Hybrid Truck-Drone Delivery Systems: A Systematic Literature Review

Owing to the continued development of e-commerce, logisticians now have an outstanding obligation to tackle last-mile delivery challenges. A number of logistics providers have suggested the incorporation of drones with trucks to provide a more flexible delivery system. This paper analyzes the content of 95 publications related to hybrid truck-drone delivery systems (HTDDS) in the context of last-mile delivery. First, a brief overview of the potential implementation of drone delivery systems is presented, including their integration with other vehicles. The overview aims to demonstrate the operational characteristics of such systems and their implications. Then, the surveyed literature is classified based on vehicles roles, system configuration, problem formulation, and solution methods. In relation to this research, several key findings and potential research directions are discussed. Despite the high level of interest in HTDDS research, it is still in its early phases and requires improvements in various areas. The payload capacity, speed, range, and energy consumption are all factors that must be considered in the modeling of drone characteristics. Almost all studies identify customer requests before the delivery operation begins. However, customer demands for immediate delivery present an opportunity for real-time optimization to provide solutions for e-commerce activities. Environmental issues are developing, as the last-mile delivery problem is regarded as the most polluting portion of the supply chain. Thus, more consideration should be given to the environmental impact of HTDDS. Finally, research on drone routing related to air traffic management has received relatively little attention.

The associate editor coordinating the review of this manuscript and approving it for publication was Frederico Guimarães .
problem, referred to as delivering goods from e-retailers hubs 32 to their final destinations [3], is one of the main issues that 33 logistics providers need to tackle. It is the most expensive, 34 most polluting, and least efficient part of the e-commerce 35 supply chain, accounting for 13%-75% of the total supply 36 chain cost [4], [5]. In a 2020 survey on identifying the most 37 significant challenges to LMD, 35% of retailers and man-38 ufacturers stated that reducing logistics costs is the main 39 obstacle in providing efficient LMD services [6]. Addition-40 ally, environmental reports urge logistics providers to shift 41 to eco-friendly solutions, where carbon dioxide emissions 42 from freight transportation account for 30% of transportation-43 related carbon emissions [7]. The transportation industry 44 must look for alternatives to tackle LMD hurdles such as high 45 cost, ecological impact, and the complexity of supply chain 46 performance. Businesses have begun competing to develop 47 medicine [25]. Zipline is now extending its drone technology 104 to Ghana in response to the current Covid-19 pandemic. 105 It allows contactless drone delivery to transport Covid-19 106 test samples and therefore help healthcare authorities to react 107 quickly to the pandemic and save lives [26]. Flirtey has 108 also provided aid kits and emergency medication delivery 109 in Australia and New Zealand [16]. Drone delivery has the 110 potential to decrease delivery costs, avoid the congestion 111 of traditional road networks, reduce carbon emissions, and 112 increase customer satisfaction by reducing the number of 113 missed deliveries caused by delivery delays [27]. 114

115
This section provides a brief review of potential drone 116 delivery systems in last-mile operations and discusses the 117 significance and the different challenges entailed by such sys-118 tems. We present the potential implementation of drone deliv-119 ery systems adopted by companies such as Amazon, UPS, 120 and others. These implementations differ in terms of their 121 operational characteristics. Table 1 outlines several potential 122 applications of drones in the LMD based on whether they 123 operate independently or in conjunction with other vehicles 124 such as trucks, trains, etc. Drones are launched from a fixed 125 location like a fulfillment center in the first scenario, while 126 they are released from a mobile vehicle, demonstrating a 127 launch from a moving location in the second case. In addition, 128 the operational characteristics of the drones along with the 129 implementation of these delivery systems are presented. 130 Despite the advantages of drones, three major restrictions 131 limit the performance of a drone-based delivery system. First, 132 drone payloads are limited in terms of permitted size and 133 overall weight of shipment [28]. Second, due to the restricted 134 capacity of existing battery technology, drone operations are 135 often hindered by short flight duration constraints [11]. This 136 means that batteries must be recharged or replaced after each 137 drone route. Third, the existing technology and regulations 138 restrict the complete dynamic and coordinated control of 139 multiple drone-based delivery models [29]. Furthermore, the 140 FAA has established restrictions that limit the use of drones 141 for commercial purposes when flying in airspace such as 142 keeping the drone within the visual line of sight [30]. Thus, 143 the use of drones for commercial purposes will tremendously 144 benefit businesses but will still enact significant limits. 145 A recent research avenue is to integrate drones with tra- 146 ditional delivery methods such as trucks to form hybrid 147 truck-drone delivery systems (HTDDS). With the combina-148 tion of the two delivery methods, the advantages of the truck, 149 such as long-range travel capability and high load capacity, 150 can offset the disadvantages of drones and vice versa [31]. 151 Figure 1 depicts an example of the HTDDS configuration. 152 In general, the truck departs from the depot to launch the 153 drone at a specific node, which could be a customer node 154 or any non-customer node (drone station, a point along the 155 truck's path, etc.). Afterward. the drone delivers to customers 156 and finally returns to the truck for recharging, maintenance, 157 approaches, vehicle synchronization, and challenges to DO 197 and DTCO implementation. The authors surveyed 68 papers 198 on DTCO in a variety of fields, including agriculture, secu-199 rity, disaster management, entertainment and media, trans-200 portation and logistics, and other areas. A total of 43 articles 201 in the reviewed literature are concerned with the application 202 of transportation and logistics. Poikonen and Campbell [36] 203 identified future directions in the research of drone opti-204 mization. They also provided ways to improve modeling in 205 the context of drone capabilities and suggested alternative 206 delivery modes. Rojas Viloria et al. [38] presented a liter-207 ature review of 79 publications on the characterization of 208 routing problems using drones in applications such as parcel 209 delivery, surveillance, entertainment, military, and internal 210 logistics. The surveyed literature is divided into categories 211 based on the goal function, solution strategy, and constraints. 212 Among the 79 articles, there are 25 related to package deliv-213 ery using drones in combination with vehicles such as trucks, 214 unmanned ground vehicles, motorcycles, and others. Thus, 215 there is no extensive discussion in the field of HTDDS. Simi-216 larly, Cheikhrouhou and Khoufi [39] reviewed the recent con-217 tributions regarding ground vehicles and UAVs, with a focus 218 on the application fields such as transportation and delivery, 219 data collection, search and rescue, multi-robot task alloca-220 tion, and scheduling. For each contribution, they discussed 221 optimization approaches like exact methods and metaheuris-222 tics to solve routing problems such as the multiple traveling 223 salesman problem (mTSP). However, there is no detailed 224 review of HTDDS as the paper presents a brief review of 225 them. Boysen et al. [11] recorded the several delivery con-226 cepts methodically in a concise notation scheme, analyzed the 227 relevant decision problems, and examined previous research 228 on operations research approaches for tackling these chal-229 lenges. Hence, their focus is not on drone operations solely. 230 Add to that, the literature on drone delivery is classified based 231 on whether it is merely on drone operations or in conjunction 232 with other vehicles, as well as solution methods. Li et al. [42] 233 reviewed the integration of ground vehicles and UAV forming 234 a two-echelon network. The authors classified two-echelon 235 network routing challenges depending on the characteristics 236 of the routing problems. The categorization supplied is lim-237 ited and considers features such as number of drones, payload 238 capacity of drones, time window, objective function and solu-239 tion methodologies. Finally, Moshref-Javadi and Winken-240 bach [44] provided a comprehensive review of the real-world 241 applications of drones in the logistics industry. They also 242 reviewed the relevant drone-based logistics systems and their 243 associated operational planning problems. In particular, the 244 systematic review covers a variety of application areas such 245 as emergency services, healthcare services, e-commerce, dis-246 aster relief, food distribution, and others.

247
Although a great number of survey papers have been con-248 ducted, none of the available contributions provide a sys-249 tematic classification approach for distinguishing HTDDS 250 models based on their detailed system configurations (num-251 ber of vehicles, role of vehicle, and model description), 252 els' application in real life is discussed by highlighting the 289 roadblocks to employing existing HTDDS models in real 290 life.

291
The remainder of the paper is structured as follows: 292 In section 2, the methodology of the systematic literature 293 review is presented. Section 3 provides the state of the art 294 of HTDDS research and summarizes papers based on the 295 roles of vehicles and system configuration. Section 4 clas-296 sifies the evaluated literature based on objective function 297 and model constraints and parameters, while section 5 298 classifies it based on solution methods. Section 6 states 299 research gaps and recommendations as well as barriers to 300 the real-world application of HTDDS. Finally, the review is 301 concluded.

303
In this paper, we follow the methodology suggested by 304 Tranfield et al. [52] for conducting a systematic review. 305 Figure 2 shows the main stages of the systematic review 306 methodology. Each of the three stages is provided in detail 307 below:  and an optimal solution is given. In some cases, the drone is Many studies have been conducted on the TSP-D problem.

522
Among them, Agatz et al. [12] were the first to propose  Ha et al.
[31] build on the work of Murray and Chu [53] and 533 Agatz et al. [12] by focusing on the cost aspect of the objec-534 tive function rather than the time aspect. To present the prob-535 lem known as min-cost TSP-D, two heuristics are proposed: 536 Greedy Randomized Adaptive Search Procedure (GRASP), 537 and TSP with Local Search (TSP-LS) which is adapted from 538 the heuristic proposed by Murray and Chu [53]. Tu et al.
[71] 539 propose the TSP with multiple drones (TSP-mD) problem 540 which extends the TSP-D by taking into account multiple 541 drones. A certain number of drones are permitted to fly 542 at the same time and be retrieved by truck at the same or 543 different location. Both vehicles cannot wait for each other 544 for more than a specific period of time. GRASP, which was 545 initially developed by Ha et al. [31], was used to solve the 546 problem. Kitjacharoenchai et al. [72] enable several trucks 547 and drones to serve customers as salesmen. The aim of the 548 multiple traveling salesmen problem with drones (mTSPD), 549 which uses a version of mTSP called min-max TSP, is to 550 reduce each salesman's overall tour length (in terms of time). 551 In this routing problem, the following solutions are permitted: 552 drones can depart and return to the depot, drones can depart 553 from the depot and fly back to the truck; or drones can be 554 launched from the truck and retrieved by either the same 555 truck; or a different truck from where they were launched.   The traveling salesman problem with a moving depot 642 (TSP-MD) presented in Madani and Ndiaye [84] considers 643 the truck acting as a moving depot for launching and col-644 lecting the drone. The drone can deliver to multiple cus-645 tomers before returning to the truck, and the goal is to 646 find the optimal locations for launching and retrieving the 647 drone while minimizing the traveling costs of both vehicles. 648 Moshref-Javadi et al. [85] introduce the STRPD problem, 649 in which a truck acting as a moving depot and a fleet of 650 drones are synchronized to serve customers. The truck can 651 only launch the drone at a customer location and retrieve 652 it back at one of the stops along its route. The truck and 653 drones move in tandem, and multiple drones can be launched 654 simultaneously. Liu et al. [86] extend the traditional two-655 echelon routing problem by incorporating a single truck and 656 a single drone into the delivery system. In the two-echelon 657 routing problem for parcel delivery by cooperated truck and 658 drone (2E-RP-T&D), the truck serves customers and acts as 659 a moving depot for the drone, which can deliver to multiple 660 customers before returning to the truck. An energy consump-661 tion model has been proposed to investigate the impact of the 662 payload on the drone's energy consumption and estimate the 663 cost of the drone's sub-routes. The objective is to optimize 664 the truck's main route and the drone's sub-routes while sat-665 isfying the drone's capacity constraints on battery and pay-666 load. The k-multi-visit drone routing problem (k-MVDRP) 667 introduced by Poikonen and Golden [87], considers a truck 668 and k drones to serve a set of customers. The truck only 669 serves as a moving depot and recharging platform for the 670 drones; it does not provide delivery services. The drones 671 can carry multiple packages of varying weights and make 672 multiple visits before returning to the truck to be recharged or 673 pick up new packages. Karak and Abdelghany [88] introduce 674 the hybrid vehicle-drone routing problem (HVDRP), which 675 consists of a single vehicle acting as a moving depot, and 676 multiple drones to pick up and/or deliver packages of different 677 weights. In this model, drone stations are used, where both 678 vehicles can wait for each other. This setting allows for 679 multiple drone dispatches from different stations, resulting 680 in a larger customer population covered. Each drone can 681 return to any of the stations, which could be similar to the 682 same or different from the dispatch station. Boysen et al. [89] 683 consider the drone scheduling problem (DSP) for a single 684 truck-multiple drone delivery system. The truck follows a 685 fixed route and can dispatch and retrieve the drones at any 686 of the predetermined stops along the route. A single drone 687 is launched from the truck at a time, and after delivering, 688 it can return to the same or a different stop. The objective is 689 to schedule the launch of the drones so that the total delivery 690 time is minimized. 691 Bai et al. [90] propose a single truck-single drone delivery 692 system configuration. The truck is restricted to following a set 693 of street-vertices while the drone serves customers. A prece-694 dence constraint is ensured to indicate which customer is 695 to be visited before the other customer. The objective is to 696 minimize the time when the last customer is served to increase 697 ring-expressway to act as a platform for UAV service, and 716 the objective is to reduce delivery time. Mathew et al. [93] 717 present the heterogeneous delivery problem (HDP) using a 718 truck and a drone. The truck does not deliver, but rather 719 transports the drone to a neighboring node within the drone's 720 flying range. The goal is to find the best truck and drone 721 routes that decrease overall delivery time. Jeong and Lee [94] 722 introduce the drone routing problem with a truck (DRP-T), 723 where the truck does not service customers but merely trans-724 ports drones to a parking location. Following that, the drones 725 will take-off to serve a single customer before returning to 726 another parking place to be collected by the truck. The goal 727 is to have the truck arrive at the depot as soon as possible so 728 that all deliveries may be completed. to the depot independently of the drone. Figure 6 illustrates 737 the parallel delivery operations of the truck and the drone. 738 In the evaluated literature, a few problems with this operation 739 are considered, and they are given below. The parallel drone scheduling traveling salesman prob-742 lem (PDSTSP) was first introduced by Murray and Chu [53], 743 and the problem considers a single truck and a fleet of iden-744 tical drones serving customers in parallel, implying that the 745 two types of vehicles are not synchronized. The truck follows 746 a TSP tour, while drones fulfill customer deliveries only 747 within the distribution center's flight range. As in FSTSP, 748 the objective of PDSTSP is to minimize the latest time to 749 return to the depot for both the truck and drones. Different 750 solution methods were used to solve the same problem in 751 [95], [96], and [97]. Kim and Moon [98] extend the PDSTSP 752 to overcome the large distance between the distribution center 753 and the customer locations, allowing the drones to cover a 754 larger number of customers. The proposed traveling salesman 755 problem with drone station (TSP-DS) considers a drone sta-756 tion located at a distance greater than the drone's maximum 757 flight distance, which was used for charging and refilling the 758 drones with parcels. Schermer et al. [99] integrate the routing 759 of the truck, location of the drone stations, and scheduling of 760 the drones to provide a general case of the PDSTSP and TSP-761 DS. The integrated traveling salesman drone station location 762 problem (TSDSLP) model incorporates drone deliveries into 763 TSP tours as well as the use of multiple drone delivery 764 stations. Each drone station can only handle a certain number 765 of drones, and each drone must return to the same station from 766 which it was launched. In some circumstances, the drone's conventional duty of serv-769 ing customers is supplanted by a different role. The drone can 770 serve as a replenishment source for the truck, which is the 771 only mode of delivery. Figure 7 provides an example of the 772 resupply operations. The drone resupplies the truck at certain 773 customer nodes and returns to the depot independently. The 774 literature that considers the resupply operation is given below. 775 Pina-Pardo et al.
[100] propose the traveling salesman 776 problem with release dates and drone resupply (TSPRD-DR). 777 The objective is to find the shortest possible time route for 778 a single truck that is resupplied by a drone while enroute 779 to fulfill customer orders. The orders and their associated 780 information are known at the beginning of the day, but they 781 may not be ready to ship. While the truck is making other 782 deliveries, the drone is loaded at the depot with ready-to-ship 783 orders and only resupplies the truck at customers' locations. 784 Dayarian et al. [101] present the vehicle routing problem 785 with drone resupply (VRPDR), which considers a fleet of 786 vehicles and a fleet of drones assisting each other to perform 787 home deliveries. Similar to Pina-Pardo et al.

820
Besides the FSTSP problem that requires synchronization, 821 Murray and Chu [53] deal with another problem. The latter 822 problem is known as the PDSTSP which considers the inde-823 pendent operation of the two vehicles.

824
The TSP-D considered in [12], [31], [102], [103], [ is illustrated in STRPD [85].    Li et al. [129] propose 923 a delivery system using multiple trucks and drones, where 924 no synchronization is required, while Pugliese et al. [130] 925 allow for the drone to wait for a specific time for the truck. 926 Wang et al. [131] consider a HTDDS consisting of multi-927 ple trucks, multiple truck-carried drones, and independent 928 drones. Each truck carries one drone, in which synchroniza-929 tion is needed between them. On the other hand, indepen-930 dent drones do not require synchronization with the trucks. 931 They depart from the depot, serve customers according to 932 their payload capacity and return to the depot. Ulmer and 933 Thomas [132] propose a same-day delivery system, where 934 a fleet of trucks and drones is used to fulfill the cus-935 tomers' changing demands. Like PDSTSP, no synchroniza-936 tion is needed between the vehicles. Table 3

944
In this section, the reviewed literature is used to categorize 945 the objective functions, followed by discussing the used con-946 straints and parameters.

948
Truck-drone delivery systems are introduced to address the 949 issues of LMD, which is the most costly, polluting, and inef-950 ficient element of the supply chain [5]. As a result, multiple 951 objective functions addressing various concerns in the LMD, 952 such as traveling costs, carbon emissions, delays, missing 953 deliveries, and others, have been identified. In this subsec-954 tion, we summarize these objective functions based on four 955 elements: total traveling time, operational and transportation 956 costs, environment, and service. can provide new opportunities to improve home delivery 965 processes. They operate without a human pilot, avoid the 966 congestion of traditional road networks by flying over them, 967 and are faster than trucks. Add to that, Amazon emphasized 968 fast delivery by deploying drones to provide deliveries within 969 30 minutes [24]. Hence, minimizing the total time would be 970 a rational objective function. Researchers used various types of parameters and constraints 1029 to define the HTDDS problems on hand. Capturing the real 1030 structure of HTDDS is reflected in the nature of the con-1031 straints and parameters considered. A detailed description of 1032 the constraints and parameters is presented below. 1033

1034
The main assumptions and parameters considered in the 1035 reviewed literature deal with operational characteristics such 1036 as service time or drone technical characteristics such as 1037 speed, weight, and energy consumption, as well as other con-1038 siderations such as customer demands, the fuel consumption 1039 of trucks, and distance metrics, as provided below.

1040
• During delivery, the drones and trucks need some time to 1041 serve the customer. This is associated with the delivery 1042 mechanisms of each drone and truck.    For instance, the customer might specify the delivery 1229 time of two orders within the same day but at different 1230 times. Thus, the same customer has two time-windows 1231 on the same day. Table 4 summarizes the most common 1232 constraints in the evaluated literature.  For the least common techniques, a brief description of 1252 each is given to understand their implementations in the con-1253 text of the HTDDS. In terms of DP, Bouman et al. [103] used 1254 a three-pass approach based on Bellman-Held-Karp DP to 1255 solve TSP-D. The purpose of the use of DP is to overcome the 1256 limitations of the exact approaches used by Agatz et al. [12], 1257 which are only capable of solving small instances (up to 1258 10 nodes). DP is known for its capability of solving larger 1259 instances. Due to the exponential number of variables and 1260 constraints of the TSP-D, DP continuously outperforms the of integrated trucks and drones with single and multiple 1264 depots. Also, CP is considered by Bai et al. [90] to model a [78], [139], the branch-and-cut algorithm [59], [60], [141], 1306 and the branch-and-price algorithm [108]. 1307 Cokyasar et al. [144] proposed the notion of locating auto- accounting for any congestion at battery swapping operations 1316 using queueing theory.

1317
As for the exact solvers, the most commonly used solvers 1318 are CPLEX, GAMS with CPLEX, and Gurobi. As shown 1319 in Figure 9, the majority of the reviewed literature (61.8%) 1320 considered exact solutions, with 27% of the articles using 1321 only exact methods. The remainder (34.8%) accounts for 1322 heuristics and/or metaheuristics along with exact methods. 1323 HTDDS routing problems are extended versions of classical 1324 VRPs with various sources of modeling challenges. This may 1325 include the number of trucks and drones, the number of 1326 depots, and drone and truck operational limitations. When 1327 the HTDDS considers a single truck stop, a single drone, 1328 and a single delivery route, the problem is reduced to a TSP. 1329 TSPs are known to be NP-hard, and consequently, HTDDS 1330 problems are NP-hard, which makes them computationally 1331 difficult to reach optimality. Due to this NP-hard nature, 1332 developing efficient heuristics and metaheuristics is required 1333 for large-size problems [152].

1335
A variety of heuristics and metaheuristics are implemented 1336 to solve HTDDS problems. For solving the FSTSP, Murray 1337 and Chu [53] used a route and re-assign heuristic, which 1338 is based on savings, nearest neighbor, and sweep, to solve 1339 problems with up to 20 customer nodes. The solution meth-1340 ods proposed by Murray and Chu [53] were able to solve 1341 small-size instances. Other studies were able to solve larger 1342 instances by improving the performance of solution methods. 1343 For example, in Kundu and Matis [58], the effect of wind 1344 and battery-power consumption are considered. As a result, 1345 the authors have modified the same heuristic to solve up to 1346 100 customer locations. To overcome the traditional heuristic 1347 methods proposed by Murray and Chu [53], de Freitas and 1348 Penna [54] used the Randomized Variable Neighborhood 1349 Descent (RVND) as a local search method. The RVND was 1350 able to solve instances with up to 100 nodes. de Freitas 1351 and Penna [56] have further improved the solution method 1352 for solving the FSTSP to handle instances with 200 nodes 1353 using a hybrid method. The work uses an exact model to 1354  and developed a three-phased iterative heuristic. The first 1366 phase considers initial customer assignments, while the sec-1367 ond creates the drones' paths. Finally, the starting times for 1368 the truck and drone operations, and the queuing of launching 1369 and collecting, are determined. A variant of the FSTSP is 1370 the mFSTSP-ECNZ, where a two-phase constructive and 1371 search heuristic is used to solve it [63]. Another variant 1372 is mFSTSP-VDS where a three-phased iterative heuristic is 1373 proposed. The heuristic includes partitioning customers and 1374 creating TSP tours; creating the drone's path; and scheduling 1375 the operations and timing [64].

1376
The TSP-D was solved using different methods, such as the 1377 route first-cluster second heuristic approach based on local 1378 search and dynamic programming [12], GRASP [31], [110]. 1379 These two methods were capable of solving instances with up 1380 to 100 nodes. Other methods used included a hybrid Genetic 1381 Algorithm (GA) [104] and CP-based heuristic [106]. The 1382 hybrid GA is a mix of a GA and 16 local search operators. 1383 It also involves a population management, diversity control, 1384 and penalization mechanism that balances the search between 1385 feasible and infeasible search areas. It is proven to provide 1386 better solution quality for instances of size 100 compared to 1387 the GRASP method

1446
In the context of HTDDS problems, scheduling problems 1447 are more concerned with the scheduling of drone operations 1448 that are in simultaneous operation with the truck [53], [96], 1449 [97], [98], [99]. These scheduling problems are analogous 1450 to the parallel machine scheduling problem, in which each 1451 customer is assigned to a drone based on the flight time 1452 required to complete the operation. For problem formula-1453 tion, the MILP approach is employed as a modeling tech-1454 nique. Ham [95], on the other hand, addressed the parallel 1455 scheduling of multiple vehicles, drones, and depots. The 1456 PDSTSP problem here is distinguished as an unrelated paral-1457 lel machine scheduling problem with a sequence-dependent 1458 structure (traveling distances), a precedence relationship (par-1459 cel delivery and pickup), and reentrant behavior (multiple 1460 visits and time window). This clearly demonstrates the prob-1461 lem's difficulty, validating the adoption of CP as a modeling 1462 technique.

1463
Task assignment problems are very much less commonly 1464 considered in the HTDDS literature. They emerge with mul-1465 tiple vehicle operations that need coordinated planning. For 1466 instance, the PCHDP in Bai et al. [90] is presented as a task 1467 assignment with precedence constraints. These constraints 1468 specify which customers should be served before others. As a 1469 result, the PCHDP is formulated as a constrained minimiza-1470 tion problem.

1471
The vast majority of the modeling techniques are of the 1472 same type. MILP models are extensively employed due 1473 to the structure of the problems introduced, which takes 1474 into account fixed drone characteristics. On the other hand, 1475 heuristic and metaheuristic solutions are more popular when 1476 the drone's characteristics such as speed and energy con-1477 sumption are changing. When considering the drone's speed 1478 as a variable, Raj [131] evaluated the drone's energy consump-1483 tion as a function of characteristics such as drone weight, 1484 package weights, speeds, and others. The authors used an 1485 enhanced artificial bee colony method, a greedy heuristic, and 1486 a hybrid heuristic, respectively. Exact solutions are imple-1487 mented in only a few circumstances. Pugliese et al. [130], 1488 for example, took the drone's power consumption to be pro-1489 portionate to its flying distance and found an ILP solution. 1490 In Poikonen and Golden [87], the problem is formulated as an 1491 ILP. However, it can only be solved via a heuristic approach. 1492 In addition, the energy consumption function of the drone is 1493 linearized using linear regression and a MILP solver is used 1494 in [63].

1552
In this section, we provide several observations regarding 1553 HTDDS problems and identify research gaps along with 1554 future research directions. We also discuss the roadblocks to 1555 HTDDS implementation.

1557
Despite the strong interest in the research of HTDDS, it is 1558 still in its early infancy and requires improvements in various 1559 dimensions. Most of the reviewed literature has considered 1560 drones with limited payload capacity (single delivery per 1561 dispatch) and fixed speed, range, and energy consumption. 1562 Adding the variability of the drone characteristics will pro-1563 vide a more accurate routing solution. Research gaps related 1564 to drone technical characteristics are discussed below.

1565
• Almost 86% of the surveyed literature limits the drone's 1566 ability to carry only a single package per dispatch. 1567 With the current advancements made in the drone's 1568 capabilities, drones can now carry multiple packages 1569 as presented in [146] and [147]. The consideration of 1570 multiple packages will certainly reduce the comple-1571 tion time of delivery operations in the HTDDS while 1572 increasing the complexity of finding the best delivery 1573 routes.

1574
• Many of the reviewed literature ignored the variability of 1575 the drone's range. Thus, models should reflect the actual 1576 drone range affected by the flight profile of the drone, 1577 which may include the vertical traveling distance and 1578 hovering. In cases where drones need to reach customers 1579 in buildings, the vertical distance should be taken into 1580 account in the drone's range calculations [79]. Further, 1581 the range of the drone is affected by the drone's bat-1582 tery capacity limit, which is influenced by the payload 1583 weight during operations [63].

1584
• The drone's speed is a critical aspect that has an 1585 impact on the drone's energy consumption, range, and 1586 endurance. The consideration of varying drone speed is 1587 scarcely limited in the literature to [64], [86] who have 1588 addressed it. Treating the drone's speed as a variable also 1589 affects the operational cost, service time, and completion 1590 time of total delivery, all of which are critical to the LMD 1591 challenges. In addition, the consideration of variable 1592 speeds is essential when allowing the launching/retrieval 1593 of the drone to occur while the truck is in motion. The 1594 truck when merging with the drone needs to increase its 1595 speed while the drone should reduce its speed. This type 1596 of operation is demonstrated by Amazon patents such 1597 as the train-mounted mobile hubs for drone delivery 1598 that allows the handover of delivery parcels to occur 1599 while the train is in motion [48]. In a similar patent, the 1600 Amazon Airborne Fulfillment Center permits the mobile 1601 replenishment of drones by small airships [48].

1602
• Realistic drone energy or power consumption is a critical 1603 aspect when modeling the HTDDS. The energy con-1604 sumption model is mainly a function of the payload 1605 weight, the speed of the drone, and the drone's self-1606 weight. It determines the battery life of the drone and 1607 thus reflects the actual performance of the UAV. It also 1608 reflects the drone's environmental impact and the asso-1609 ciated cost savings [80]. Therefore, it will be interesting 1610 to see more studies focusing on considering the drone's 1611 technical characteristics in novel ways.  Drone technology is rapidly evolving. It is important to 1684 develop the infrastructure and tools required to ensure proper 1685 handling in the delivery of different product types. The type of 1686 product being delivered dictates the requirements needed to 1687 maintain the quality of the goods. For example, assessing the 1688 stresses that may be encountered during drone delivery such 1689 as vibration, humidity, and temperature excursions, should 1690 be included as all of these can affect the critical properties 1691 of shipped products. Add to that, docking hubs or drone 1692 stations used for storing and facilitating the coordination 1693 between trucks and drones are attracting market for drone 1694 delivery [148]. This is because drones may need special 1695 conditions for a safe and secure landing [149]. However, 1696 a limited number of studies have introduced this concept and 1697 have predefined the locations of the hubs. One remarkable 1698 consideration is to determine the location, size, configuration 1699 of the hubs, and the allocation of drones to achieve an effec-1700 tive HTDDS.

1701
As most of the literature focuses on the use of a single truck 1702 with a single drone, attention should be given to extending 1703 these systems to multiple trucks and multiple drones as well 1704 sequence of drone launches/collections). mercial application of HTDDS is near [47], [48], [157] which  The LMD is a very complex process that needs proper plan-1729 ning and managerial decisions to achieve optimal outcomes.   Another aspect is the energy consumption of the drone, 1759 which provides a better understanding of the cost and 1760 environmental impact of the drone. The modeling of the 1761 drone's energy consumption is affected by the design of the 1762 drone, weather conditions, and limited drone battery capacity. 1763 Current battery-powered drones have significant range and 1764 endurance limitations. Future research should concentrate on 1765 alternate energy sources such as fuel cells and solar pan-1766 els. This will alleviate the battery limitation that presently 1767 restricts the use of UAVs efficiently and result in significant 1768 cost savings when combined with traditional delivery vehi-1769 cles for LMD.

1770
The usability of drones is limited by the need for fre-1771 quent recharging. In HTDDS literature, battery recharging is 1772 assumed to occur instantaneously and has a fixed amount of 1773 time rather than optimizing the required charging of the drone 1774 in a cost-optimal way. The optimization of dynamic battery 1775 recharging will result in more accurate routing. To do so, 1776 locating charging stations with a peer-to-peer network can be 1777 a highly promising solution [158]. The peer-to-peer network 1778 can assist drones in reserving a recharging slot at the nearest 1779 charging station on their route at the lowest possible cost. 1780 Thus, allowing drones to fly for extended periods of time. Delivery processes are disrupted by unpredictable situations 1783 that may occur such as congestion, severe weather condi-1784 tions, service time, and the dynamic restrictions of the UTM 1785 imposed on the drones. Adequate route planning is necessary 1786 for avoiding delivery delays, customer dissatisfaction, and 1787 high costs. Incorporating spatial and temporal constraints is 1788 critical for creating more realistic models. Specific regions 1789 will almost stay off-limits to drone fly-over, therefore ideal 1790 routes must account for these constraints. In addition, such 1791 uncertainty sources have a significant influence on drone 1792 safety and security. However, these critical uncertainty issues 1793 are not adequately addressed in the literature.  The most crucial requirement for ensuring customer loyalty 1803 is visibility into the delivery operations. Customers want to 1804 know exactly where their package is and when it will arrive. 1805 Drones have distinct advantages in addressing the LMD 1806 obstacles. However, their use raises a number of security 1807 concerns. These security concerns pervade the whole deliv-1808 ery business process, from package pickup to final payment 1809 by the consumer upon successful parcel receipt. Tradition-1810 ally, each company's business data is stored independently 1811 in logistics systems. When disagreements arise, third-party 1812 arbitration is frequently required. In practice, however, few reliable third parties can guarantee the impartiality of the 1814 arbitration. Currently, the majority of existing UAV delivery 1815 systems are based on cloud computing, which cannot match 1816 the requirements of many real-time services in UAV delivery systems efficiently [162]. Security difficulties in UAV delivery systems have also been raised due to the presence 1819 of numerous parties who may not have a mutual trust rela-1820 tionship. In the UAV delivery system, drones will not be able  [167], disaster management [168], [169], and others 1847 to explore the different modeling aspects that could affect