedure: The main procedure for
c ···1-2 is the flying cost of the original subroute (—1 − 2)
D the HH is shown in Algorithm 1. At first, according to the
and cD ···1-2-3 is the flying cost of the subroute with Customer actual road distance, the truck is assigned to leave from the
3 added. depot and choose to serve the nearest customer every time
In condition 3, a truck visiting node, which is the taking until all the parcels haven been delivered. When the truck
off node of a drone’s adjoint subroute, is added at the start of returns to the depot, a truck-only route is generated (line 1).
this adjoint subroute. As illustrated in Fig. 6(c), Customer 2 is Then loops would be executed following the sequence of the
visited by the truck and is the taking off node of the drone. If customer list in this main route. In every loop, every cus-
the parcel in Customer 2 can be delivered by drone and added tomer node would be judged whether they can be served by
to the start of the adjoint subroute, the saved cost can be calcu- drone. We calculate the saved cost if it can be accessed by
lated as (cT 12 + cT 24 + cD 2-3··· ) − (cT 14 + cD 1-2-3··· ), where the drone (line 3). If one or more customers can be found,
cD 2-3··· is the flying cost of the original subroute (2−3—) and choose the most cost-saving customer and adjust two-echo
cD 1-2-3··· is the flying cost of the subroute with node 1 added. routes to change the delivery way of the chosen customer
In condition 4, two drone’s adjoint subroutes are merged, (line 5). If not, end the loop (line 7) and output the solu-
which is illustrated in Fig. 6(d). It can be seen that, in Fig. 6(d), tion, which is a two-echelon route for the truck and drone
Customer 2 is visited by the truck, which is the landing node (line 10).
Authorized licensed use limited to: Indian Institute Of Technology (Banaras Hindu University) Varanasi. Downloaded on February 06,2023 at 20:29:36 UTC from IEEE Xplore. Restrictions apply.
7458 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 12, DECEMBER 2021
Authorized licensed use limited to: Indian Institute Of Technology (Banaras Hindu University) Varanasi. Downloaded on February 06,2023 at 20:29:36 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: TWO-ECHELON ROUTING PROBLEM FOR PARCEL DELIVERY BY COOPERATED TRUCK AND DRONE 7459
Fig. 8. Sketch map for 2-exchange. (a) Exchange one truck node with drone
take off or land with another drone node. (b) Exchange one truck node with
drone take-off or land with another truck node without drone take-off or land.
Fig. 7. Sketch map for reinsertions of nodes visited by the truck. (a)
Reinsertion of node 3 has no effect on the drone route. (b) Reinsertion of
node 2 does not change the direction of drone route. (c) Reinsertion of node
2 reverses the direction of drone route.
Authorized licensed use limited to: Indian Institute Of Technology (Banaras Hindu University) Varanasi. Downloaded on February 06,2023 at 20:29:36 UTC from IEEE Xplore. Restrictions apply.
7460 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 12, DECEMBER 2021
TABLE IV TABLE V
PARAMETERS OF T RUCK AND D RONE PARAMETERS OF I NSTANCES
Authorized licensed use limited to: Indian Institute Of Technology (Banaras Hindu University) Varanasi. Downloaded on February 06,2023 at 20:29:36 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: TWO-ECHELON ROUTING PROBLEM FOR PARCEL DELIVERY BY COOPERATED TRUCK AND DRONE 7461
TABLE VI
R ESULTS OF 30 R ANDOM I NSTANCES IN T HREE S CALES
For small-scale instances, the decreasing range between the the optimization proportions in Table VI differ greatly. The
costs of the improved solution presented by SA-TS and the density of customers would increase as enlarging the scale,
costs of the initial solutions by HH is from 57.14% to 13.47%. which is more favorable to the drone delivery.
The cost of initial solution (obtained by HH) for other two As for the computational time of SA-TS, it is controlled
scales of instances are also reduced by 27.35% to 49.33% in 12 s for all small instances, and a little longer for medium
for medium ones and 29.61% to 54.16% for large ones. As instances, which is still controlled in 23 s. Although the
the final cost (obtained by SA-TS) for most instances is over calculating time for the large instances is much longer than
30% lower than the initial cost (obtained by HH), it can be that for the medium instances, it is still acceptable and
proved that the neighborhood operators do be an effective local controlled in 62 s.
optimization method. Finally, the cost and computational time of traditional SA
Besides, it can be observed that the adoption of cooper- and TS are also presented. It can be observed that SA-TS
ated truck and drone for parcel delivery can greatly reduce can provide similar solutions as traditional SA while taking
the delivery costs, compared with the truck-only mode. With less time. Especially for the medium-scale and large-scale
the drone introduced, the final costs obtained by SA-TS for the instances, the calculating time of SA is more than twice as long
truck-drone delivery mode lower the cost ratio from 65.63% to as that of SA-TS. The reason can be inferred that, as the cal-
20.92% comparing to the costs of truck-only mode in small- culation after operations would involve calculating the energy
scale instances. In many instances, the costs of the truck-drone consumption of the whole drone routes, traditional SA will
mode are even less than the half of that with truck only. waste too much time due to repeated and unnecessary cal-
Furthermore, the performance of the whole algorithm culation. Moreover, compared with the traditional TS, SA-TS
(SA-TS) is more stable with larger scale instances. In small- can provide better solution while consuming a little more time.
scale instances, the distribution of customers is relatively In specific, TS can solve most of small-scale instances in less
scattered. At this time, the optimization process is directly than 8 s while SA-TS may need more than 8 even 9 s. This
affected by the layout of the nodes, which is the reason why gap would get widened with the increase of instance size.
Authorized licensed use limited to: Indian Institute Of Technology (Banaras Hindu University) Varanasi. Downloaded on February 06,2023 at 20:29:36 UTC from IEEE Xplore. Restrictions apply.
7462 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 12, DECEMBER 2021
Fig. 11. Layout of road network and the instance for 30 customers in
Changsha.
However, the spent time can lead to better results. Take large-
scale instances as examples, the cost of solutions obtained by
SA-TS is approximately 20% to 45% less than that of solutions
given by TS, while the time difference between two algorithms
is no more than 15 s. It is because that the traditional TS can-
not avoid the solution falling into the local optimal, whereas
our algorithm can overcome this weakness and provide better
solutions in acceptable time.
B. Case Study
1) Case Description: A case is built based on the practical
road network and customer nodes in Changsha, China, which
is used for the sensitivity analysis. As Fig. 11 displays, in
the east urban area of Changsha, 95 crossings of the main
roads are drawn and 30 customer nodes are marked. Through
connecting these crossings, there forms a road network for the
delivery truck.
Specifically, the latitude and longitude coordinates of the
crossings are obtained from the Google map. As for the road
distance of the truck, a tool provided by Google is applied
for the distance calculation of the actual road segments, gen-
erating the initial distance matrix. Then the crossing distance
matrix is filled by the Floyd algorithm to acquire the shortest
distance between any two crossings based on the road network.
The weight for each customer is randomly generated with the Fig. 12. Comparison of solutions for different delivery modes. (a) Truck-
mentioned weight ratio, 90%. only solution. (b) Truck & drone solution with only one parcel in each drone
To display the result visually, the SA algorithm and the flight. (c) Final truck & drone solution with multiple parcels in each drone
flight.
constructed algorithm are both applied to solve the example
instance in Fig.12. Besides, the SA algorithm is also adjusted
to solve the cooperated truck and drone problem with only solution, only ten customers are delivered by the truck, includ-
one parcel in the drone flight. These three results are all illus- ing six customers with heavy parcels. The other 20 customers
trated in Fig. 12. Fig. 12(a) presents the truck-only solution are visited by drone, which vastly saves the delivery cost and
with the cost of 28.63, Fig. 12(b) is the truck & drone solution reduces the delivery time. The computational time is 9.39 s.
with only one parcel in each drone flight and its cost is 23.29, 2) Sensitivity Analysis: The sensitivity tests are conducted
while Fig. 12(c) is the final truck & drone solution with the on the practical case in Fig. 11 with three critical factors,
cost of 16.72. When the drone can only deliver one parcel which are: 1) the ratio of light parcels; 2) drone’s capacity on
in one flight, 12 customers are served by the drone, and there the payload; and 3) the capacity of drone’s battery power. All
exist several cases that the truck stays in the place and waits for the reported results are average values after running algorithms
the drone to reduce the drone’s flight distance and save cost. for ten times.
However, this delivery mode has not took full advantage of a) Impact analysis for the ratio of light parcels: When
drone’s endurance and consumes more time for delivery. These generating the weight of customers’ parcels, a ratio of light
shortages can be made up by our delivery mode. In the final parcels is defined, which means the percentage of light parcels
Authorized licensed use limited to: Indian Institute Of Technology (Banaras Hindu University) Varanasi. Downloaded on February 06,2023 at 20:29:36 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: TWO-ECHELON ROUTING PROBLEM FOR PARCEL DELIVERY BY COOPERATED TRUCK AND DRONE 7463
Fig. 13. Computational results under different ratios of light parcels. Fig. 14. Computational results under different capacities of drone’s payload.
Authorized licensed use limited to: Indian Institute Of Technology (Banaras Hindu University) Varanasi. Downloaded on February 06,2023 at 20:29:36 UTC from IEEE Xplore. Restrictions apply.
7464 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 51, NO. 12, DECEMBER 2021
running out of its power. In this case, relying on only one more environmental than the truck, it is also an important
truck equipped with one drone may not be able to complete extension to solve the pollution routing problem with truck and
the delivery task. Thus, multiple combinations of one truck drone and discuss the contribution of this cooperative delivery
and one drone are taken into account. mode to the environment.
To solve the multiple truck-drone routing problem, the
proposed algorithms are improved. When constructing the R EFERENCES
vehicle route, the constraint of truck endurance is considered. [1] Y.-H. Jia et al., “A dynamic logistic dispatching system with
It is ensured that each truck can return to the depot within its set-based particle swarm optimization,” IEEE Trans. Syst.,
Man, Cybern., Syst., vol. 48, no. 9, pp. 1607–1621, Sep. 2018,
traveling range. Then we apply the cost saving strategy to con- doi: 10.1109/TSMC.2017.2682264.
struct adjoint drone routes for each truck route, which would [2] C. C. Murray and A. G. Chu, “The flying sidekick traveling salesman
generate a feasible solution. Furthermore, SA-TS is used to problem: Optimization of drone-assisted parcel delivery,” Transp. Res.
C, Emerg. Technol., vol. 54, pp. 86–109, May 2015.
optimize the initial solution. Not only can routes within a sin- [3] Q. M. Ha, Y. Deville, Q. D. Pham, and M. H. Hà, “On the min-cost trav-
gle combination be optimized but the operations can also work eling salesman problem with drone,” Transp. Res. C, Emerg. Technol.,
on different nodes from different combinations. vol. 86, pp. 597–621, Jan. 2018.
[4] The Vision Van—The Design. Accessed: Nov. 2016. [Online]. Available:
In order to verify the effectiveness of the modified algo- https://www.mercedes-benz.com/en/design/vehicles/the-vision-van-the-
rithm, 25 customers close to depot are selected from the case design/
in Fig. 11. The result is illustrated in Fig. 16. Two trucks [5] UPS Tests Package Delivery Drone in Residential
Area in Florida. Accessed: Feb. 2017. [Online]. Available:
are employed and 12 customers are served by drones, which https://www.trucks.com/2017/02/21/ups-drone-delivery-florida-test/
totally requires the cost of 14.23. [6] N. Agatz, P. Bouman, and M. Schmidt, “Optimization approaches for
the traveling salesman problem with drone,” Transp. Sci., vol. 54, no. 4,
pp. 739–1034, Jul./Aug. 2018.
[7] E. E. Yurek and H. C. Ozmutlu, “A decomposition-based iterative
VI. C ONCLUSION optimization algorithm for traveling salesman problem with drone,”
Transp. Res. C, Emerg. Technol., vol. 91, pp. 249–262, Jun. 2018.
In this article, recognizing the importance and necessity of [8] K. Dorling, J. Heinrichs, G. G. Messier, and S. Magierowski,
the coordination of the truck and the drone in parcel delivery, “Vehicle routing problems for drone delivery,” IEEE Trans. Syst.,
the drone-assisted parcel delivery problem is studied. Since Man, Cybern., Syst., vol. 47, no. 1, pp. 70–85, Jan. 2017,
doi: 10.1109/TSMC.2016.2582745.
multiple parcels are allowed to be delivered by the drone in [9] K. Sundar and S. Rathinam, “Algorithms for routing an unmanned
one flight, the investigated problem can be viewed as a new aerial vehicle in the presence of refueling depots,” IEEE Trans.
variant of two-echelon routing problem. Besides, the effect of Autom. Sci. Eng., vol. 11, no. 1, pp. 287–294, Jan. 2014,
doi: 10.1109/TASE.2013.2279544.
payload on the energy consumption is estimated by a proposed [10] K. T. San, E. Y. Lee, and Y. S. Chang, “The delivery assignment solution
energy consumption model. Then an algorithm based on SA is for swarms of UAVs dealing with multi-dimensional chromosome rep-
presented. A heuristic combining nearest neighbor algorithm resentation of genetic algorithm,” in Proc. IEEE 7th Annu. Ubiquitous
Comput. Electron. Mobile Commun. Conf. (UEMCON), New York, NY,
and saving algorithm is designed to generate the initial solu- USA, 2016, pp. 1–7, doi: 10.1109/UEMCON.2016.7777839.
tion. And the variable local search is conducted with removal [11] A. M. Ham, “Integrated scheduling of m-truck, m-drone, and m-depot
constrained by time-window, drop-pickup, and m-visit using constraint
operators and insertion operator in every iteration to find a programming,” Transp. Res. C, Emerg. Technol., vol. 91, pp. 1–14,
better solution. Jun. 2018.
Radom test experiments with different scales are con- [12] V. K. Shetty, M. Sudit, and R. Nagi, “Priority-based assignment and
routing of a fleet of unmanned combat aerial vehicles,” Comput. Oper.
ducted and indicated the efficiency of the proposed algorithm. Res., vol. 35, no. 6, pp. 1813–1828, Jun. 2008.
Furthermore, three factors are considered in sensitivity tests [13] G. S. C. Avellar, G. A. S. Pereira, L. C. A. Pimenta, and P.Iscold,
based on a practical case, which are the ratio of light parcels, “Multi-UAV routing for area coverage and remote sensing with min-
imum time,” Sensors, vol. 15, no. 11, pp. 27783–27803, Nov. 2015,
maximum payload of drone, and drone’s battery power. The doi: 10.3390/s151127783.
experimental results show that the employment of the drone [14] S. G. Manyam, S. Rasmussen, D. W. Casbeer, K. Kalyanam, and
can save more costs when there are more light parcels for S. Manickam, “Multi-UAV routing for persistent intelligence surveil-
lance & reconnaissance missions,” in Proc. Int. Conf. Unmanned
delivery. Also, proper improvement on technologies for enlarg- Aircraft Syst. (ICUAS), Miami, FL, USA, 2017, pp. 573–580,
ing the capacities of the drone’s payload and battery power doi: 10.1109/ICUAS.2017.7991314.
would help save more delivery cost. [15] F. Mufalli, R. Batta, and R. Nagi, “Simultaneous sensor selection
and routing of unmanned aerial vehicles for complex mission plans,”
It is an emerging phenomenon in recent years that the Comput. Oper. Res., vol. 39, no. 11, pp. 2787–2799, Nov. 2012.
drone is employed in parcel delivery to cooperate with the [16] B. D. Song, K. Park, and J. Kim, “Persistent UAV delivery logistics:
truck. There are many research topics that are not studied MILP formulation and efficient heuristic,” Comput. Ind. Eng., vol. 120,
pp. 418–428, Jun. 2018.
in the management and operation research area, such as the [17] S. Kim and I. Moon, “Traveling salesman problem with a drone sta-
two-echelon truck & drone routing problems with time win- tion,” IEEE Trans. Syst., Man, Cybern., Syst., vol. 49, no. 1, pp. 42–52,
Jan. 2019, doi: 10.1109/TSMC.2018.2867496.
dow, with multiple vehicles, and with multiple depots. More [18] J. C. de Freitas and P. H. V. Penna, “A randomized variable neighbor-
interesting and valuable variants of this problem are waiting hood descent heuristic to solve the flying sidekick traveling salesman
to be explored. Due to the complexity for the optimization of problem,” Electron. Notes Discr. Math., vol. 66, pp. 95–102, Apr. 2018.
[19] B. Grocholsky, J. Keller, V. Kumar, and G. Pappas, “Cooperative air
the two-echelon routes, more efficient algorithms also have to and ground surveillance,” IEEE Robot. Autom. Mag., vol. 13, no. 3,
be developed in future work. pp. 16–25, Sep. 2006, doi: 10.1109/MRA.2006.1678135.
Furthermore, with increasing awareness of environmental [20] P. Maini and P. B. Sujit, “On cooperation between a fuel constrained
UAV and a refueling UGV for large scale mapping applications,” in
protection, there have been many researches considering the Proc. Int. Conf. Unmanned Aircraft Syst. (ICUAS), Denver, CO, USA,
pollution routing problem [36]–[38]. Since the drone is much 2015, pp. 1370–1377, doi: 10.1109/ICUAS.2015.7152432.
Authorized licensed use limited to: Indian Institute Of Technology (Banaras Hindu University) Varanasi. Downloaded on February 06,2023 at 20:29:36 UTC from IEEE Xplore. Restrictions apply.
LIU et al.: TWO-ECHELON ROUTING PROBLEM FOR PARCEL DELIVERY BY COOPERATED TRUCK AND DRONE 7465
[21] H. Savuran and M. Karakaya, “Route optimization method for unmanned Jianmai Shi received the Ph.D. degree in man-
air vehicle launched from a carrier,” Lecture Notes Softw. Eng., vol. 3, agement science and engineering from the National
no. 4, pp. 279–284, Nov. 2015. University of Defense Technology (NUDT),
[22] Z. Luo, Z. Liu, and J. Shi, “A two-echelon cooperated routing problem Changsha, China, in 2011.
for a ground vehicle and its carried unmanned aerial vehicle,” Sensors, From 2008 to 2009, he was a visiting Ph.D. stu-
vol. 17, no. 5, p. 1144, May 2017, doi: 10.3390/s17051144. dent with the University of Windsor, Windsor, ON,
[23] R. Cuda, G. Guastaroba, and M. G. Speranza, “A survey on two-echelon Canada. From 2014 to 2015, he was a Visiting
routing problems,” Comput. Oper. Res., vol. 55, pp. 185–199, Mar. 2015. Scholar with McGill University, Montreal, QC,
[24] R. D’Andrea, “Guest editorial can drones deliver?” IEEE Trans. Canada. He is currently an Associate Professor
Autom. Sci. Eng., vol. 11, no. 3, pp. 647–648, Jul. 2014, with the School of System Engineering, NUDT.
doi: 10.1109/TASE.2014.2326952. He has authored more than 30 refereed papers,
[25] T. Cover and P. Hart, “Nearest neighbor pattern classification,” including those published in the IEEE T RANSACTIONS ON I NTELLIGENT
IEEE Trans. Inf. Theory, vol. 13, no. 1, pp. 21–27, Jan. 1967, T RANSPORTATION S YSTEMS, Computers & Operations Research, and
doi: 10.1109/TIT.1967.1053964. Applied Energy. His current research interests include vehicle routing, supply
[26] G. Clarke and J. W. Wright, “Scheduling of vehicles from a central depot chain management, and heuristic design.
to a number of delivery points,” Oper. Res., vol. 12, no. 4, pp. 568–581,
1964.
[27] S. Kirkpatrick, C. D. Gelatt, Jr., and M. P. Vecchi, “Optimization by sim-
ulated annealing,” Science, vol. 220, no. 4598, pp. 671–680, May 1983,
doi: 10.1126/science.220.4598.671.
[28] G. Reinelt, The Traveling Salesman: Computational Solutions for TSP
Applications, Berlin, Germany: Springer-Verlag, 1994.
[29] A. Ponza, Optimization of Drone Assisted-Parcel Delivery, Univ. Padova,
Padua, Italy, 2016.
[30] A. El-Bouri, N. Azizi, and S. Zolfaghari, “A comparative study of a
new heuristic based on adaptive memory programming and simulated Guohua Wu received the B.S. degree in information
annealing: The case of job shop scheduling,” Eur. J. Oper. Res., vol. 177, systems and the Ph.D. degree in operations research
no. 3, pp. 1894–1910, Mar. 2007.
from the National University of Defense Technology,
[31] G. Wu et al., “An adaptive simulated annealing-based satellite obser-
Changsha, China, in 2008 and 2014, respectively.
vation scheduling method combined with a dynamic task clustering
From 2012 to 2014, he was a visiting Ph.D. stu-
strategy,” Comput. Ind. Eng., vol. 113, pp. 576–588, Jan. 2014.
dent with the University of Alberta, Edmonton,
[32] R. Allain. Physics of the Amazon Octocopter Drone. Accessed: Dec.
AB, Canada. From 2014 to 2017, he was
2013. [Online]. Available: https://www.wired.com/2013/12/physics-of-
a Lecturer with the College of Information Systems
the-amazon-prime-air-drone/
[33] G. P. Cachon, “Retail store density and the cost of greenhouse gas and Management, National University of Defense
emissions,” Manag. Sci., vol. 60, no. 8, pp. 1907–1925, 2014. Technology. He is currently a Professor with the
[34] C. Guglielmo. Turns Out Amazon, Touting Drone Delivery, School of Traffic and Transportation Engineering,
Does Sell Lots of Products That Weigh Less Than Central South University, Changsha. His current research interests include
5 Pounds. Accessed: Dec. 2013. [Online]. Available: planning and scheduling, evolutionary computation, and machine learning. He
https://www.forbes.com/sites/connieguglielmo/2013/12/02/turns-out- has authored more than 50 refereed papers, including those published in the
amazon-touting-drone-delivery-does-sell-lots-of-products-that-weigh- IEEE T RANSACTIONS ON S YSTEMS , M AN , AND C YBERNETICS : S YSTEMS,
less-than-5-pounds/#6823b14d455e IEEE T RANSACTIONS ON C YBERNETICS, and Information Sciences.
[35] J. J. Ikoba. JD.com CEO Sets February 2018 Target to Change Its Fleet Prof. Wu served as an Associate Editor of Swarm and Evolutionary
of Delivery Vans in Beijing to Electric. Accessed: Dec. 2017. [Online]. Computation Journal and an Editorial Board Member of the International
Available: https://www.gizmochina.com/2017/12/13/jd-com-ceo-sets- Journal of Bio-Inspired Computation. He is a Regular Reviewer of more
february-2018-target-change-fleet-delivery-vans-beijing-electric/ than 20 journals, including the IEEE T RANSACTIONS ON E VOLUTIONARY
[36] T. Bektaş and G. Laporte, “The pollution-routing problem,” Transp. Res. C OMPUTATION, IEEE T RANSACTIONS ON C YBERNETICS, Information
B, Methodol., vol. 45, no. 8, pp. 1232–1250, 2011. Sciences, and Applied Soft Computing.
[37] E. Demir, T. Bektaş, and G. Laporte, “The bi-objective pollution-routing
problem,” Eur. J. Oper. Res., vol. 232, no. 3, pp. 464–478, 2014.
[38] O. Dukkanci, B. Y. Kara, and T. Bektaş, “The green location-routing
problem,” Comput. Oper. Res., vol. 105, pp. 187–202, May 2019.
Authorized licensed use limited to: Indian Institute Of Technology (Banaras Hindu University) Varanasi. Downloaded on February 06,2023 at 20:29:36 UTC from IEEE Xplore. Restrictions apply.