Autonomous Exploration and Mapping using a Mobile Robot

Running ROS

Rachael N. Darmanin and Marvin Bugeja

Department of Systems and Control Engineering, University of Malta, Msida, Malta

Keywords:

Wheeled Mobile Robots, ROS, Autonomous Exploration, SLAM.

Abstract:

This paper proposes a practical solution to the autonomous exploration and mapping problem using a single

mobile robot. Moreover, the authors implement the proposed scheme within the Robot Operating System

(ROS), and validate it experimentally using PowerBot, a real wheeled mobile robot equipped with a 2D laser

scanner. In essence, the proposed scheme integrates an efﬁcient particle-ﬁlter-based SLAM algorithm, two

different exploration strategies, and a path-planning and navigation module. The modular nature of the pro-

posed scheme is intentional and advantageous. It allows the authors to compare the two exploration strategies

under investigation objectively and with ease. Finally, hypotheses testing is also used to strengthen the results

of the comparative analysis.

1 INTRODUCTION

Exploration and mapping by an autonomous agent

have been of particular interest to the robotics com-

munity for decades. Research shows that the au-

tonomous mapping of an unknown environment by a

mobile robot is a non-trivial task (Thrun et al., 2005).

Global Positioning Systems (GPS) and the use of bea-

cons for mappingan unknownenvironmentmay facil-

itate the task of mapping (Dissanayake et al., 2011).

However, such systems suffer from a number of in-

herent limitations, namely, the inaccessibility of GPS

indoors and the laborious setting up of strategically

placed beacons. Hence, the concept of Simultaneous

Localization and Mapping (SLAM), was developed in

order to amalgamate sensory data into a world model

(or a map). Through probabilistic techniques, SLAM

algorithms are capable of processing sensory data to

estimate the location of a robotic system and a map of

its environment at the same time.

In robotics, maps are mainly acquired to enable

autonomous navigation. However, mapping using

robots can also be useful when it comes to explore

and map dangerous environments, or places that are

inaccessible to human beings. Hence, autonomous

exploration strategies were developed, which enable

the robot to determine the next location to explore

in order to improve its map and localization esti-

mates. Exploration strategies place an emphasis on

autonomously mapping as much of an unknown en-

vironment as possible, and in the shortest time pos-

sible. Other researchers argue that rather than just

exploratory actions, accurate mapping requires ex-

ploitation actions, such as place revisiting actions

(Makarenko et al., 2002). This gave rise to Active

SLAM strategies, that seek to improve the localiza-

tion estimate of the robot rather than explore as much

of the environment as possible in the shortest time.

However, both exploration and Active SLAM strate-

gies ultimately provide the robot with a sequence of

locations that it needs to visit in order to meet the

speciﬁed exploration criteria.

For a robot to autonomously explore and map an

environment, it must follow a set of navigation in-

structions that allow it to proceed from one location

to the next safely, even in the presence of obstacles.

During the years, several path planning and motion

control algorithms have been proposed which enable

the robot to move to the chosen location. The integra-

tion of a SLAM algorithm, an exploration or Active

SLAM strategy and a path planning/navigation algo-

rithm allows the robot to autonomously explore and

map an unknownenvironmentusing just sensory data.

The purpose of this work was to develop a sin-

gle robot system, running Robot Operating System

(ROS), that can autonomously explore and map un-

known environments efﬁciently. The system needed

to be modular, in that the SLAM, path planning and

motion control modules, can be used with differ-

ent exploration strategies, without any modiﬁcations.

208

Darmanin, R. and Bugeja, M.

Autonomous Exploration and Mapping using a Mobile Robot Running ROS.

DOI: 10.5220/0005962102080215

In Proceedings of the 13th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2016) - Volume 2, pages 208-215

ISBN: 978-989-758-198-4

ROS facilitates this modular implementation since it

allows a set of independent processes to communicate

through a message passing structure.

The main contributions of this work are: 1) the de-

velopment and experimental evaluation of a new ROS

package consisting of two autonomous exploration

strategies for mobile robots. To the best of the au-

thors’ knowledge, no other ROS package offering the

same functionality is available to date; 2) a compara-

tive study of the two mentioned exploration strategies,

based on physical experiments in a real-life environ-

ment.

The rest of this paper is structured as fol-

lows. Section 2 provides a review of the exist-

ing autonomous exploration and mapping systems.

Section 3 describes the main hardware used in this

implementation and the underlying software frame-

work that makes up the whole robotic system.

Section 4, presents a set of experimental results while

Section 5 contains a comparative analysis, backed up

by statistical hypothesis testing, of the two explo-

ration strategies under investigation. Finally, conclu-

sions from this work are drawn in Section 6.

2 RELATED WORK

Being aware of one’s environment means that: one

is able to perceive the environment, localize himself

in it, and be able to navigate effectively in this envi-

ronment. In the ﬁeld of mobile robotics, this involves

learning maps of an unknown environment, which is

the integration of three non-trivial tasks that must be

solved concurrently (Stachniss et al., 2005). Figure 1

illustrates this concept as the integration of mapping,

localization and path planning and motion control.

Figure 1: Illustration of the integrated approaches, adapted

from (Makarenko et al., 2002).

Given a known map, several path planning algo-

rithms such as Dijkstra’s algorithm (Dijkstra, 1959),

A* algorithm (Dechter and Pearl, 1985), D* algo-

rithm (Stentz, 1994) and Dynamic Window Approach

(DWA) (Fox et al., 1997), among others, have been

proposed to allow a robot to navigate autonomously

in the environment. During the years, there have

been many advancements in the ﬁeld of path planning

algorithms that can be used seamlessly within the

integrated problem of autonomous exploration and

mapping. Moreover, the Dijkstra, A* and D* algo-

rithms are all global path planners, in that they plan a

path from the current robot location to the destination

based on the current global map (the whole map ac-

quired so far). The DWA was designed to make use of

the velocity control space to search for control com-

mands that must be sent to the robot in order to reach

a goal destination. Hence, the DWA does not make

use of a global map, but rather it uses the local map

obtained through sensory data.

Perhaps the most complex problem that must be

solved in the task illustrated in Figure 1, is SLAM. In

autonomous exploration, the map estimate produced

by SLAM is used by the exploration algorithms in

order to identify which is the best location that the

robot must visit next. Overall, SLAM algorithms

may be organized into three main categories as fol-

lows: Extended Kalman Filter (EKF) SLAM, Particle

Filter SLAM, and Graph Based SLAM. EKF SLAM

was for many years the preferred choice since it re-

quires a simple and straight-forward implementation

and performs well in small areas having a limited

number of landmarks. However, since EKF SLAM

bases its estimate on just one hypothesis and assumes

that the posterior is normally distributed, it may be-

come overconﬁdent in its estimate, leading to less ac-

curate mapping (Dissanayake et al., 2011). Hence,

Particle Filtering became the preferred probabilis-

tic technique for many researchers. Particularly, in

(Thrun et al., 2004), the authors integrate the common

Sequential Importance Resampling, (SIR) ﬁlter with

EKF to produce FastSLAM, which results in a faster

algorithm, when compared to the common SIR ﬁlter

on its own. Furthermore, Rao-Blackwellized Particle

Filters are more commonly used to solve the SLAM

problem since they attempt to reduce the degeneracy

and particle depletion problems presented by less ad-

vanced particle ﬁlters such as the Sequential Impor-

tance Sampling (SIS) and SIR ﬁlters (Grisetti et al.,

2007). In (Grisetti et al., 2007) the authors propose

two improved techniques for Rao-Blackwellized Par-

ticle Filter SLAM, that make the algorithm more efﬁ-

cient and accurate in estimating the most likely pose

as well as having a reduction in the particle depletion

problem. The implementation of this improved Rao-

Blackwellized Particle Filter is often referred to as

GMapping, which is directly accessible online from

openslam.org

. Moreover, the use of Graph-Based

Autonomous Exploration and Mapping using a Mobile Robot Running ROS

209

techniques in SLAM is fast gaining popularity. In

Graph-Based SLAM, each node in the graph repre-

sents an estimated robot pose or the estimated posi-

tion of a landmark, while the edges connecting these

nodes encode in them sensor observations that are

constrained by the estimated robot pose. The perfor-

mance of Graph-Based SLAM is said to be compara-

ble to that of Particle Filter-Based SLAM algorithms

such as GMapping (Burgard et al., 2009).

Although the task of choosing the next loca-

tion that the robot should visit seems to be intu-

itive, several researchers have developed several dif-

ferent exploration and Active SLAM algorithms for

this purpose. One very popular and simple explo-

ration strategy is the Nearest Frontier approach (Ya-

mauchi, 1997). This algorithm simply analyses an

Occupancy Grid Map and detects all potential borders

(called candidate frontiers) between the cells marked

as OPEN (free from obstacles) and those marked as

UNKNOWN (not yet explored). Such a map is seg-

mented into cells, where each cell is assigned with a

probability of occupancy,with zero probability mean-

ing that it is free from obstacles, and with a probabil-

ity of one, meaning that it is OCCUPIED. Normally,

a probability of 0.5 means that there is no informa-

tion about the occupancy of the cell, and hence it is

still UNKNOWN (Yamauchi, 1997). The distance be-

tween the current robot pose and each frontier, is anal-

ysed, and the frontier closest to the robot is chosen as

the next location to explore.

Another algorithm that is designed to explore as

much of the environment as possible is the Next Best

View (NBV) approach proposed in (Gonzalez-Ba˜nos

and Latombe, 2002). In this case, the evaluation cri-

teria of the candidate destinations, attempt to strike a

balance between the amount of unknown area that the

robot can explore from that location, and the distance

that the robot must travel to arrive there.

Unlike the Nearest Frontier and the NBV strate-

gies just described, in (Makarenko et al., 2002) and

(Stachniss et al., 2005), among others, the authors

propose techniques that do not only focus on explor-

ing as much of the environment as possible, but in

addition aim to improve the localization estimate of

the robot (Active SLAM). Particularly, in (Stachniss

et al., 2005), the authors propose a technique that

makes use of the entropy of a Rao-Blackwellized Par-

ticle Filter together with the distance from the current

robot pose in order to evaluate candidate destinations.

To date, the robotics community is still in de-

bate on which of these exploration strategies is best

to solve the problem of autonomous exploration and

mapping. To this end, several works such as those

presented in (Amigoni, 2008; Carlone et al., 2014),

among others, seek to compare such strategies. In

(Amigoni, 2008), the author shows that the simpler

algorithms, such as the NBV approach, yield the best

results in the shortest time possible, and the robot

travelling the shortest distance. However, in (Carlone

et al., 2014), the authors show that when the focus is

placed on the quality of the map, then those strate-

gies that involve Active SLAM produce the best re-

sults, even if not in the shortest time possible. Such

works indicate that the choice of exploration or Ac-

tive SLAM algorithm is highly dependent on the ap-

plication and on the criteria that shall be used in the

evaluation of the results.

3 SYSTEM DESIGN AND

IMPLEMENTATION

The concept of autonomous exploration and map-

ping represented in Figure 1 was implemented in

ROS on PowerBot

, a differentially-drivenwheeled

mobile robot designed and manufactured by Adept

Mobilerobots for research and rapid prototyping.

PowerBot is equipped with a SICK LMS200 laser

scanner, an Advanced Robotics Control and Opera-

tions (ARCOS) robot controller for low-level motion

control, as well as an on-board computerfor algorithm

development. The on-board computer is interfaced

with the robot controller via Advanced Robotics In-

terface for Applications (ARIA), which is a library of

functions that are capable of handling the lowest level

details of the client-server interactions, such as con-

trolling the speed of the robot.

Robot Operating System (ROS) is a software

framework that is ideal for the development of robotic

applications. It consists of a number of software pack-

ages that are used to perform particular tasks, for ex-

ample, SLAM. ROS facilitates modular implementa-

tion of different functions, in that processes that are

currently running on the system (called nodes), in-

teract between themselves through a message passing

system, as shown in Figure 2. Each packet of data is

sent between nodes according to the topic that it has

been assigned. The topic deﬁnes the type of data mes-

sage that is sent. Furthermore, a node may either be

a publishing node (broadcasting data) or a subscriber

node (receiving data). Hence any node subscribing

to the same topic shall receive the same data being

broadcast by the publisher. For instance, in Figure 2

the publishing node “talker” is communicating with

the subscriber node “listener” by sending data mes-

sages over the topic “chatter”. These interactions are

controlled by the master node called “roscore”.

The modularity of ROS leads to a rather simple

ICINCO 2016 - 13th International Conference on Informatics in Control, Automation and Robotics

210

Figure 2: ROS architecture of nodes and topics.

design that emerges from the integration of the fun-

damental concepts and the system block diagram de-

picted in Figure 3. Going from the block diagram in

Figure 3 to a system designed in ROS requires the

identiﬁcation of the speciﬁc package to be used in

place of each component in the block diagram. The

parameters of each package must then be tuned for

the speciﬁc robot platform and application.

Figure 3: ROS architecture of nodes and topics.

To perform SLAM, the gmapping package, which

implements the Rao-Blackwellized particle-ﬁlter-

based approach developed in (Grisetti et al., 2007),

was chosen due to its ability to map an environment

in an accurate and efﬁcient manner. In contrast to

EKF-based SLAM, this technique uses multiple hy-

potheses represented by the particle set in the parti-

cle ﬁlter. The node, called “slam

gmapping”, sub-

scribes to the raw odometry data and laser scan data

over the “/RosAria/pose” and “/scan” topics, respec-

tively. The package “ROSARIA” acts as a wrapper

for the ARIA libraries in ROS. It allows the transmis-

sion of high level commands to the robot controller

from ROS. In fact, the odometric pose estimate of the

robot is published by the “RosAria” node. In turn, the

“slam

gmapping”node publishes the map data, which

is then used by the exploration algorithm to compute

the next location that the robot needs to explore.

The published map is an Occupancy Grid Map

vector. Moreover, some of the “slam

gmapping” pa-

rameters had to be tuned in order to extract the best

performance in the given environment with the given

robot. For example: two particular parameters were

used to enable the SLAM package to execute a scan

every time the robot translates by 0.2 metres or rotates

by 0.1 radians. The settings were chosen in an attempt

to obtain the best map with the available computa-

tional power, as more frequent scans naturally require

more computational power. Furthermore, the laser

range was also tuned in accordance with the speci-

ﬁcations of the laser scanner, namely the maximum

range setting was set to ﬁve metres. Finally, the map

was set to be updated every three seconds.

Although there are some packages that implement

the Nearest Frontier exploration approach in ROS,

these presented certain limitations, such as not being

fully autonomous, and thus could not be used for the

intended purpose. Hence, the Nearest Frontier ap-

proach proposed in (Yamauchi, 1997) and the Next

Best View approach proposed in (Gonzalez-Ba˜nos and

Latombe, 2002) were both implemented from scratch

as ROS nodes in a new package called “exploration”.

To the best of the authors knowledge, this is the ﬁrst

time that these two algorithms are being made avail-

able in ROS.

For both exploration strategies, the potential can-

didates for exploration emerge through the use of

frontier detection which was also proposed in (Ya-

mauchi, 1997). Frontier detection is performed in two

steps:

1. The search for frontier cells, and

2. The grouping of said frontier cells into frontiers

comprising adjacent frontier cells only.

In the search of frontier cells, the algorithm traverses

the map data and creates an vector of frontier cells.

These frontier cells are then grouped into frontiers ac-

cording to their adjacency. Each frontier is then stored

in a two-dimensional vector which is used by the ex-

ploration algorithm for further analysis. Eventually,

the exploration process terminates when the frontier

detection algorithm can no longer detect frontiers that

are at least ten cells long. This applies to both explo-

ration approaches that were examined in this work.

The difference among the two approaches lies

in how they evaluate the candidate destinations and

choose the location that the robot must explore next.

In the Nearest Frontier approach, a global path is

planned from the current robot pose to each candidate.

The candidate that is closest to the robot (i.e. the one

having the shortest path) is then chosen. On the other

hand, the Next Best View approach implements the

following equation to rank the potential candidates,

g(q) = A(q)exp(−λL(q, q

)) (1)

where g(q) represents the ranking function, A(q) is

the total unknown area that is expected to be covered

by the laser scanner from a candidate location, and

L(q, q

) is the length of the path planned from the cur-

rent robot pose to each candidate location, using the

global occupancy grid map published by the SLAM

algorithm.

The path length, L(q, q

) is computed in the same

manner as it is computed in the Nearest Frontier ap-

proach. This means that the same global path planner

used to plan accessible paths for navigation is used to

Autonomous Exploration and Mapping using a Mobile Robot Running ROS

211

plan a path for the robot in an incomplete occupancy

grid map. The length of this path is then computed.

The unknown area that the robot can potentially “see”

from a candidate location, A(q), is calculated by a ray

casting technique, which is also used in (Gonzalez-

Ba˜nos and Latombe, 2002). Ray casting attempts to

simulate the data obtained through a single scan of

the laser scanner. Given a scan resolution, the max-

imum laser range and the scanning angle of the real

laser scanner, a number of rays are projected from a

candidate location. If the ray encounters a known cell

(OPEN or OCCUPIED), then the length of the ray

from that point to the maximum ray length is cropped

out. Consequently, the endpoint of the ray is the point

at which the ray hits a known cell, and therefore, the

length of that ray is the Euclidean distance between

the candidate location coordinatesand the coordinates

of the ray endpoint. Thus, the unknown area that

may be visible from a candidate point is the summa-

tion of the lengths of the rays projected outwards over

the unknown cells only, as seen in (2), where (x

, y

)

are the map coordinates of the candidate location and

end

, y

end

) are the coordinates of the ray endpoint.

A(q) =

∑

No. of rays

end

− x

)

+ (y

end

− y

)

(2)

Furthermore, the parameter λ plays an important role

in the selection of the next best view since it allows

the robot to either prefer shorter routes over a larger

area of visibility, or vice-versa. If the value of λ

is small, the algorithm prefers the candidates from

which a larger unknown area may be explored and

hence gives them a higher rank. If λ is large, the algo-

rithm gives a higher rank to those candidates that are

closer to the robot. Eventually the candidate that has

the largest ranking value, g(q), is chosen as the next

location that the robot must navigate to and explore.

As soon as the nextdestination is chosen, the robot

must navigate to that location. To do so, path plan-

ning and motion control algorithms are required to

allow the robot to navigate effectively in an environ-

ment cluttered with both static and dynamic obstacles.

For this purpose, the package “move

base” was used.

This package requires the pose of the robot within the

map to be able to plan a global path from the current

robot pose to the goal destination. In order to take the

dimensions of the robot into consideration, the path

is planned in a global costmap that inﬂates the cost

around each OCCUPIED cell. If the robot traverses

the inﬂated circle around such a cell, it may be either

close to colliding with an obstacle or in fact, it may be

in collision, depending on the cost of each cell. This

global costmap is constructed on the static occupancy

grid map that is being built by SLAM. Moreover, hav-

ing just a global path planned does not enable the

robot to avoid dynamic obstacles such as people walk-

ing around it. To mitigate this problem, the package

makes use of a local path planner, which plans inter-

mediate paths using a local map. This local map is

another costmap which is constructed using the laser

scan data directly. Therefore, if a dynamic obstacle

appears in the local path planned of the robot, this

would be detected by the laser scanner and the local

path planner changes course accordingly. However,

the local path planner always seeks to guide the robot

as close to the global path as possible. For this effect,

the global and local costmap parameters were tuned to

the dimensions of the robot. Furthermore, the global

path planner adopted in this case is the Dijkstra’s Al-

gorithm (Dijkstra, 1959), while the local path plan-

ner adopted is the Dynamic Window Approach (Fox

et al., 1997). The implementation of these path plan-

ners were both available in ROS.

4 RESULTS

One of the aims of this study was to compare the

mapping accuracy of two exploration strategies which

were implemented as part of a complete autonomous

exploration and mapping system in ROS. The ﬁrst

set of results presented in Section 4.1 shows how the

different exploration strategies choose different ex-

ploratory locations from a set of candidates. Further-

more, Section 4.2 presents a set of results obtained

through a number of real life autonomous exploration

and mapping experiments. The map obtained in each

experiment was compared to a ground truth map, il-

lustrated in Figure 4, which was obtained by manually

steering the robot in the environment. Hence, the Ac-

ceptance Index metric could be calculated. This com-

parativemetric was introduced in (Carpin, 2008). The

acceptance index, ω, is calculated as the ratio of the

agreement between the known cells of a ground truth

map, M1, and a resulting map, M2, to the total sum

of agreed and disagreed cells, as shown in (3). When

ω = 0 it means that the two maps are completely dis-

tinct and when ω = 1 it means that the two maps are

completely identical.

ω(M1, M2) =

agr(M1, M2)

agr(M1, M2) + dis(M1, M2)

(3)

4.1 Case Study

In this case study, an incomplete map of the Univer-

sity of Malta Faculty of Engineering ground ﬂoor was

ICINCO 2016 - 13th International Conference on Informatics in Control, Automation and Robotics

212

Figure 4: Ground truth map obtained by manually steering

the robot in the laboratory environment.

used, where three candidate locations were detected

by the frontier detection technique as shown in Figure

5. Candidate 3 shows a frontier that is made up of a

collection of connected smaller frontiers. This is due

to the diffused laser rays in that area which resulted in

several gaps of unknown cells between the free cells.

For this scenario, the ranking values of each approach

are tabulated in Table 1. In order to show how the

NBV approach can be used to rank candidates differ-

ently, the parameter λ was set to four different values

from the set {0.02, 0.15, 0.5, 1.0}.

Figure 5: Incomplete map of the University of Malta Fac-

ulty of Engineering ground ﬂoor showing three exploratory

candidates, together with the respective path length (in me-

tres) from the current robot pose and the expected unknown

area (in metres squared) that is visible from each candidate.

Table 1: Case Study results showing the value of the rank-

ing function for each candidate, {1, 2, 3}, where the ranking

function of the Nearest Frontier is the path length, which

must be minimized, and that of the NBV is g(q) as shown in

(1), which must be maximized. Note that the ranking value

of the chosen candidate for each approach is italicized.

Approach 1 2 3

Nearest Frontier 4.84 35.65 7.04

NBV λ = 0.02 23.70 40.55 144.1

NBV λ = 0.15 12.64 0.39 57.74

NBV λ = 0.5 2.33 1.49 ×10

−6

4.92

NBV λ = 1.0 0.21 2.67 ×10

−14

0.15

As can be seen in Table 1, the Nearest Frontier

approach chose Candidate 1 as the next exploratory

location since it offered the smallest path length. The

algorithm of interest is NBV. When the value of λ was

small (i.e. {0.02, 0.15, 0.5}), Candidate 3 was always

preferred since it offered the largest area of visibil-

ity even though it was not the closest location to the

robot. When λ was equal to 1.0, the algorithm pre-

ferred Candidate 1 since it is the location closest to the

robot. Hence, one can conclude that as the value of λ

increases, the NBV approach starts preferring candi-

dates that are closer to the current robot pose and thus,

this approach becomes similar to the Nearest Frontier

approach.

4.2 Completely Autonomous

Exploration and Mapping Results

In order to validate and compare the algorithms,

the whole system was used in a set of four ex-

periments where for each experiment, seven trials

were recorded. In each experiment, the robot au-

tonomously explored, navigated and mapped a con-

trolled environment (a laboratory) consisting of sev-

eral static obstacles such as desks and chairs. In

the ﬁrst experiment the Nearest Frontier approach

was used while in the other three experiments, the

NBV approach was used with different values of

λ = {0.15, 0.5, 1.0}. These values of λ were cho-

sen heuristically depending on the visual inspection

of the robot behaviour. Figure 6 illustrates a typical

map resulting from one trial for each of the adopted

approaches while Table 2 presents the average (across

the seven trials in each experiment) acceptance index,

distance travelled and total exploration time for each

approach.

(a) Nearest Frontier (b) NBV λ = 0.15

Figure 6: Sample map results for each approach showing

the robot trajectory. ‘A’ and ‘B’ indicate the initial and ﬁnal

robot poses respectively.

Autonomous Exploration and Mapping using a Mobile Robot Running ROS

213

Table 2: Autonomous exploration and mapping results

showing the average acceptance index, average distance

travelled (metres) and the average exploration time for each

approach (seconds).

Acceptance

Index

Distance

Travelled

Exploration

Time

Nearest

Frontier

0.82 25.41 132.48

NBV

λ = 0.15

0.86 30.95 156.28

NBV

λ = 0.5

0.84 40.44 140.90

NBV

λ = 1.0

0.83 35.15 160.56

Visual inspection of the resulting maps, compared

to the ground truth map in Figure 4, and trajectories,

reveals that the Nearest Frontier approach keeps the

robot from visiting certain locations that are quite far

from its current pose. This then leads to unclear fea-

tures and regions in the map, as shown in Figure 6a.

On the other hand, in general, with the NBV approach

the robot seems to cover a larger portion of the envi-

ronment, hence leading to more accurate maps. It is

important to note that the two approaches are subject

to the same termination condition, speciﬁed in Sec-

tion 3, to end exploration. Furthermore, in this study,

although the focus is placed on the map accuracy, the

distance travelled by the robot and the total time of ex-

ploration were also recorded to analyse the efﬁciency

of the algorithms.

5 DISCUSSION

A question arises, of whether it is better to use a sim-

ple and fast exploratory algorithm like the Nearest

Frontier (NF) approach, or the more complex and less

time-efﬁcient Next Best View (NBV) approach. To

evaluate the map accuracy objectively, the differences

between the acceptance indices of the seven trials for

each approach were statistically tested. The boxplot

in Figure 7 illustrates the distribution of these indices

for each of the four adopted approaches. Intuitively,

Figure 7 indicates that the best accuracy was obtained

by the NBV approach since this exhibited the high-

est mean and the smallest variance. However, as the

value of λ is increased, the NBV approach starts con-

verging towards the NF approach. This intuition was

veriﬁed statistically by using the One-Way ANOVA

hypothesis test, where the null (H

) and alternative

) hypotheses were formed as shown in Table 3.

The main result of ANOVA reported that there

exists a signiﬁcant difference in the acceptance in-

Figure 7: Boxplot of the acceptance index distributions.

Table 3: ANOVA statistical test: The Hypotheses.

The NF approach and the NBV approach

with λ = {0.15, 0.5, 1.0} all perform equally

well.

Some of the approaches have better perfor-

mance than the others in terms of map accu-

racy.

dex performance of the four experiments under test.

Hence, this means that H

was rejected. However,this

test does not reveal where the signiﬁcant difference

occurs, and hence the Tukey HSD post-hoc test was

performed. The outcome of this test shows that while

there is a signiﬁcant difference between the NF algo-

rithm and the NBV algorithm with λ = 0.15, there

is no signiﬁcant difference among the other groups.

Since the mean of NBV with λ = 0.15 is larger than

that of NF, then the NBV approach with λ = 0.15

is signiﬁcantly better than NF. Furthermore, the lack

of signiﬁcant difference between NF and NBV with

λ = 0.5, and NF and NBV with λ = 1.0 indicates that

for higher value of λ the NBV algorithm converges

to the NF algorithm. One may argue that this is in

line with the theoretical expectations since as λ in-

creases, the NBV approach prefers the closer candi-

dates (i.e. shorter path lengths) over candidates that

offer a higher area of visibility. Furthermore, one

must note that the parameter λ is application-speciﬁc.

Thus, the threshold value at which the NBV algorithm

is not statistically different from the NF approach may

be different for different environments.

From the means tabulated in Table 2, one can im-

mediately identify a difference between the distance

travelled by the NF approach and the NBV approach.

This difference is also apparent in the time taken to

ﬁnish the exploration and mapping process. Although

such metrics have been used to compare different ex-

ploratory algorithms, it is important to note that in this

case, the distance travelled and the time taken do not

depend on just the adopted strategy. The actual dis-

ICINCO 2016 - 13th International Conference on Informatics in Control, Automation and Robotics

214

tance travelled by the robot does not only depend on

the destination, but it also depends on the local tra-

jectories planned by the local path planner, which ac-

counts for the robot kinematic constraints. Hence, the

length of the travelled path is typically longer than

that which is considered as the global path, L(q, q

) in

both algorithms. Furthermore, the time taken to ﬁnish

the whole process does not only depend on the algo-

rithm or the distance travelled. It also depends on the

speed of the robot while it is navigating. This speed

may be reduced signiﬁcantly if the robot passes close

to an obstacle, which is a precaution taken by the lo-

cal path planner in order to avoid collision with obsta-

cles. For these reasons, this study mainly focused on

the accuracy of the map obtained by the two different

algorithms which were implemented from scratch as

ROS packages for the ﬁrst time.

6 CONCLUSIONS

In this paper, a modular design concept was used

in order to implement a robotic system that can au-

tonomously explore, navigate and map an unknown

environment in ROS. Furthermore, this work con-

tributes a new package to the ROS community. This

package consists of the implementation of two explo-

ration algorithms which can be used independently

of the navigation and the SLAM components. More-

over, the experimental evaluation of the Nearest Fron-

tier (NF) and the Next Best View (NBV) approach re-

vealed that in general, the NBV approach produces

more accurate maps than the NF approach. Further-

more, from this study one can also conclude that as

the parameter λ in the NBV approach is increased, the

NBV algorithm converges to the NF approach. More-

over, the results clearly conﬁrm that the best choice

of an exploration strategy, is highly dependent on the

problem at hand and the environment in question.

REFERENCES

Amigoni, F. (2008). Experimental evaluation of some ex-

ploration strategies for mobile robots. In IEEE In-

ternational Conference on Robotics and Automation

(ICRA), 2008, pages 2818–2823.

Burgard, W., Stachniss, C., Grisetti, G., Steder, B., Kum-

merle, R., Dornhege, C., Ruhnke, M., Kleiner, A., and

Tard´os, J. D. (2009). A comparison of SLAM algo-

rithms based on a graph of relations. In IEEE/RSJ In-

ternational Conference on Intelligent Robots and Sys-

tems, (IROS)., pages 2089–2095.

Carlone, L., Du, Jingjing, K., Ng, M., Bona, B., and Indri,

M. (2014). Active SLAM and exploration with parti-

cle ﬁlters using Kullback-Leibler Divergence. Journal

of Intelligent and Robotic Systems, 75(2):291–311.

Carpin, S. (2008). Fast and accurate map merging for multi-

robot systems. Autonomous Robots, 25(3):305–316.

Dechter, R. and Pearl, J. (1985). Generalized best-ﬁrst

search strategies and the optimality of A*. Journal

of the ACM (JACM), 32(3):505–536.

Dijkstra, E. W. (1959). A note on two problems in connex-

ion with graphs. Numerische mathematik, 1(1):269–

271.

Dissanayake, G., Huang, S., Wang, Z., and Ranasinghe, R.

(2011). A review of recent developments in Simul-

taneous Localization and Mapping. In 6th Interna-

tional Conference on Industrial and Information Sys-

tems, pages 477–482.

Fox, D., Burgard, W., Thrun, S., et al. (1997). The

Dynamic Window Approach to collision avoidance.

IEEE Robotics & Automation Magazine, 4(1):23–33.

Gonzalez-Ba˜nos, H. H. and Latombe, J.-C. (2002). Naviga-

tion strategies for exploring indoor environments. The

International Journal of Robotics Research, 21(10-

11):829–848.

Grisetti, G., Stachniss, C., and W.Burgard (2007). Improved

techniques for grid mapping with Rao-Blackwellized

particle ﬁlters. IEEE Transactions on Robotics,

23(1):34–46.

Makarenko, A. A., Williams, S. B., Bourgault, F., and

Durrant-Whyte, H. F. (2002). An experiment in inte-

grated exploration. In Proc. of the IEEE/RSJ Int. Conf.

on Intelligent Robots and Systems, (IROS), pages 534–

539.

Stachniss, C., Grisetti, G., and Burgard, W. (2005).

Information gain-based exploration using rao-

blackwellized particle ﬁlters. In Robotics: Science

and Systems, volume 2, pages 65–72.

Stentz, A. (1994). Optimal and efﬁcient path planning for

partially-known environments. In Proc. of IEEE In-

ternational Conference on Robotics and Automation,

1994., pages 3310–3317.

Thrun, S., Burgard, W., and Fox, D. (2005). Probabilistic

robotics, chapter 10, pages 309 – 330. MIT press.

Thrun, S., Montemerlo, M., Koller, D., Wegbreit, B., Ni-

eto, J., and Nebot, E. (2004). FastSLAM: An efﬁcient

solution to the simultaneous localization and mapping

problem with unknown data association. Journal of

Machine Learning Research, 4(3):380–407.

Yamauchi, B. (1997). A frontier-based approach for au-

tonomous exploration. In IEEE International Sympo-

sium on Computational Intelligence in Robotics and

Automation, 1997., pages 146–151.

Autonomous Exploration and Mapping using a Mobile Robot Running ROS

215