Accepted Manuscript

Scheduling last-mile deliveries with truck-based autonomous robots

Nils Boysen, Stefan Schwerdfeger, Felix Weidinger

PII: S0377-2217(18)30477-6
DOI: 10.1016/j.ejor.2018.05.058
Reference: EOR 15171

To appear in: European Journal of Operational Research

Received date: 1 August 2017

Revised date: 23 May 2018
Accepted date: 27 May 2018

Please cite this article as: Nils Boysen, Stefan Schwerdfeger, Felix Weidinger, Scheduling last-mile
deliveries with truck-based autonomous robots, European Journal of Operational Research (2018),
doi: 10.1016/j.ejor.2018.05.058

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Highlights
• We consider a novel last-mile concept relying on autonomous robots.

• A novel scheduling problem for launching robots from trucks is derived.

• Efficient solution procedures for this problem are provided.

• We benchmark the novel concept with traditional truck deliveries.

T
IP
CR
US
AN
M
ED
PT
CE
AC

1
 ACCEPTED MANUSCRIPT

Working Paper

T
IP
Scheduling last-mile deliveries with truck-based autonomous robots

CR
Nils Boysen1,∗ , Stefan Schwerdfeger2 , Felix Weidinger1

US
August 2017
AN
Revised: March 2018
M
ED
PT
CE

1
: Friedrich-Schiller-Universität Jena
Lehrstuhl für Operations Management
Carl-Zeiß-Straße 3, 07743 Jena, Germany
AC

http://www.om.uni-jena.de
{nils.boysen,felix.weidinger}@uni-jena.de

2
: Friedrich-Schiller-Universität Jena
Lehrstuhl für Management Science
Carl-Zeiß-Straße 3, 07743 Jena, Germany
http://www2.wiwi.uni-jena.de/Entscheidung/
[email protected]

∗
Corresponding author, phone +49 3641 9-43100
 ACCEPTED MANUSCRIPT

Abstract

To reduce the negative impact of excessive traffic in large urban areas, many inno-
vative concepts for intelligent transportation of people and freight have recently been
developed. One of these concepts relies on autonomous delivery robots launched from
trucks. A truck loads the freight dedicated to a set of customers in a central depot and
moves into the city center. Also on board are small autonomous robots which each can
be loaded with the freight dedicated to a single customer and launched from the truck.
Then, the autonomous robots move to their dedicated customers and, after delivery,
autonomously return to some robot depot in the city center. The truck can replenish
robots at these decentralized depots to launch further of them until all its customers are
supplied. To assess the potential of this innovative concept, this paper develops schedul-

T
ing procedures which determine the truck route along robot depots and drop-off points

IP
where robots are launched, such that the weighted number of late customer deliveries
is minimized. We formulate the resulting scheduling problem, investigate computational
complexity, and develop suited solution methods. Furthermore, we benchmark the truck-

CR
based robot delivery concept with conventional attended home delivery by truck to assess
the potential of this novel last-mile concept.

Keywords: Scheduling; Transportation; City logistics; Autonomous robots

1 Introduction
US
AN

To attenuate the negative effects of transportation on congestion, safety, and environment in

large city centers, plenty of innovative concepts for moving people and freight have recently
M

been developed. Among those concepts, especially focusing on freight transportation, are,
for instance, goods distribution with electric vehicles [17], drone-based freight transport [16],
delivery into the trunk of a parked car [18], and crowdsourcing of deliveries [4]. An overview of
ED

the latest developments and concepts in city logistics is, for instance, provided by Savelsbergh
and Van Woensel [19]. The novel concept focused in this paper relies on autonomous robots
launched from a truck (see Figure 1).
PT
CE
AC

Figure 1: Truck-based autonomous robots1

In September 2016, German truck producer Mercedes-Benz Vans announced a strategic part-
nership with Starship Technologies [8]. The latter is an Estonian start-up company which
1
Source: Daimler https://www.daimler.com/innovation/specials/future-transportation-vans/

1
 ACCEPTED MANUSCRIPT

develops autonomous robots for last-mile deliveries. Starship’s robots move along sidewalks
and weigh no more than 40 pounds, fully loaded. They can be applied to deliver parcels
or groceries directly from stores or specialized hubs. Customers can monitor deliveries via
smartphones which are also applied to open the locked cargo bay of the robots upon arrival.
Afterwards, the robots autonomously return to their store or hub. Due to safety reasons, the
robots are only permitted to move at pedestrian speed, so that either a dense (and costly)
network of stores or depots is required or (comparatively) long delivery times have to be ac-
cepted. To avoid these drawbacks, the aforementioned alliance advocates a concept where
trucks are used as mobile launching platforms for the robots.
The truck-based robot delivery concept works as follows. A truck loads the shipments for a
set of customers at a central depot where the goods to be shipped are stored. A fixed part of

T
the truck’s loading capacity is reserved to also load autonomous robots on board. The truck,

IP
then, moves into the city center and, once a drop-off point is reached, one or multiple robots
are loaded with shipments and launched to autonomously deliver their goods to customers.

CR
Each robot has a capacity for a single shipment and after delivery, returns to a decentralized
robot depot within the city center. Note that at these decentralized depots, only robots are
stored but not goods, so that only a small garage is required but no complete distribution
center. The truck moves onwards to successive drop-off points until all robots are released.

US
If the truck still has further shipments for additional customers on board, it can move to one
of the decentralized robot depots to load another batch of robots. In this way, the process
goes on until all shipments on board of the truck are launched with a robot and the truck can
AN
return to the central depot in order to load more shipments for the next set of customers.
This paper introduces scheduling procedures for an efficient truck-based robot delivery. Specif-
ically, we aim at a truck route along a sequence of stops consisting of drop-off points and
M

decentralized robot depots along with a launching plan of robots, such that the weighted
number of late customer deliveries, after the announced delivery date, is minimized. This
problem is defined, computational complexity is proven, and exact as well as heuristic solution
ED

methods are introduced. The devil’s advocate could say that the second problem is treated
before the first. First, all technological challenges should be solved and government admission
for the robots on public roads should be reached (first problem). Afterwards, there is still
PT

enough time to develop suited scheduling procedures (second problem). However, it seems
hard to properly anticipate the potential gains of the truck-based robot delivery concept with-
out a detailed scheduling procedure which is able to exactly quantify important performance
CE

indicators, such as the resulting travel distances. Therefore, to get a first impression of the
efficiency of a truck-based robot delivery, we apply our scheduling procedures to different data
sets and benchmark them with traditional attended home delivery by truck. Furthermore, we
compare our approach to an alternative process of the truck-based robot delivery concept.
AC

Instead of applying decentralized robot depots, the truck could also wait for the return of
launched robots. Our computations, however, show that, due to the robots’ low travel speed,
this alternative mode of operation is considerably inferior.
The remainder of the paper is structured as follows. A brief literature review is provided
in Section 2. Then, Section 3 defines our optimization problem and Section 4 elaborates
on suited exact as well as heuristic solution procedures, whose computational performance is
tested in Section 5. The comparison of the two alternative modes of operation of truck-based
robot deliveries and their benchmarking with conventional attended home deliveries by truck
is provided in Section 6. Finally, Section 7 concludes the paper.

2
 ACCEPTED MANUSCRIPT

2 Literature review
Several innovative concepts for the transport of people and goods have recently been devel-
oped. Instead of trying to summarize the vast body of literature which has accumulated in
this area, we refer to the recent survey papers on trends in transportation [21], distribution
with electric vehicles [17], shared transportation [15], and city logistics [19]. One of the latest
concepts, announced in September 2016 [8], is truck-based robot delivery. To the best of the
authors’ knowledge, this paper is the first to derive scheduling procedures for this concept.
Therefore, our literature survey will only address two related fields: (i) truck-based drone
delivery and (ii) related transportation problems with similar mathematical structure.

T
(i) Another innovative last-mile concept where smaller delivery vehicles are launched from a
truck serving as a mobile depot is truck-based drone delivery (see [2]). Here, unmanned aerial

IP
vehicles (also denoted as drones) are launched from the top of a truck to deliver shipments to
customers. In spite of all physical differences between the delivery vehicles, from the planning

CR
perspective, both concepts are very similar. Drones also have a small capacity and can service
just one customer at a time. The main differences, however, which lead to distinct routing
problems, are the following:

US
• Our autonomous robots require the customers to be at home, so that a person can unlock
the robot’s cargo bay and withdraw the shipment. Drones can deliver unattended, for
AN
instance, just be releasing their shipment on the balcony of a customer’s apartment.
Therefore, existing papers (i.e., [2, 16, 22, 5]) on truck-based drone deliveries do not
consider customer due dates, but rather minimize the makespan until all customers are
serviced.
M

• Furthermore, drones travel unobstructed at a high speed through the air, so that the
truck can wait for the return of a drone after each delivery. This is also a possible mode
ED

of operation for our autonomous robots and we benchmark this alternative in Section
6. However, our robots only travel at pedestrian speed and unavailable customers may
further delay the robots. Therefore, this mode of operation would cause excessive waiting
PT

times for the truck. Small robot depots, to which the robots can autonomously return
and where the truck can refill its capacity for robots, seem to be the better choice. To
the best of the authors’ knowledge, comparable drone depots have not been considered
CE

in the literature yet.

(ii) As the routing of vehicles has a long tradition, it is anything but astounding that there
AC

are quite a few related problems which share some characteristics with our truck-based robot
delivery:

• Time-related versions of the traveling salesman problem (TSP), such as the traveling
repairman problem (also known as minimum latency problem [20]), come pretty close to
our problem. In this extension of the TSP, the travel between a pair of cities is related
with a travel time and the objective is to minimize the sum of travel times along the tour
towards each city (e.g., [1]). There are also other variants like the traveling repairman
problem with time windows [12] and the time-dependent TSP [11]. Our problem also
has a time-dependent objective function and aims to minimize the weighted number of
late deliveries. However, in contrast to the TSP (and its aforementioned variants) we

3
 ACCEPTED MANUSCRIPT

do not have to visit each drop-off point but just a selection of them. Moreover, the
same robot depots may be visited multiple times and customers (cities) are not directly
approached by the truck but only by launched robots.

• Another variant of the TSP related to our problem is the covering salesman problem
introduced by Current and Schilling [7]. Here, the shortest tour along a subset of given
nodes is sought, such that every node that is not on the tour is within a predefined
covering distance of a node on the tour (see also [3]). Similar to our problem, not every
node is to be visited and reducing the distances for the robots between their launching
point and the customer they serve is certainly a good idea. In our problem, however,
each customer has an individual due date for parcel delivery, so that a detailed launching

T
schedule of robots is required.

IP
• In our problem, the route of the truck and the launching schedule of robots have to
be synchronized, so that our problem falls into the domain of vehicle routing problems

CR
with synchronization (for a survey see [9]). Among these problems, the truck and trailer
routing problem (TTRP) comes quite close to our problem. In the TTRP, a vehicle
composed of a truck with or without a detachable trailer serves the demand of a set of
customers reachable by truck and coupled trailer or just by the truck (e.g., see [14]).

US
The main difference between TTRP and our problem is that the robots, once launched,
separately serve customers, while the trailer cannot serve customers on its own and
remains where it is parked.
AN

To conclude, this paper introduces a novel routing problem that has not been treated in the
literature yet.
M

3 Problem definition
ED

Consider a single truck loaded with shipments for a set C of customers to be supplied. Initially,
the truck is positioned at location γ with δ robots on board. This location may be the central
PT

depot where the customers’ shipments are loaded and an initial delivery schedule is determined.
In case of unforeseen events, such as congestion, however, a short-term adaption of the initial
plan may become necessary, so that γ and δ can also represent the truck’s current status-quo
CE

on a half-finished tour once a replanning is initiated. Each customer j ∈ C is associated with

a deadline dj , representing the latest desirable delivery time, and a weight wj . This weight
quantifies the penalty term if the customer’s shipment arrives after the deadline.
AC

Our problem assumes a given discrete set P = C ∪ D ∪ R of locations which can be visited
by the truck and the robots. This set contains the following location types:

• Set C defines the customer locations, so that Cj defines the location of customer j.
Customers get exclusively supplied by robots and not directly by the truck, so that only
robots may access these locations.

• Set D represents the drop-off points where the truck stops, loads robots with shipments,
and launches them towards a customer.

4
 ACCEPTED MANUSCRIPT

• Finally, we have set R representing decentralized robot depots. Here, the truck can also
launch (a non-restricted number of) robots and, before departure, replenish robots. Up
to K new robots can be taken on board where K defines the truck’s maximum robot
capacity.

Given the speed of the truck as well as the much lower velocity of the robots and the distances
among all locations P, parameters ϑtv,v0 and ϑrv,v0 define the travel time of truck and robot,
respectively, between two points v, v 0 ∈ P.
One part of the solution for the truck-based robot delivery scheduling (TBRD) problem is
a truck route. A route is defined by a sequence of stops starting at initial location γ. Each

T
following stop of the truck can only be at a drop-off point D or a robot depot R. Furthermore,
a solution for TBRD requires the definition of a robot launching schedule. Such a schedule

IP
defines the number of robots launched from a stop of the truck and the subset of customers
supplied from there. Each robot can only load one shipment at a time, so that the number of

CR
robots launched has to equal the cardinality of the customer subset supplied. The objective of
TBRD is to find a truck route and a launching schedule that minimizes the weighted number
of late deliveries. More formally, TBRD can be defined as follows:

US
A solution π for TBRD consists of a sequence of tuples (v, L) defining the subset L ⊆ C of
customers supplied by the robots launched from location v ∈ D ∪ R. We say a solution π is
feasible, if the following conditions hold:
AN
S
• We have (v,L)∈π L = C, that is, all customers are supplied.

• For any pair of distinct tuples (v, L) ∈ π and (v 0 , L0 ) ∈ π, we have L ∩ L0 = ∅, that is

M

each customer is serviced by a single robot launched from a single location.

• For the very first tuple (v, L) of π, we have v = γ, that is the truck route starts at the
ED

truck’s initial position γ.

• Let π 0 be the set of tuples referring to the foremost drop-off

P points of π before the
first decentralized robot depot is accessed. We have (v,L)∈π0 |L| ≤ δ, that is, no
PT

more robots can be launched before reaching the first robot depot than those that were
initially loaded.
CE

• Let π i be the set of tuples referring to all drop-off points following on the i-th visit at
a robot depot before either the next robot depot is visited P or the end of the route is
reached. For all visits i of robot depots within π we have (v,L)∈πi |L| ≤ K, that is at
AC

most K robots can be launched after any robot depot visit. Once the capacity K for
robots is depleted and further customers have to be supplied, another depot has to be
visited to replenish the robots on board of the truck.

Let σ(j) and vi return the number of the stop in the truck route of π where the robot supplying
customer j ∈ C is launched and the location of the i-th stop within π, respectively. Then,
the delivery time ϕj of a customer j is

σ(j)−1
X
ϕj = ϑtvi ,vi+1 + ϑrvσ(j) ,j .
i=1

5
 ACCEPTED MANUSCRIPT

Delivery time ϕj amounts to the driving time of the truck from its initial location up to the
location where the robot supplying customer j is launched plus the driving time of the robot
from its launching location up to the location Cj of the customer. Consider U being the set of
customers supplied late, i.e., those customers with ϕj > dj , TBRD seeks one feasible solution
that minimizes the weighted number of late deliveries, i.e., minimizes

X
Z(π) = wj .
j∈U

T
IP
CR
US
AN
Figure 2: Two alternative solutions for an example instance of TBRD

Example: Consider the example of Figure 2 where two alternative solutions for an instance
M

of TBRD are depicted. Here, eight customers are to be supplied which are represented by
the (non-)filled circles. The deadlines of the customers are given above the circles and all
customers are assumed to have unit weights. The filled triangle defines the start position
ED

of the truck which starts without robots on board and has a capacity of K = 3. Drop-off
points (robot depots) are symbolized by non-filled triangles (squares). The arcs starting from
the truck’s starting point γ represent the truck routes which only go through drop-off points
and robot depots. Robot deliveries either from drop-off points or depots to customers are
PT

symbolized by dashed lines. The duration of truck and robot movement along the respective
relations are given by the weights. Customers supplied late, where the shipments arrive after
the deadline, are highlighted by filled circles. Thus, solution A is better than solution B with
CE

two delayed shipments less.

Our TBRD is based on some (simplifying) assumptions that are elaborated in the following:
AC

• The truck exclusively acts as a mobile depot which has the shipments on board and
launches robots. It does not access customer locations to also service customers directly.
We leave the evaluation of this alternative mode of operation, where we can choose
between robot and truck delivery, up to future research.

• We presuppose that robots are not a bottleneck resource. This implies that we have
no shortages of robots at depots and trucks can always be replenished with robots to
capacity. If robots are scarce, they need to be modeled individually which adds plenty of
complexity to the problem. Their return to the robot depots would have to be considered,
so that either only fewer robots can be loaded by the truck or additional waiting time

6
 ACCEPTED MANUSCRIPT

for returning robots occurs. We leave this extension of our basic TBRD up to future
research.

• If we assume that there are always enough robots, this also impacts the objective. The
robots have to be unloaded at each customer location, so that we have attended home
delivery. This necessitates that a delivery time window has been agreed with each
customer in which he/she has to be at home to receive the shipment. If robots are no
bottleneck, however, a premature arrival at a customer location is unproblematic because
the robot can wait until the time window is reached and the customer is available.
Only an arrival after the time window is to be avoided because otherwise unsatisfied
customers result. Therefore, the end of the delivery time window agreed with the

T
customer constitutes the deadline dj within our TBRD.

IP
• We restrict our problem to a single delivery vehicle serving a given set of customers.
Thus, we assume that a distribution of customers among multiple trucks has already

CR
been planned on a higher planning level. A dynamic adaption of the customer set, which
may occur if a customer cancels a shipment on short notice, is not considered. However,
in this case a new planning run of TBRD for the altered situation can be executed even
when the truck is already on its way.

US
• We presuppose deterministic travel times of truck and robots. A simple way to integrate
the inevitable uncertainty coming along with these values is to integrate safety buffers,
AN
e.g., by reducing either the customers’ deadlines or travel speeds of the vehicles. Recall
that also a new planning run of TBRD can be executed whenever travel times have
(considerably) altered.
M

• To keep the problem structure as basic as possible, we do not consider the stop times
of the truck at drop-off points and depots explicitly. Compared to the driving times
stop times are short, so that they seem negligible and not that much precision should
ED

be lost. However, waiting times of the truck can be integrated implicitly by adding a
service time s (stop times, time to replenish/launch robots) to the travel times ϑtv,v0 as
a constant factor, i.e. ϑ̄tv,v0 = ϑtv,v0 + s. Furthermore, in a preprocessing step the time
for launching robots can be either added to travel times ϑrv,v0 or respectively subtracted
PT

from the customers’ deadlines dj .

CE

Given this problem definition, the following theorem constitutes the complexity status of
TBRD.

Theorem 1. TBRD is N P-hard in the strong sense.

AC

The proof is by reduction from the feasibility version of the traveling salesman problem (TSP)
which asks whether a feasible tour exists that visits each of the given n cities exactly once,
returns to the origin city, and has a tour length not exceeding T . The TSP is well-known to
remain N P-complete even if only a path through all cities is sought and the starting city is
fixed [10].

Proof. Our transformation scheme from TSP to TBRD is defined as follows. The starting
city of the TSP is also the starting location γ of the truck which, initially, has δ = n − 1
robots on board. At each location of the n − 1 remaining cities of the TSP, we also have a

7
 ACCEPTED MANUSCRIPT

drop-off point within TBRD and a customer location in direct vicinity. Any pair of customer
location and drop-off point referring to the same city of TSP have no travel distance between
each other. Note that if per problem definition drop-off points and customers cannot reside
at the same location, this proof also works with a tiny distance among them. To keep this
proof as simple as possible, we, however, assume that they may have the same location. The
truck moves with a velocity of one distance unit per time unit and the speed of the robots
is prohibitively slow, i.e., smaller than d/T where d denotes the smallest distance among any
pair of distinct cities of the TSP instance. The deadlines and weights of all customers j ∈ C
are unanimously set to dj = T and wj = 1, respectively. The question we ask is whether a
feasible solution exists where all customers are timely supplied, i.e., Z = 0.
First, we show that each feasible solution to TBRD is also a feasible TSP solution. Due to

T
the prohibitively low robot speed, each customer’s robot has to be launched at the drop-off

IP
point directly next to the customer. Otherwise, even if launched from the closest alternative
drop-off point the respective customer will inevitably be supplied too late. Furthermore, we

CR
have no robot depot and enough robots initially on board, so that each pair of drop-off point
and customer, corresponding to a city of TSP, can be visited exactly once without robot
replenishment. The delivery time of each customer is only determined by the travel time of
the truck up to the respective customer. As all customers have to be timely supplied (even the

US
last one), the total truck route through all pairs of drop-off point and customer may therefore
not exceed T . Just the same tour through all corresponding cities is also a feasible TSP
solution with a tour length smaller than or equal to T .
AN
Finally, each feasible solution of TSP is also a YES-instance of TBRD. The truck route with
TBRD just goes through the corresponding pairs of drop-off points and customers and the
tour length is identical.
M

4 Solution methods
ED

This section is dedicated to suited solution procedures of our TBRD problem. After present-
ing a MIP model in Section 4.1, we provide an efficient approach to determine an optimal
PT

assignment of customers to drop-off points and robot depots for a predetermined truck route
in Section 4.2. Based on this, we introduce a tailor-made local search procedure in Section
4.3.
CE

4.1 Mixed-integer model for TBRD

AC

Applying the notation as summarized in Table 1, our mixed-integer program (TBRD-MIP)

consists of objective function (1) and constraints (2) to (16).

X
TBRD-MIP: Minimize F (L, S, T, X, Z) = zk · wk (1)
k∈C

8
 ACCEPTED MANUSCRIPT

C set of customers
D set of drop-off points
R set of robot depots
dk deadline of customer k
wk weight of customer k
δ number of robots initially loaded on the truck
γ, γ e initial position, final (dummy) position of the truck
K the truck’s maximum loading capacity for robots
ϑtv,v0 , ϑrv,v0 travel time of truck and robots between two points v, v 0
lj continuous variables: amount of robots on board at the departure from
location j

T
si,j binary variables: 1, if location j is the direct successor of location i; 0,

IP
otherwise
tj continuous variables: arrival time at location j
xj,k binary variables: 1, if customer k is supplied from location j; 0, otherwise

CR
zk binary variables: 1, if customer k is supplied late; 0, otherwise

Table 1: Notation for MIP model

X

j∈D∪R∪{γ}
xj,k = 1
X
US ∀k ∈ C (2)
AN
lγ ≤ δ − xγ,k (3)
k∈C
X
lj ≤ K + M · (1 − si,j ) − xj,k ∀ i ∈ R; j ∈ D (4)
M

k∈C
X
lj ≤ li + M · (1 − si,j ) − xj,k ∀ i ∈ D ∪ {γ} ; j ∈ D \ {i} (5)
k∈C
ED

tγ = 0 (6)
tj ≥ ti − M · (1 − si,j ) + ϑti,j ∀ i ∈ D ∪ R ∪ {γ} ; j ∈ D ∪ R \ {i}
(7)
PT

M · zk ≥ tj − M · (1 − xj,k ) + ϑrj,k − dk ∀ j ∈ D ∪ R ∪ {γ} ; k ∈ C (8)

X
sγ,j ≤ 1 (9)
CE

j∈D∪R∪{γ e }
X X
sj,i = si,j ∀j ∈ D ∪ R (10)
i∈D∪R∪{γ e }\{j} i∈D∪R∪{γ}\{j}
AC

X X
xj,k ≤ M · si,j ∀j ∈ D ∪ R (11)
k∈C i∈D∪R∪{γ}\{j}
X X
xj,k ≥ si,j ∀j ∈ D (12)
k∈C i∈D∪R∪{γ}\{j}
X X X
xj,k + sj,i ≥ si,j ∀j ∈ R (13)
k∈C i∈D i∈D∪R∪{γ}\{j}

si,j ∈ {0, 1} ∀ i ∈ D ∪ R ∪ {γ} ; j ∈ D ∪ R ∪ {γ e } \ {i}

(14)
xi,k , zk ∈ {0, 1} ∀ i ∈ D ∪ R ∪ {γ} ; k ∈ C (15)
lj ≥ 0 ∀ j ∈ D ∪ {γ} (16)
9
 ACCEPTED MANUSCRIPT

Objective function (1) minimizes the weighted number of late deliveries which are determined
by constraints (8). Constraints (2) ensure a one-to-one mapping between customers and
locations where the robots were launched. The amount of robots on board is controlled by
constraints (3) to (5), i.e., the initial loading (3) and the loading after leaving a drop-off
point having a robot depot (4) or a drop-off point (5) as direct predecessor, respectively.
While constraints (6) and (7) bound the arrival times at each location from below, subtour
elimination constraints are established by (9) and (10). Constraints (11) guarantee that robots
are launched from visited locations only. Note that constraints (12) and (13) are useful to
eliminate redundant subtours. Specifically, these constraints enforce that drop-off points are
part of the tour only if robots are launched from there and that depots are visited only if either
robots are launched from there or a drop-off point is visited next. Finally, the domains of the

T
variables are set by (14) to (16).

IP
Solving TBRD instances with the proposed model requires the following points to be consid-
ered:

CR
• TBRD-MIP can directly be applied if the truck’s starting point γ is a drop-off point.
However, in case γ is a robot depot, we have to execute the following preprocessing
steps. First, we remove all customers k ∈ C from our customer set which can be

the truck is loaded withP US

supplied in time from the current depot (i.e., ϑrγ,k ≤ dk ). Then, we set δ = K, so that
robots to capacity and ensure that no customer is served from
the starting point (i.e., k∈C xγ,k = 0). Afterwards, we execute the model as described
AN
above.
• Our problem setting allows to visit locations, e.g., robot depots, more than once. To
enable this property, we can simply duplicate locations. More precisely, if all deliveries
M

are executed from a single drop-off point, d|C|/Ke visits are necessary. Thus, depending
on the initial amount of loaded robots, d|C|/Ke + 1 duplicates
l mfor each drop-off point
|C|
are required. In addition, a depot is visited at most d(K+1)/2e times because in the
ED

worst case d(K + 1)/2e robots are required at each drop-off point.
• To keep the model as simple as possible, it is desirable to reduce the number of duplicates
PT

to a minimum. First, we are allowed to remove  all customers

 k ∈ C who cannot be
delivered in time. This is given whenever min ϑrγ,k , min j ∈ D ∪ R|ϑtγ,j + ϑrj,k > dk
holds. Next, if it is not possible to supply a set of customers C̄ ⊆ C from drop-off point
CE

j ∈ D in time, i.e., we have ϑtγ,j + ϑrj,k > dk , k ∈ C̄, then we duplicate drop-off point
l m
j ∈ D only |C\KC̄| + 1 times.
AC

In spite of these considerations, however, it seems most likely that TBRD-MIP struggles
with large instances of real-world size. Therefore, the following sections provide an efficient
heuristic solution procedure. This heuristic repeatedly solves a subproblem where we derive
the launching plan of robots for a given truck route. This subproblem is shown to be solvable
to optimality in polynomial time first.

4.2 An efficient solution procedure for given truck routes

Once the truck route, i.e., a sequence of visits at drop-off points and robot depots starting
in γ, is given, TBRD reduces to find a launching plan of robots. The optimal assignment

10
 ACCEPTED MANUSCRIPT

of customer deliveries to the given stops of the truck can be determined by transferring the
problem to a Hitchcock transportation problem, which is well-known to be efficiently solvable
in polynomial time [13, 6]. Specifically, the transformation goes as follows.
Given a fixed truck route v, where vi denotes the i-th stop, we first define R(v) and S(v)
whichnrepresent theoset of all robot depots within truck route v and the set of all subtours
Sk = vk1 , . . . , vkik along drop-off points (vk1 , . . . , vkik ∈ D) either between two consecutive
depots (vk1 −1 , vkik +1 ∈ R) or the beginning or end of truck route v (k1 = 1/kik = |v|),
respectively. Given these definitions the supply and demand nodes as well as the transportation
costs among them, which define the Hitchcock transportation problem, are constructed as
follows:

T
IP
• A supply node with supply quantity ai = |C|, i = 1, . . . , |R(v)|, is added for each depot
ri ∈ R(v).

CR
• A supply node with supply quantity ak = K, k = |R(v)| + 1, . . . , |R(v)| + |S(v)|, is
added for each subtour Sk ∈ S(v). In case γ ∈ / R, the supply quantity of the first
subtour S|R(v)|+1 equals the initial amount of loaded robots a|R(v)|+1 = δ.

US
• Each customer j ∈ C constitutes a demand node Cj with demand quantity bj = 1.

We introduce a single dummy demand node d with demand quantity b0

• P =
AN
|R(v)|+|S(v)|
i=1 ai − |C|.

• Finally, we specify the transportation costs ci,j between supply and demand nodes. The
costs from each supply node to the dummy demand node are set to zero. Costs ci,j
M

between nodes referring to a depot ri ∈ R(v) at stop vr(i) and a customer j ∈ C are
set to zero if the supply is in time and wj otherwise, i.e.,
(
ED

Pr(i)−1 t
0 , if l=1 ϑvl ,vl+1 + ϑrvr(i) ,j ≤ dj
ci,j =
wj , else.
PT

Furthermore, costs ck,j between nodes referring to a subtour Sk ∈ S(v) and a customer
j ∈ C are set zero if there is at least one drop-off point vki (i = 1, . . . , ik ) within Sk ,
so that the supply is in time and wj otherwise, i.e.,
CE

( nP o
i−1 t
0 , if mini=k1 ,...,kik ϑ
l=1 vl ,vl+1 + ϑr
vi ,j ≤ dj
ck,j =
wj , else.
AC

Given this bipartite graph, which consists of supply and demand nodes, solving the Hitchcock
transportation problem yields the optimal min-cost transportation plan in polynomial time. By
considering which drop-off points and robot depots the supply nodes and which customers the
demand nodes refer to, we can directly derive the optimal launching plan by minimizing the
weighted sum of delayed deliveries (for the given truck route).
Example: Consider the example of Figure 3. On the left side, a problem instance of TBRD is
given with a fixed route v, a truck capacity for K = 1 robot, and an initial amount of loaded
robots δ = 1. The two customers C = {1, 2} have weights w1 = w2 = 1 and deadlines
d1 = 3 and d2 = 6, respectively. The truck route v = {v1 , v2 , v3 } contains a depot r1 at

11
 ACCEPTED MANUSCRIPT

T
IP
Figure 3: An instance of TBRD with a fixed route and the corresponding transportation

CR
problem

stop v3 with supply a1 = |C| = 2 and two drop-off points v1 = γ and v2 which form the

US
subtour S2 = {v1 , v2 } with supply a2 = δ = 1. The demand nodes (customers) have demands
b1 = b2 = 1 and dummy node d has a demand of b0 = 1. While customer 1 can be satisfied in
time only if its corresponding robot is launched within subtour S2 (at stop v2 ), customer 2 can
timely be supplied either from subtour S2 (at stop v2 ) or from depot r1 (at stop v3 ). Thus,
AN
we obtain the corresponding instance of the Hitchcock transportation problem given on the
right side of Figure 3. Solving this problem reveals that the min-cost transportation plan (i.e.,
having an objective value of 0) corresponds to a launching plan where customer 1 is served
M

from subtour S2 and customer 2 from depot r1 .

An optimal assignment of customers to depots and drop-off points for a given truck route
v may contain unnecessary drop-off points, where not even a single robot is launched, and
ED

depots, where robots are neither launched nor taken on board of the truck. If the triangle
inequality for our travel times ϑtv,v0 and ϑrv,v0 holds, which we assume throughout this paper,
then we can remove these superfluous stop from the truck route. According to the triangle
PT

inequality removing a stop will not enlarge the tour and therefore, will not increase the objective
function value. First, we are allowed to remove each superfluous drop-off point where no
robot is launched. Next, we remove each superfluous depot, which requires that no customer
CE

is served from this depot and the removal will not combine two subtours Sk , Sk+1 ∈ S(v) for
k > |R(v)| + 1 (k = |R(v)| + 1) with more than K (δ) customers to be served.
AC

4.3 A multi-start local search procedure

Based on the findings of Section 4.2, we propose a straightforward multi-start local search
procedure. This simple yet efficient optimization approach can be subdivided into two steps:
(1) Generation of a first tour employing priority rules and (2) ensuring feasibility of the tour
by considering capacity constraints and improving the feasible solution via local search.
In the first step, a start solution is generated where we considerably simplify the problem by
ignoring capacity constraints. The priority rules PR1 and PR2, employed to generate a first
tour of the truck, are solely based on travel times and deadlines. Note that we investigated a

12
 ACCEPTED MANUSCRIPT

larger selection of rules, but these were the two working best. Our formal description is based
on the notation introduced in Section 3.

• PR1: Move to position with most satisfiable customers

1. Start from initial position vn = γ at time t = 0, such that πn = (γ, ∅) with n = 1.

Initialize the set of already processed customers C̄ = ∅.
2. Determine all customers Cvtn satisfiable in time from current position vn at the
current point in time t as follows: Cvtn = {k ∈ C \ C̄ | t + ϑrvn ,k ≤ dk }. Update the
current tuple πn = (vn , Cvtn ) of customers served from the current stop and the
set of already processed customers C̄ = C̄ ∪ Cvtn . If all customers are processed,

T
that is C̄ = C, terminate.

IP
3. Determine the number of (not yet processed) customers satisfiable in time for each
t+ϑtvn ,i
depot and drop-off point Ci ∀i ∈ D ∪ R.

CR
t+ϑt
4. If no customer can be satisfied in time anymore, that is Ci vn ,i =
∅ ∀i ∈ D ∪ R, visit the nearest depot, set the current tuple as fol-
lows πn+1 = (arg mini∈R {ϑtvn ,i }, C \ C̄), and terminate. Otherwise, visit
the location with the highest number
(arg maxi∈D∪R {|Ci
t+ϑtvn ,i US of satisfiable customers next πn+1 =
|}, ∅), t = t + ϑtvn ,vn+1 , n = n + 1 and go to step 2.
AN
• PR2: Move to position with highest urgency

1. Start from initial position vn = γ at time t = 0, such that πn = (γ, ∅) with n = 1.

Initialize the set of already processed customers C̄ = ∅ and assign each customer
M

to its nearest depot or drop-off point: ĩk = arg mini∈D∪R {ϑri,k }.

2. (same as for PR1 )
ED

3. For each depot and drop-off point i ∈ D ∪ R determine the set of customers
t+ϑt
timely satisfiable assigned to i, which have not been processed yet C̃i vn ,i = {k ∈
C \ C̄ | ĩk = i ∧ t + ϑtvn ,i + ϑri,k ≤ dk }.
PT

t+ϑt
4. If no customer can be satisfied in time anymore, that is C̃i vn ,i = ∅ ∀i ∈
D ∪ R, visit the nearest depot, set the current tuple as follows πn+1 =
CE

(arg mini∈R {ϑtvn ,i }, C \ C̄), and terminate. Otherwise, visit the location with
the highest urgency, e.g., the smallest (positive) time budget to reach an as-
signed customer in time considering the velocity of truck and robot, πn+1 =
(arg mini∈D∪R {min t+ϑtvn ,i {dk −(t+ϑtvn ,i )−ϑri,k }}, ∅), t = t+ϑtvn ,vn+1 , n = n+1
AC

k∈C̃i
and go to step 2.

A solution generated in the first step may be infeasible as it can happen that either no depot is
visited at all or more than K customers are assigned to the same subtour. In the second step
we, therefore, fix infeasible solutions and try to further improve the obtained feasible solution
by a local search procedure. In the following, we name the exact, polynomial time procedure
introduced in Section 4.2 as function HT P (v) = π and the solutions generated by priority
rules PR1 and PR2 as πP R1 and πP R2 , respectively. The function Z(π) returns the objective
value of solution π. Applying this notation the second step proceeds as follows:

13
 ACCEPTED MANUSCRIPT

1. Try to improve πP R1 and πP R2 by applying HT P , such that πP R1 = HT P (vP R1 ) and

πP R2 = HT P (vP R2 ).

2. Generate solution pools ΠP R1 and ΠP R2 using depot insertions. A depot insertion is

possible at all index positions n > 1 of tour v, for which vn−1 , vn ∈ D hold. At
each potential insertion position, we only consider the one depot that causes the least
additional distance when added to the tour. In this way, we analyze all combinations of
possible depot insertions, such that having a tour with m successively visited pairs of
drop-off points 2m tours are evaluated. Pretests have shown that HT P is fast enough,
so that we can easily afford the evaluation of all additional tours. All these solutions are
added to solutions pools ΠP R1 or ΠP R2 , provided that a solution π 0 = (v 0 , HT P (v 0 ))

T
has an objective value less than the respective initial solution πP R1 and πP R2 that is
Z(π 0 ) < Z(πP R1 ) ∀π 0 ∈ ΠP R1 and Z(π 0 ) < Z(πP R2 ) ∀π 0 ∈ ΠP R2 , respectively.

IP
3. Evaluate solutions pools ΠP R1 and ΠP R2 multiple times in a parallelized manner using a

CR
multi-thread framework. Each pool is evaluated κ2rs times and each evaluation starts with
the initial solution πlocal = πP R1 or πlocal = πP R2 , respectively. During an evaluation,
the procedure iteratively picks a solution π 0 of the local copy of the selected pool by
random choice. Afterwards, a local search procedure as described below, is performed

US
for κls1 iterations starting from π 0 and πlocal is updated according to the result of this
search. π 0 and all solutions with objective values higher than or equal to πlocal are then
removed from the local copy of the selected pool. Once a pool contains no solution
AN
anymore, an additional run of local search is performed for κls2 iterations starting from
πlocal and the result of this search run is returned to the framework.

4. Return the best solution found during all pool evaluations.

M

The local search procedure performed multiple times during pool evaluation is based on the
following neighborhoods:
ED

• N1 : Remove a random position of the tour if it is not the single robot depot of the tour
(otherwise, feasibility is not guaranteed).
PT

• N2 : Insert a random depot or drop-off point into the tour at a random position.
CE

• N3 : Swap two randomly chosen positions of the tour.

• N4 : Select a position of the tour. If the selection is a depot, do nothing. Otherwise,

insert the depot closest to the drop-off point right before the selected position into the
AC

tour.

The resulting tour v 0 is transformed into a full solution of our TBRD (v 0 , HT P (v 0 )). If the
generated solution is better than the current one, it is accepted and set as the starting point of
preceding search steps. Otherwise, it is rejected and another neighbor of the current solution
is evaluated. The search is executed until the given number of iterations has been performed.
The whole multi-start local search procedure is summarized in Algorithm 1.

14
 ACCEPTED MANUSCRIPT

Algorithm 1: Multi-start local search procedure

1 solution πbest = null ; // best solution found during all restarts
2 πP R1 = HTP(solution found by PR1 );
3 ΠP R1 = set of tours generated by depot insertions into πP R1 ;
4 πP R2 = HTP(solution found by PR2 );
5 ΠP R2 = set of tours generated by depot insertions into πP R2 ;
6 for (i = 1; i ≤ κrs ; i += 1) do
7 if (i ≤ κ2rs ) then
8 πlocal = πP R1 ; Π0 = ΠP R1 .Clone();
9 else

T
10 πlocal = πP R2 ; Π0 = ΠP R2 .Clone();
11 while (|Π0 | > 0) do

IP
12 π 0 = randomly selected solution in Π0 ;
13 πlocal = LS(#iterations: κls1 , start solution: π 0 ); // start local search

CR
0 0 0 00 0 00
14 Π = Π \ {π } ∪ {π ∈ Π |Z(π ) ≥ Z(πlocal )};
15 πlocal = LS(#iterations: κls2 , start solution: πlocal ); // start local search
16 if (πbest = null OR Z(πlocal ) ≤ Z(πbest )) then
πbest = πlocal ; // update best solution during all restarts
17
18 return πbest ;
US
// return best solution during all restarts
AN
5 Performance of algorithms
In this section, we test the performance of our solution methods. Since no established testbed is
M

available for our TBRD, we first elaborate how our instances have been generated (see Section
5.1). Afterwards in Section 5.2, we benchmark the performance results of our heuristic solution
procedure with a standard solver solving TBRD-MIP.
ED

All computations have been executed on a 64-bit PC with an Intel Core i7-6700K CPU (4
x 4.0 GHz), 64 GB main memory, and Windows 7 Enterprise. The procedures have been
implemented using C# (Visual Studio 2015) and off-the-shelf solver Gurobi (version 7.0.2)
PT

has been applied for solving the MIP models.

CE

5.1 Instance generation

We generate two different sizes of test instances. The dataset labeled as small consists of
AC

instances which can be solved to proven optimality in a sufficient percentage of cases employing
a standard solver, while the dataset labeled large represents instances of real-world size. The
parameters handed over to our data generator are presented in Table 2. The procedure of
instance generation is summarized in the following.
Layout: First, we generate a square with side length w. The square is divided into a grid of
squares of side length 1/6 kilometers (about 1/10 miles). A reasonable policy when planning a
depot network is to equally spread the facilities in the urban area to be serviced. Therefore, we
allocate the given number of robot depots in an equidistant manner within our grid. Afterward,
we randomly scatter drop-off points and customers, making sure that each point is assigned
only once. Finally, a randomly chosen depot or drop-off point is set as the initial truck position,
such that γ ∈ D ∪ R (with δ = K).

15
 ACCEPTED MANUSCRIPT

symbol description small large

w side length of main square 2 km (1.3 miles) 5 km (3.1 miles)
|R| number of robot depots 4 16
|D| number of drop-off points 6 30
|C| number of customers 6 40
deadline factor interval
[ρmin ; ρmax ] {[2, 4]; [3; 5]} {[4; 12]; [4; 15]}
(named {tight, wide})
weights interval
[wmin ; wmax ] {[1; 1]; [1; 3]}
(named {homo, hetero})
K trucks robot capacity 2 8
km km
speed truck/robot 30 h (18.6 mph) / 5 h (3.1 mph)

T
Table 2: Parameter values for instance generation

IP
CR
Distances and travel times: We use the Euclidean metric to compute distances between points.
Based on these distances and given truck and robot speeds, travel times are calculated. Note
that using the Euclidean metric ensures that the triangle inequality is never violated considering
a single mode of transportation, e.g., trucks or robots.

US
Customer deadlines: Customer deadlines are generated on the basis of the travel times. First,
we determine the minimum travel time to each customer k ∈ C by ϑmin (k) = minj∈D∪R {ϑtγ,j +
ϑrj,k }. Second, we draw a random number rk from the interval (0, 1]. Then we determine the
AN
deadline of customer k as dk = ϑmin (k) · (ρmin + (ρmax − ρmin ) · rk ).
Weights: For each customer k ∈ C, we draw a weight wk randomly from interval [wmin ; wmax ]
applying a uniform distribution.
M

We generate ten geographical configurations per dataset. For each of these layouts, we com-
bine different deadlines and weights in a full factorial approach and repeat instance generation
ED

five times per parameter setting. In total, we receive 200 small and 200 large instances in each
dataset.
PT

5.2 Computational study of algorithm performance

In this section, we evaluate the computational performance of our solution methods. Table 3
CE

summarizes the results for our small dataset. For both solution methods, i.e., our multi-start
local search procedure presented in Section 4.3 and standard solver Gurobi solving TBRD-MIP
of Section 4.1, we report the number of best solutions found (best), e.g. how many times the
AC

solution approach was able to find the minimal known objective value, the average relative gap
to the optimal solution for all instances a proven optimum is available (gap), as well as the
average CPU-time in seconds (sec) required by the respective solution method per instance
of the whole dataset. Additionally, we report the number of proven optima found by Gurobi
(prov. opt.) and the number of optima found (but not proven) by the heuristic procedure
(opt.). Gurobi is applied with a time limit of 30 minutes. Based on the results presented in
Table 3, the following conclusions can be drawn:
Applying standard solver Gurobi solving TBRD-MIP, obviously, becomes easier if customer
deadlines are rather generously set and heterogeneous weights are chosen. However, a signif-
icant correlation between hetero- and homogeneous weights and the performance of Gurobi

16
 ACCEPTED MANUSCRIPT

Gurobi Heuristic
(time limit=1800sec) (κrs =100;κls1 =1000;κls2 =500)

prov.
best

best
opt.

opt.

opt.
gap
sec

sec
deadlines weights
thight homo 50 36 578.38 49 35 2.78% 0.77
hetero 50 37 513.88 50 37 0.00% 0.85
wide homo 50 43 278.60 50 43 0.00% 0.99
hetero 50 44 286.57 50 44 0.00% 1.04
total/average 200 160 414.36 199 159 0.63% 0.91

Table 3: Solution performance for the small dataset

is only available for the large dataset. For the small dataset, our heuristic procedure is able

T
to find all but one proven optima Gurobi found within its time limit. Indeed, except for this
single outlier, the heuristic solutions are all of the same quality than the ones found by Gurobi,

IP
proven optimum or not. Please note that in the single case our heuristic approach was not
able to achieve a best known solution, two instead of one customer are delivered late, which

CR
is the lowest deviation possible. Nonetheless, the average time consumption of the heuristic
is just 0.91 seconds. In comparison, Gurobi’s mean computational time amounts to 414.36
seconds.

Gurobi
US
(time limit=1800sec)
Heuristic
(κrs =80;κls1 =4000;κls2 =500)
prov.
best

best

best

best
opt.

opt.
gap

gap

gap

gap
AN sec

sec
LB

LB
deadlines weights
tight homo 7 6 248.60% 658.80% 1614.06 50 6 0.00% 117.60% 53.92
hetero 13 11 308.20% 758.60% 1496.51 50 11 0.00% 110.30% 55.04
wide homo 10 9 316.00% 550.00% 1530.82 50 9 0.00% 56.30% 39.45
hetero 20 19 258.10% 470.40% 1245.04 50 19 0.00% 59.30% 43.39
total/average 50 45 283.10% 614.60% 1471.61 200 45 0.00% 86.50% 47.95
M

Table 4: Solution performance for the large dataset

ED

For the large dataset (see Table 4), a substantial percentage of instances could not be solved to
optimality. Therefore, we compare solution values to the best solution available per instance
(gap best) and to a lower bound (gap LB). Please note that we determined lower bounds
PT

by combining an LP-relaxation and Gurobi’s lower bound after 1800 seconds of computation
time. For all 200 instances, Gurobi’s bound was tighter and therefore employed for evaluation.
However, the bound does not seem to be tight enough for a qualified evaluation of the solution
CE

procedures. For the large dataset the gap to the bound amounts to nearly 90% for our heuristic
procedure and even when applied to the small dataset the gap to the best found solution, of
which 80% are proven optima, is about 11%.
AC

When comparing to the best solution found, our heuristic is successful for all instances. Gurobi
obtains a solution of the same quality in 50 out of 200 cases (25%). For 45 of these 50 instances
Gurobi proves optimality. However, Gurobi’s solution values for the remaining instances are
far behind those found by our heuristic, so that the average gap of Gurobi’s solutions amounts
to almost 290%. Moreover, Gurobi’s average CPU time is 1470 seconds (time limit 1800
seconds) while the heuristic’s mean CPU time amounts to merely 48 seconds.
It can be concluded that our heuristic seems well suited to solve even larger instances of TRBD
in reasonable time at a good quality.

17
 ACCEPTED MANUSCRIPT

symbol parameter value

w side length of main square 4 km (2.5 miles)
|R| number of depots 1, 2, 4, 12, 24
|D| number of drop-off points 6, 12, 24, 36
|C| number of customers 25
[ρmin ; ρmax ] deadline factor interval [4; 10]
[wmin ; wmax ] weights interval [1; 1]
K trucks robot capacity 3, 5, 8, 10
µr handling time per robot [sec] { 20, 40, 60 }
µt handling time per truck delivery [sec] { 60, 120, 180 }
speed truck [km/h(mph)] 30(18.6)
speed robot [km/h(mph)] 3.5(2.2), 5(3.1), 6(3.7),
10(6.2), 15(9.3), 30(18.6)

T
Table 5: Parameter values used in managerial study

IP
6 Managerial aspects

CR
Beyond the pure computational performance, this section is dedicated to managerial aspects.
Specifically, we report on the results of a sensitivity analysis in Section 6.1. Here, we explore

US
how the network density of robot depots and drop-off points as well as the velocity of the
robots and the truck’s capacity for robots impact the timeliness of deliveries. Furthermore, in
Section 6.2 we benchmark our delivery policy where the robots return to decentralized robot
AN
depots with two alternative policies. One alternative is that the truck waits for the return
of the robots, so that they are taken on board again and no decentralized robot depots are
required. Our other competitor is a traditional delivery by truck without the assistance of
autonomous robots. In this way, some decision support on the right delivery policy is provided.
M

The computational tests of this section are based on a basic dataset of 10 instances generated
using the procedure described in Section 5.1. The parameters employed are listed in Table
ED

5. Note that default values are given in bold. If not explicitly specified otherwise, these
default values are applied. Whenever parameters are changed, the standard setting of a basic
instance remains identical and only properties influenced by the new parameter setting are
PT

recomputed, such that 10 instances derived from the basic dataset are solved for each setting
investigated. Consequently, the whole dataset solved in Section 6.1 contains a total of 200
instances, where 190 are obtained from the basic dataset. The study in Section 6.2 involves
CE

another 90 instances, which are each solved twice by differing MIP models, e.g., TBRD-MIP
and R2T-MIP or MinTruck-MIP, respectively.
Parameters µr and µt representing fixed handling times per robot and truck delivery, respec-
AC

tively, are included in the instance data as described in Section 3. These fixed times required
for loading a robot with a shipment and handing over a shipment by a deliveryman are added
to the travel time matrices. Further note that we assume the waiting and handling times at
each customer required to unload a robot being equal to the fixed robot loading time µr when
applying the policy where the robots have to return to the truck. Thus, twice µr adds to the
driving time of the robots after their launch from the truck under this policy.
All instances are solved employing Gurobi with a time-limit of one hour per instance. Please
note that, with only a few randomly scattered exceptions, all instances could be solved to
optimality.

18
 ACCEPTED MANUSCRIPT

6.1 Impact of network density, robot speed, and truck capacity

The performance of the truck-based robot delivery concept heavily depends on the following
four factors, whose impact is explored in a sensitivity analysis:

(a) The truck has to visit a decentralized robot depot from time to time in order to take
another batch of robots on board. Clearly, the denser the network of robot depots, the
shorter the detours of the truck to visit them. To explore the impact of the depot density
on the delivery performance, we place different number |R| of depots in our grid in the
equidistant manner defined in Section 5.1.

T
(b) Loading the robots with customer shipments may take some minutes, especially, if a

IP
complete batch of K robots is to be handled. Thus, stopping the truck, loading the robots
and launching them safely may not be possible just anywhere, so that regulatory authorities
may restrict robot handling to specifically designated parking spaces. Consequently, it

CR
seems interesting to also explore the impact of the network density of potential drop-off
points |D|.

US
(c) Due to safety reasons robots driving along sideways are bound to pedestrian speed. How-
ever, there may still be some flexibility whether robots are allowed to rather travel at 4.5
or at 6 km/h (2.8 to 3.7 mph). Therefore, we also explore the impact of the robot speed
on the delivery performance.
AN
(d) The limited capacity of the truck has to be partitioned among the space reserved for
shipments and that for robots. Thus, we also explore the impact of the truck’s maximum
loading capacity for robots K on the solution performance.
M

Delivery performance of our truck-based robot delivery concept is measured by the mean
ED

number of customers supplied late (averaged over all 10 instances per data-point) and the
results of our sensitivity analysis are depicted in Figure 4. The following conclusions can be
drawn from these results:
PT

(a) As expected, a denser network of robot depots increases delivery performance. However,
the good news is that marginal returns of additional depots diminish the more depots are
CE

added. Therefore, not too much investment costs for decentralized garages where robots
can be parked and recharged need to be spent.

(b) An analogous impact can be observed for the density of drop-off points. In our tests,
AC

more than 1.5 drop-off points per square kilometer (about 0.4 square mile) could not
further increase delivery performance. Within such a relatively large area it should be
easily possible to find secure launching points for robots in any city center.

(c) Also, the robot speed shows great influence on delivery performance. The gray interval
marks the range of pedestrian speeds, i.e., between 4.5 and 6 km/h (2.8 to 3.7 mph),
which is often attributed as walking speed. Our results reveal that choosing either the
lower or higher range of this interval is rather critical. For instance, if the speed is increased
from 5 km/h to 6 km/h in our tests, the number of delayed deliveries can be reduced by
75% from 0.4 to 0.1 on average. Due to the decreasing marginal utility, a further increase
of speed is less effective.

19
 ACCEPTED MANUSCRIPT

1 0.7

number of late deliveries

number of late deliveries

0.8
0.6

0.6
0.5
0.4

0.4
0.2

T
0 5 10 15 20 25 5 10 15 20 25 30 35
|R| |D|

IP
2 |R| = 1

number of late deliveries

3

CR
number of late deliveries

|R| = 12
1.5
2
1

0.5
US 1
AN
0

0 5 10 15 20 25 30 2 4 6 8 10 12
robot speed km
truck capacity K
M

Figure 4: Sensitivity analysis: Impact of network density of depots (top left) and drop-off
points (top right) as well as robot speed (bottom left) and truck capacity (bottom right)
ED

(d) The last impact factor investigated is the truck capacity K. It shows to be highly inter-
dependent especially with the depot density. Having a moderate depot density of 0.75
PT

depots per square kilometer (about 1.94 per square mile), i.e., 12 depots in total, the
solution quality shows astonishingly robust to changing truck capacity, because depots
are always close by and, thus, repeatedly visited anyway. Having a very low density of
CE

only 0.06 depots per square kilometer (about 0.16 per square mile), i.e., just a single
depot, truck capacity K becomes a critical factor and is able to reduce the number of
delayed customers by about 70%. However, our tests reveal that even in the latter case a
AC

decreasing marginal utility of additional capacity is observable. Consequently, a capacity

for K = 8 robots per truck, which is currently planned in the case of Mercedes-Benz
Vans and Starship Technologies, rather seems at the upper limit of reasonable capacities.
Consequently, it might be advisable to reduce truck capacity for robots in favor of a higher
capacity for shipments, especially if quite a few robot depots are available.

Our results show that, especially, weighing off the investment costs for a denser network of
robot depots and the security gains by slower robots need to be carefully traded off against
the delivery performance. A suitable truck capacity has to be chosen based on several other
setup factors, such as the depot density. In the next section, we explore whether or not it is
indeed a good idea to let the robots return to decentralized depots.

20
 ACCEPTED MANUSCRIPT

6.2 Benchmarking different operational policies

To explore the value of last-mile deliveries with truck-based autonomous robots, we benchmark
three alternative organizational settings for delivering goods to customers:

• First, we have our truck-based robot delivery policy where robots are launched from the
truck but return autonomously to decentralized robot depots. We dub this policy robot-
to-depot or R2D for short. Using this policy, truck stops remain pretty short because
robots only have to be loaded and launched by the driver. On an operational level, these
brief stops come for the price of additional detours towards robot depots. Additionally,
we have the investment and operational costs for the decentralized robot depots. To

T
evaluate the performance of the R2D policy, we solve TBRD-MIP with standard solver

IP
Gurobi and record the average number of delayed customers.

• To save these costs, the truck-based robot delivery concept can be altered. If the truck

CR
waits for the return of launched robots, they are taken on board again and can be
launched at the same or another truck stop without the need for robot depots. We
dub this policy robot-to-truck or R2T for short. On the negative side, truck stops

US
become much longer under this policy as the truck has to wait for the last robot to
return. Additionally, robot delays due to longer service times or unsuccessful waiting at
customer locations have to be considered. To determine the average number of delayed
AN
customers for the R2T policy, we solve the MIP model presented in Appendix A.

• Finally, we also benchmark the status quo which is a home-attended truck-based deliv-
ery without further assistance of autonomous robots. Here, each customer is supplied
M

directly by the driver of the truck, so that customer locations are directly accessed and
unnecessary stops at drop-off points or robot depots are avoided. Nonetheless, addi-
tional driving time to each single customer is required and each single truck stop takes
ED

comparatively long time. The deliveryman has to find a suited parking space, walk the
remaining distance, wait for the customer, and then return to the vehicle. To bench-
mark the traditional truck delivery with our R2D-policy, we determine the minimum
truck fleet that is required to also reach the service level of R2D. Specifically, we hand
PT

over the number Θ of delayed customers determined when solving TBRD-MIP for an
instance and determine the minimum fleet size of trucks, so that at most Θ customers
CE

are supplied late by solving the MIP model presented in Appendix B.

Clearly, the performance of each policy is heavily dependent on the service times, e.g., how
AC

long it takes a robot or the driver to hand over a shipment, so that we benchmark the three
delivery policies and quantify how their performance is impacted by different delivery times.
Specifically, we vary parameters µr and µt which define the delivery time of the robot and the
truck, respectively. Loading a robot at a drop-off point takes µr time units which – in addition
to the robot’s driving time between drop-off point and customer location – is the only delivery
time to be considered under the R2D policy. For the R2T policy, we have the time of the R2D
policy for loading the robot and driving to the customer and have to add the unloaded time
by the customer and additionally the driving time back to the truck. Unloading the robot by
a customer is assumed to take another µr time units. For the traditional truck-based delivery,
we presuppose the complete delivery process, i.e., parking, walking to the customer location,
handing over the shipment, and returning to the truck, to take µt time units. The results

21
 ACCEPTED MANUSCRIPT

of the three delivery policies and their dependency of different delivery times are depicted in
Figure 5. Each data-point represents the mean value of the respective objective function,
averaged over all 10 instances of the represented scenario. From these results, the following
conclusions can be drawn:
6
15
number of late deliveries

µr = 20
5.5 µr = 60

number of trucks
10 R2T
R2D 5

T
5 4.5

IP
4
0

CR
20 30 40 50 60 60 80 100 120 140 160 180
µr sec µt sec

Figure 5: Comparison of policies for varying delivery times for robots (µr ) and truck (µt )

US
• First, it can be concluded that our R2D policy where the robots return to depots con-
siderably outperforms the R2T policy where the truck has to wait for the robots’ return.
AN
The latter policy leads to significantly more dissatisfied customers whose shipments are
supplied late. The huge gap between both policies as depicted on the left of Figure 5 re-
mains rather unaffected by larger delivery times µr . This is not that astounding because
M

both policies equally suffer if loading (and unloading) of robots takes more time.
• Moreover, in our setting, also, the traditional truck-based delivery is clearly outperformed
by the R2D policy (see Figure 5, right). To reach the same service level as R2D, which
ED

means that at most the same number of customers are supplied late, at least three
additional trucks are required if no autonomous robots are deployed. Even more trucks
are required if the loading time µr of the robots is short and the delivery time µt of the
PT

deliveryman is large.

These results are a bit biased in favor of our R2D because we assume that robots are no
CE

bottleneck and the truck can always take the desired number of robots on board in each robot
depot. To realize this assumption in the real world, this would require additional investment
into a sufficiently sized robot fleet. If – to avoid the investment costs – our premise is not given
AC

in reality and robots may become scarce, then the advantage of R2D decreases. However, the
large advantage of R2D in our computational study compared to R2T and traditional truck-
based deliveries is, nonetheless, a strong indicator that R2D is a promising delivery policy when
applying a truck-based robot delivery.

7 Conclusion
This paper investigates an innovative last-mile concept where autonomous robots are launched
from trucks to deliver shipments towards customers. After delivery, the robots return to de-
centralized robot depots where a delivery truck can take them on board again. This paper

22
 ACCEPTED MANUSCRIPT

focuses on the scheduling of the delivery truck and the launching of robots along the truck
route, such that customers are timely delivered. We formulate the resulting scheduling prob-
lem, prove computational complexity, and derive an efficient heuristic solution procedure. Our
computational study shows that the decentralized robot depots greatly contribute to an effi-
cient delivery process. If the truck has to wait for the return of the robots instead, this leads
to considerable waiting times for the truck and much more unsatisfied customers. We also
benchmark our novel delivery concept with traditional attended-home delivery by truck. Our
computational study reveals that the truck fleet can considerably be reduced if autonomous
robots support the delivery process.
There are quite a few opportunities to extend our first findings on the truck-based robot
delivery concept. First, our study only considers a single truck where the set of customers

T
to be serviced is already given. On a preceding planning level, however, all current customer

IP
orders have to be partitioned among multiple vehicles. Deriving a larger solution framework,
where our solution procedure for the single truck case could be a valid building block, would be

CR
a challenging task. Furthermore, our (simplifying) assumption that robots are no bottleneck
resource and the truck can always load the demanded number of them on board in each robot
depot could be reinvestigated. If robots are scarce, each individual robot and its tour towards
customers and back to depots has to be tracked in detail, which adds a lot of complexity

US
to the problem. Furthermore, recent developments in smart-lock technologies (e.g., Amazon
Key), make unattended home delivery with autonomous ground vehicles a tangible possibility.
The implications of unattended home deliveries on the problem and our heuristics are sketched
AN
in Appendix C, but require further research effort. Decision support in these directions could
help to understand the economics of the truck-based robot delivery concept and whether it is
indeed a promising concept for future last-mile deliveries.
M

References
ED

[1] Afrati, F.; Cosmadakis, S.; Papadimitriou, C.H.; Papageorgiou, G.; Papakostantinou, N.
(1986): The complexity of the travelling repairman problem. Informatique théorique et
Applications 20, 79-87.
PT

[2] Agatz, N.; Bouman, P.; Schmidt, M. (2016): Optimization approaches for the traveling
salesman problem with drone. Working Paper Rotterdam School of Management.
CE

[3] Arkin, E.M.; Hassin, R. (1994): Approximation algorithms for the geometric covering
salesman problem. Discrete Applied Mathematics 55, 197-218.
AC

[4] Arslan, A.; Agatz, N.; Kroon, L.G.; Zuidwijk, R.A. (2016): Crowdsourced delivery: A
dynamic pickup and delivery problem with ad-hoc drivers. Working Paper Rotterdam School
of Management.

[5] Boysen, N.; Briskorn, D.; Fedtke, S.; Schwerdfeger, S. (2017): Drone delivery from trucks:
Drone scheduling for given truck routes. Working Paper Friedrich-Schiller-University Jena.

[6] Brenner, U. (2008): A faster polynomial algorithm for the unbalanced Hitchcock trans-
portation problem. Operations Research Letters 36, 408-413.

[7] Current, J.R.; Schilling, D.A. (1989): The covering salesman problem. Transportation
Science 23, 208-213.

23
 ACCEPTED MANUSCRIPT

[8] Daimler (2017): Vans & Robots Paketbote 2.0. https://www.daimler.com/innovation/

specials/future-transportation-vans/paketbote-2-0.html.

[9] Drexl, M. (2012): Synchronization in vehicle routing - A survey of VRPs with multiple
synchronization constraints. Transportation Science 46, 297-316.

[10] Garey, M.R.; Johnson, D.S. (1979): Computers and intractability: A guide to the theory
of NP-completeness. New York, Freeman.

[11] Gouveia, L.; Voss, S. (1995): A classification of formulations for the (time-dependent)
traveling salesman problem. European Journal of Operational Research 83, 69-82.

T
[12] Heilporn, G.; Cordeau, J.F.; Laporte, G. (2010): The delivery man problem with time
windows. Discrete Optimization 7, 269-282.

IP
[13] Kleinschmidt, P.; Schannath, H. (1995): A strongly polynomial algorithm for the trans-

CR
portation problem. Mathematical Programming 68, 1-13.

[14] Lin, S.W.; Vincent, F.Y.; Chou, S.Y. (2009): Solving the truck and trailer routing problem
based on a simulated annealing heuristic. Computers & Operations Research 36, 1683-1692.

US
[15] Miller, H.J. (2013): Beyond sharing: Cultivating cooperative transportation systems
through geographic information science. Journal of Transport Geography 31, 296-308.
AN
[16] Murray, C.C.; Chu, A.G. (2015): The flying sidekick traveling salesman problem: Opti-
mization of drone-assisted parcel delivery. Transportation Research Part C 54, 86-109.

[17] Pelletier, S.; Jabali, O.; Laporte, G. (2016): 50th anniversary invited article – Goods
M

distribution with electric vehicles: Review and research perspectives. Transportation Science
50, 3-22.
ED

[18] Reyes, D.; Savelsbergh, M.W.P.; Toriello, A. (2017): Vehicle routing with roaming deliv-
ery locations. Transportation Research Part C 80, 71-91.

[19] Savelsbergh, M.W.P.; Van Woensel, T. (2016): 50th anniversary invited article – City
PT

Logistics: Challenges and opportunities. Transportation Science 50, 579-590.

[20] Silva, M.M.; Subramanian, A.; Vidal, T.; Ochi, L.S. (2012): A simple and effective
CE

metaheuristic for the minimum latency problem. European Journal of Operational Research
221, 513-520.

[21] Speranza, M.G. (2018): Trends in transportation and logistics. European Journal of
AC

Operational Research 264, 830-836.

[22] Wang, X.; Poikonen, S.; Golden, B. (2017): The vehicle routing problem with drones:
several worst-case results. Optimization Letters 11, 679-697.

24
 ACCEPTED MANUSCRIPT

Appendix A: A MIP model if the truck waits for the

robots’ return
To compare the usage of decentralized depots, i.e., the R2D policy, with the concept of the
truck waiting for the return of launched robots, i.e., the R2T policy, we introduce a MIP
model, which minimizes the weighted number of late deliveries under the R2T policy. Within
this setting, there is no need for robot depots, so that we do not differentiate between depots
and drop-off points, i.e., we set D̄ = D ∪ R. The new setting allows us to relinquish all
l variables of model TBRD-MIP provided in Section 4.1. Instead, we introduce variable Ck
representing the delivery time at customer location k. Furthermore, we have to keep track

T
which robot serves a customer, so that we have to adapt our x variables (see Table 6). Our
modified model R2T-MIP consists of objective function (17) and constraints (18) to (31).

IP
R̄ set of robots (i.e. R̄ = {1, . . . , K})

CR
Ck continuous variables: delivery time at location k
xj,k,r binary variables: 1, if customer k is supplied from location j by robot r; 0,
otherwise
yk1 ,k2
US
binary variables: 1, if customer k1 is supplied before customer k2 from the
same location by the same robot; 0, otherwise

Table 6: Additional and altered notation for R2T-MIP

AN

X
R2T-MIP: Minimize F (S, T, X, Y, Z) = (17)
M

zk · wk
k∈C
ED
PT
CE
AC

25
 ACCEPTED MANUSCRIPT

X X
xj,k,r = 1 ∀k ∈ C (18)
r∈R̄ j∈D̄∪{γ}
X X
xj,k,r · ϑrj,k ≥ xj,k,r−1 · ϑrj,k ∀ j ∈ D̄ ∪ {γ} ; r ∈ R̄ \ {1} (19)
k∈C k∈C
X
Cj ≥ 2 · xj,k,K · ϑrj,k ∀ j ∈ D̄ ∪ {γ} (20)
k∈C
tγ = 0 (21)
tj ≥ ti − M · (1 − si,j ) + ϑti,j + Ci ∀ i ∈ D̄ ∪ {γ} ; j ∈ D̄ \ {i} (22)
xj,k1 ,r + xj,k2 ,r ≤ 1 + yk1 ,k2 + yk2 ,k1 ∀ j ∈ D̄ ∪ {γ} ; k1 < k2 ∈ C; r ∈ R̄
(23)

T
X
M · zk ≥ tj − M · (1 − xj,k,r ) − dk

IP
r∈R̄
X
+2· ϑrj,k1 · yk1 ,k + ϑrj,k ∀ j ∈ D̄ ∪ {γ} ; k ∈ C (24)

CR
k1 ∈C\{k}
X
sγ,j ≤ 1 (25)
j∈D̄∪{γ e }
X

i∈D̄∪{γ e }\{j}
XX
sj,i =
X

i∈D̄∪{γ}\{j}
X
si,j US ∀ j ∈ D̄ (26)
AN
xj,k,r ≤ M · si,j ∀ j ∈ D̄ (27)
r∈R̄ k∈C i∈D̄∪{γ}\{j}
XX X
xj,k,r ≥ si,j ∀ j ∈ D̄ (28)
M

r∈R̄ k∈C i∈D̄∪{γ}\{j}

si,j ∈ {0, 1} ∀ i ∈ D̄ ∪ {γ} ; j ∈ D̄ ∪ {γ e } \ {i}

(29)
ED

xi,k,r , zk ∈ {0, 1} ∀ i ∈ D̄ ∪ {γ} ; k ∈ C; r ∈ R̄ (30)

yk1 ,k2 ∈ {0, 1} ∀ k1 ∈ C; k2 ∈ C \ {k1 } (31)
PT

Objective function (17) as well as constraints (18), (21) to (28) and (29) to (30) directly
correspond to objective function (1) as well as constraints (2), (6) to (12) and (14) to (15) of
CE

TBRD-MIP described in Section 4.1. Furthermore, constraints (19) and (20) determine the
delivery time at each location. Note that (19) eliminates significant symmetry from the model.
According to the deadline of each customer it might become necessary to visit a location
l m more
than once. To comply with that requirement we can duplicate each location |C|
AC

2
times.
At worst, the truck has to drive back and forth between two drop-off points and, each time,
merely launches a single robot.

Appendix B: A MIP model for traditional truck-based

delivery
To benchmark our R2D policy, we compare the robot-based last-mile delivery with the tradi-
tional usage of trucks where no autonomous robot is available. To benchmark these policies

26
 ACCEPTED MANUSCRIPT

we introduce the following MIP model, which determines the minimum number of trucks
needed under the classical policy to obtain a solution with the same service level as R2D, i.e.,
at most the same number of late customers deliveries are allowed. The number of allowed
deadline violations Θ is given by the objective value of TBRD-MIP presented in Section 4.1.
The resulting MIP model consists of objective function (32) and constraints (33) to (40). The
additional notation is summarized in Table 7.

V set of vehicles (|V | = |C|)

Θ maximum number of deadline violations allowed
tk continuous variables: arrival time at customer k
xk1 ,k2 ,v binary variables: 1, if customer k1 visited direct before customer k2 by truck

T
v; 0, otherwise
zk binary variables: 1, if customer k is supplied late; 0, otherwise

IP
Table 7: Additional notation for the MIP modelling the traditional truck-based delivery

CR
XX
MinTruck-MIP: Minimize F (T, X, Z) = xγ,k,v (32)

X X
US k∈C v∈V
AN
xk1 ,k2 ,v = 1 ∀ k2 ∈ C (33)
k1 ∈C∪{γ} v∈V
X
xγ,k,v = 1 ∀v ∈ V (34)
M

k∈C∪{γ}
X X
xk1 ,k2 ,v = xk2 ,k1 ,v ∀ k2 ∈ C; v ∈ V (35)
k1 ∈C∪{γ} k1 ∈C∪{γ}
ED

tk1 + ϑtk1 ,k2 − M · (1 − xk1 ,k2 ,v ) ≤ tk2 ∀ k1 , k2 ∈ C; v ∈ V (36)

ϑtγ,k − M · (1 − xγ,k,v ) ≤ tk ∀ k ∈ C; v ∈ V (37)
tk − M · zk ≤ dk ∀k ∈ C (38)
PT

X
zk ≤ Θ (39)
k∈C
CE

xk1 ,k2 ,v , zk1 ∈ {0, 1} ∀ k1 , k2 ∈ C ∪ {γ}; v ∈ V (40)

The objective function (32) minimizes the number of utilized trucks. Constraints (33) to (35)
AC

make sure that each customer is delivered by a vehicle and a well-defined sequence of visits is
determined. The arrival times at the customers are set in constraints (36) and (37). Delays
are detected by constraints (38) and the number of delays is limited by constraints (39). The
domains of the binary variables are finally set by constraints (40).

Appendix C: Unattended home delivery

In the current planning stage, the delivery robots are not able to drop off loaded freight
without human assistance, so that their application is restricted to the attended home delivery

27
 ACCEPTED MANUSCRIPT

context. Recent developments in smart-lock technologies (e.g., Amazon Key), however, make
unattended robot delivery a possible option for the future. This would be advantageous for
both, customers, who can order goods without worrying about suited delivery times, and
couriers, who are relieved from tight delivery time windows. Once unattended home delivery
by robots is realized, there is no need to agree delivery time windows with customers. For
our problem setting, this means that delivery deadlines are superfluous and other objectives,
e.g., traditional vehicle routing objectives, need to be incorporated into our MIP model and
the heuristic.
First, we explore one of the most popular vehicle routing objectives of minimizing the total
travel distance. Adapting the introduced model TBRD-MIP necessitates to remove constraints
(8) and to substitute the objective function (1) by

T
X X X X

IP
Minimize F (L, S, T, X) = xj,k · ϑrj,k + α · si,j · ϑti,j ,
j∈D∪R∪{γ} k∈C i∈D∪R∪{γ} j∈D∪R∪{γ e }\{i}

CR
where α weights the travel distance of the truck compared to the robots.
To adapt our heuristic procedure to this problem setting, the cost matrix of the transportation
problem has to be determined in the following way:

US
• The costs from each supply node to the dummy demand node d are set to zero.
• The costs ci,j between nodes referring to a depot ri ∈ R(v) at stop vr(i) and a customer
AN
j ∈ C are set to
ci,j = ϑrvr(i) ,j .

• Furthermore, costs ck,j between nodes referring to a subtour Sk ∈ S(v) and a customer
M

j ∈ C are set to the minimum travel distance of the robots, such that
 r
ck,j = min ϑvi ,j .
ED

i=k1 ,...,kik

Finally, the weighted truck route length has to be added to the value determined by the
transportation problem. Note that we do not integrate the distances between customers
PT

and robot depots into our calculations. They are constant, since robots have to return to
depots irrespective of the current assignment decision. In consequence, distances traveled
after customer delivery can be neglected.
CE

Another prominent vehicle routing objectives is the minimization of the trip duration. We can
deal with this objective by introducing a further decision variable z representing the completion
time of the tour. For our TBRD-MIP we, then, have to replace objective function (1) by
AC

Minimize F (L, S, T, X, Z) =z
and constraints (8) by
z ≥ tj − M · (1 − xj,k ) + ϑrj,k ∀ j ∈ D ∪ R ∪ {γ} ; k ∈ C.

For a given truck route, the minimal duration can be determined by binary search on the
completion time z solving the introduced Hitchcock transportation problem. In this way, our
heuristic solution procedure can be adapted to solve the considered problem by solving the
binary search each time a truck route is evaluated. To assess if a given trip duration z can be
kept for a predefined tour, the cost matrix has to be computed as follows:

28
 ACCEPTED MANUSCRIPT

• The costs from each supply node to the dummy demand node are set to zero.

• The costs ci,j between nodes referring to a depot ri ∈ R(v) at stop vr(i) and a customer
j ∈ C are: ( Pr(i)−1 t
0 , if l=1 ϑvl ,vl+1 + ϑrvr(i) ,j ≤ z
ci,j =
1 , else.

• Additionally, costs ck,j between nodes referring to a subtour Sk ∈ S(v) and a customer
j ∈ C are:
( nP o
i−1 t
0 , if mini=k1 ,...,kik ϑ
l=1 vl ,vl+1 + ϑr
vi ,j ≤z
ck,j =

T
1 , else.

IP
If the given trip duration z is satisfiable, i.e., each customer is timely supplied within trip
duration z, the optimal objective value of the transportation problem amounts to 0. Otherwise,

CR
it is greater than zero. Based on this, the binary search can be continued to achieve the minimal
trip duration for a given truck tour.

US
Note that in this configuration, the trip duration is measured based on the delivery time at the
customer. This approach is valid whenever a fast delivery of customers is of high importance
and robots are not a scarce resource. Whenever robots are scarce, however, the return time
to the nearest depot should be modeled, too, to avoid shortages. Such an objective can easily
AN
be implemented by computing the return time to the next depot ϑ̄depot k = minj∈R {ϑrk,j } for
each customer k ∈ C and considering this value in the formulas above.
M
ED
PT
CE
AC

29