Towards Total Coverage in Autonomous Exploration for UGV in 2.5D

Dense Clutter Environment

Evgeni Denisov

1 a

, Artur Sagitov

1 b

, Konstantin Yakovlev

2 c

Kuo-Lan Su

3 d

, Mikhail Svinin

4 e

and Evgeni Magid

1 f

Department of Intelligent Robotics, Higher Institute for Information Technology and Intelligent Systems,

Kazan Federal University, 35 Kremlyovskaya street, Kazan, Russian Federation

Federal Research Center ”Computer Science and Control” of Russian Academy of Sciences, Moscow, Russian Federation

Department of Electrical Engineering, National Yunlin University of Science and Technology, Tainan City, Taiwan

Robot Dynamics and Control Laboratory, College of Information Science and Engineering, Ritsumeikan University,

Noji Higashi 1-1-1, Kusatsu 525-8577, Japan

Keywords:

Mobile Robot, Path Planning, Autonomous Exploration and Coverage Algorithm, Next-best-view, Dense

Clutter Environment, Environment Reconstruction.

Abstract:

Recent developments in 3D reconstruction systems enable to capture an environment in great detail. Several

studies have provided algorithms that deal with a path-planning problem of total coverage of observable space

in time-efﬁcient manner. However, not much work was done in the area of globally optimal solutions in dense

clutter environments. This paper presents a novel solution for autonomous exploration of a cluttered 2.5D

environment using an unmanned ground mobile vehicle, where robot locomotion is limited to a 2D plane,

while obstacles have a 3D shape. Our exploration algorithm increases coverage of 3D environment mapping

comparatively to other currently available algorithms. The algorithm was implemented and tested in randomly

generated dense clutter environments in MATLAB.

1 INTRODUCTION

In recent years, 3D reconstruction systems have

evolved towards highly-detailed and accurate recon-

structions of 3D environment, which is now possible

even with monocular cameras (Engel et al., 2014).

Developments in parallel GPU-based computing en-

abled online processing of incoming sensory data,

which allows moving away from ofﬂine data process-

ing. Applications of 3D reconstruction systems in-

clude autonomous navigation in mobile robotics, 3D

scanning and augmented reality applications. 3D re-

construction requires to face various challenges such

as dealing with dynamic environments, small range

sensing, featureless monotone and reﬂective surfaces,

https://orcid.org/0000-0003-0868-0197

https://orcid.org/0000-0001-8399-460X

https://orcid.org/0000-0002-4377-321X

https://orcid.org/0000-0003-2807-2121

https://orcid.org/0000-0003-2459-2250

https://orcid.org/0000-0001-7316-5664

dynamic lighting conditions etc.

Although many of the above mentioned problems

still do not have robust solutions, researchers attempt

to go one step ahead and develop algorithms for 3D

exploration and coverage planning under assumption

that existing SLAM systems are robust and close to

its ﬁnal development. Notable results in these scien-

tiﬁc areas can lead to such advantages as fully au-

tonomous 3D scanning and intelligent reasoning in

3D space. Similarly to 2D exploration path planning

(Gonz

alez-Banos and Latombe, 2002), major issue in

3D exploration is development of a globally optimal

planning algorithm (Lavrenov et al., 2017) that pro-

vides total coverage of visible environment in free

conﬁguration space without a-priori knowledge of en-

vironment (even for static environments). We further

discuss those issues in Section 2.

In this paper, we present a novel algorithm that

optimizes existing exploration and coverage strategies

for a simple 3-DoF UGVs in 2.5D dense clutter envi-

ronments. To evaluate our results, we developed en-

vironment generation algorithm that creates random

Denisov, E., Sagitov, A., Yakovlev, K., Su, K., Svinin, M. and Magid, E.

Towards Total Coverage in Autonomous Exploration for UGV in 2.5D Dense Clutter Environment.

DOI: 10.5220/0007923304090416

In Proceedings of the 16th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2019), pages 409-416

ISBN: 978-989-758-380-3

409

dense clutter environments. Additionally, we intro-

duced a new type of voxels that we refer as ”hard

to observe” to formally deﬁne dense clutter environ-

ments and to provide a better tool for evaluation of

algorithms in this area.

The remainder of this paper is organized as fol-

lows. Section 2 overviews related work and highlights

limitations of existing solutions. Section 3 describes

problem deﬁnition, system setup and our proposed so-

lution approach. Section 4 presents simulation setup

while Section 5 analyzes and evaluates simulation re-

sults. Finally, we discuss our future work plans in

Section 6 and conclude the paper in Section 7.

2 RELATED WORK

The problem of 3D space exploration and coverage is

fairly recent in scientiﬁc research. Heng et al. were

among the ﬁrst authors who introduced the idea of

combining 3D space exploration and coverage prob-

lems together (Heng et al., 2015). They pointed out

that majority of works (further discussed in (Heng

et al., 2015)) on exploration attempted to avoid 3D

space, while research works on coverage were based

on a-priori available entire map of environment and

next-best-view algorithms assumed only single ob-

jects of known sizes. They presented an information

gain-based heuristic solution for unmanned aerial ve-

hicles (UAVs), which relies on selection of closest

frontiers with high information gain, mostly known

as next-best-view gain (see Section 3.2.2). Next-best-

view gain remains a highly popular tool within solu-

tions for 3D space exploration and coverage problem.

In (Ad

an et al., 2015) a similar next-best-view so-

lution was presented for 3-DoF unmanned ground ve-

hicle (UGV) with omnidirectional sensor. Coverage

planning with a known map (Dornhege et al., 2013)

can be used incrementally, but requires additional

planning in order to obtain time efﬁciency. Most

widespread solution to 3D space exploration and cov-

erage problem was proposed by (Bircher et al., 2016)

with RRT* path planning algorithm (LaValle, 1998).

They improved RRT* algorithm (with regard to the

search results and adaptiveness to changes relatively

to other existing approaches) by stopping a current

best branch execution after one node. A newer ver-

sion of this paper introduced an ability to explore vis-

ible space in surfaces instead of voxel s(Bircher et al.,

2018). (Senarathne and Wang, 2016) introduced an

idea of searching edges of known surfaces instead of

unobserved frontiers, which performs well for orthog-

onal environment. (Mendez et al., 2017) proposed to

take into account paths that provide a gain in qual-

ity for modelling results inside voxels. They also

made a multi-robot implementation for cooperative

stereo-pair planning. (Dang et al., 2018) implemented

human-inspired visual attention model that plans ex-

ploration towards visually salient areas in RGB im-

age. (Meng et al., 2017) used genetics algorithms as

an extra step to reﬁne robot movement between se-

lected frontier viewpoints.

Vast majority of research in this area concentrate

on algorithms for UAVs because of a limited tele-

operation control and low capacity batteries that de-

mand automation for UAV-based search and coverage

tasks. Even though design of these algorithms is usu-

ally platform-independent and may suit various sim-

ple robots (including 3-DoF UGVs), a few works, in-

cluding this paper, design solutions specially for 3-

DoF UGVs without targeting for generalized solu-

tions.

We highlighted two limitations of the above men-

tioned works that in our opinion are the most impor-

tant:

1. The algorithm evaluation was not performed in

a highly dense clutter unstructured environment. The

exception is (Zhang et al., 2017) where (unstructured)

3D structures of caves were used, but typical caves

lack a density of cluttered structures and the proposed

solution was not fully autonomous.

2. Low interest towards total capture of an envi-

ronment as a main target of exploration. For UAVs

this arises from expensiveness of viewpoint sampling

in 3D free conﬁguration space. Moreover, most pa-

pers stated that dense viewpoint sampling in 2D free

conﬁguration space, which might be necessary for to-

tal coverage, is too expensive for UGVs, and thus is

ignored.

To test against these limitations for a 3-DoF

UGV setup, we have created a random environment

generation algorithm with a novel heuristic planner

that outperforms current notable RRT*-based algo-

rithm (Bircher et al., 2016).

The motivation behind using a dense clutter en-

vironment is that it’s not clear how well existing

and newly introduced algorithms would perform in

complex environments. Majority of authors simply

test their algorithms inside a limited set of indoor

environments (including own laboratories or univer-

sity buildings). When they use simulated environ-

ment it’s mostly not signiﬁcantly more complicated

than their real working environment. But there ex-

ist a variety of signiﬁcantly more complicated types

of environments, which can be considered important

for application of autonomous 3D exploration algo-

rithms, especially for USAR (Urban search and res-

cue): caves, forests, mountains, construction sites,

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

410

partially or fully destroyed buildings, junkyards, etc.

Aside from that, making dense or even simply high

clutter environment for real experiments is challeng-

ing, especially if a researcher wants to add some ex-

otic objects that come out of environment walls or

hang down from a ceiling.

3 OVERVIEW

This section describes our approach in details. Af-

ter formal problem deﬁnition, an overview for both

path-planning and random environment generation al-

gorithms are presented.

3.1 Formal Problem Deﬁnition

Our problem deﬁnition is similar to other works in

this research ﬁeld. The goal is to autonomously ex-

plore a bounded 3D space V ⊂ R

while minimiz-

ing total coverage time. We classify V as free, occu-

pied, non-observed or residual voxels. Areas that are

not observable from any conﬁguration ξ by a given

robot and its sensor(s) are denoted as V

res

and the

goal of exploration is considered to be achieved when

f ree

∪ V

occ

= V −V

res

To deﬁne dense clutter environments we specify

”hard to observe” voxels or simply V

hto

. The environ-

ment is considered densely cluttered if it contains a

large amount of voxels that are observable only from

a very limited number of free conﬁguration space lo-

cations. Figure 1 demonstrates is a simple example

of V

hto

concept. Green cells denote occupied voxels,

red cells are unobservable voxels, white cells are ob-

servation space, grey cells are non-observation space

and the yellow cells is the only voxel in the picture

that is observable from a limited, but not an empty set

of viewpoints.

Current state-of-the-art algorithms consider a path

that can observe more unobserved voxels as the best

one, which is not the case when we take total cover-

age as a main goal, because these leftovers from paths

without highest count of unobserved voxels will lead

to extra planning steps closer to the end of coverage

process with huge distance costs. The fault of these

approaches becomes more obvious if we consider this

problem as a variation of the traveling salesman prob-

lem. In our research goal we are looking for every

visible voxel (even if it could be observed only from

a single free space conﬁguration) and therefore it is

important not to postpone small unobserved environ-

ment patches for a future exploration.

Aside from that, another possible application of

hto

concept comes from its ability to determine pro-

posed algorithms overall progress without perform-

ing total coverage exploration, because in both real

and simulated experiments process may require sig-

niﬁcant (scale of hours and days) exploration time.

Therefore, for evaluation of algorithms in Section V

we simply count V

hto

based on a-priori knowledge

with a user-deﬁned threshold and provide information

about how many of V

hto

were missed leftovers. Sec-

tion 5 evaluates both algorithms.

Figure 1: An example of V

hto

concept: occupied (green) and

unobservable voxels (red), observation space (white), non-

observation space (grey) and the only voxel that is observ-

able from a limited non-empty set of viewpoints (yellow).

3.2 Path-planning Algorithm Overview

To approach this problem we had to reformulate clas-

sic frontier solutions (for example, in (Gonz

alez-

Banos and Latombe, 2002)) as they do not suited

for coverage in complex 2.5D and 3D environments.

First, step towards a more brute force solution than

a heuristic one. Still our algorithm densely scans

only viewpoints inside a limited squared area around

a robot to avoid high computation costs. Every view-

point inside this area is checked for visibility of unex-

plored surface voxels (V

). A viewpoint with a mini-

mum distance to the robot is considered as a next path.

After reaching the best viewpoint the robot rotates to-

wards unexplored voxels. In order not to emphasize

space exploration priority over obstacle examination,

we ignore ﬂoor surface voxels like if there is no ob-

stacle between them and the robot.

After all contact space inside a scanning area is

fully explored or the robot can not reach a next cell

we switch to RRT* algorithm with NBV gain. A con-

tact cell is marked as fully explored when there is not

a single V

that is visible through this cell in a free

space within robot scanning area. Note that our al-

gorithm searches V

outside the scanning area since

this area contains only viewpoints and so all V

the scanning range from every viewpoint are consid-

ered. Algorithm 1 is visualized in Fig. 2. The robot

(3x3 cell size with yellow dot at the center as its depth

Towards Total Coverage in Autonomous Exploration for UGV in 2.5D Dense Clutter Environment

411

sensor) explores space behind contact space around

(highlighted as red square area). On the left image

white cells are fully explored contact space, blue cells

are partially explored contact space and light grey

cells are free space from which unexplored surface

voxels are visible.

To improve effectiveness of our approach we

added a step that produces a limiting line for convex

regions. Without this step our algorithm tends to leave

unexplored corners if searching area size is smaller

than robot vision range. Limiting line is considered

optimal if it is the shortest line that paths through the

robot location. A cooldown feature is added to lets

the robot moving few extra steps before producing a

next line (Fig. 3: limiting lines are shown in red, yel-

low line is the robot path, which is produced by our

planner; other colors are the same as in Fig. 2).

Algorithm 1: Overview of our planner concept.

Initialization;

while termination condition not met do

if area around robot contains not entirely

explored contact space then

ﬁnd V

hto

and free space;

go to a free cell with highest gain;

produce a limiting line if possible;

else

plan exit strategy with RRT*;

end

Figure 2: Algorithm 1 concept (2D view from above): robot

(yellow) explores space behind contact space (red square

area).Left image: fully explored contact space (white cells),

partially explored contact space (blue cells) and free space

(light grey) from which unexplored surface voxels are visi-

ble.

Estimate of information gain for a viewpoint cell

that we get after evaluation is calculated using the fol-

lowing equation:

Gain

hto

= HTO ∗ e

−d

∗ LIMIT (1)

, where d is the distance cost, HTO is the sign of a

number of visible V

from a viewpoint and LIMIT is

zero if a cell is placed outside of convex region, which

Figure 3: Example of limiting lines (red) sampling. Yellow

line is the robot path.

is closed with a limiting line. Termination condition

for our algorithm is a state with no empty frontiers left

and all of contact space is being fully explored in 2D

top-view with some error. It is possible to continue

with RRT* planner if necessary.

Next subsections brieﬂy explain the concepts of

RRT* and next-best-view gain that were used in our

planner’s step when there is no fully examined contact

space left near the robot.

3.2.1 Exit Strategies with RRT*

Rapidly-exploring random tree (RRT, (LaValle,

1998)) is a path-planning algorithm that expands a

random tree towards unexplored areas until it reaches

a goal. The algorithm samples a collection of con-

nected nodes in SE(3) space without intersections of

non-free space and between their connections. A

branch is considered the best for execution if it has a

highest next-best-view gain relatively to other nodes

of the branch

Gain

nbv

= NBV ∗ e

−λd

(2)

, where d is a distance cost, NBV is a gain from a node

by the count of visible empty voxels and λ is a param-

eter that regulates the gain from nodes in a far dis-

tance. After the robot movement execution, the max-

ima of the second best branch becomes a new goal for

a next random tree. RRT* algorithm improves con-

vergence (of the original algorithm) towards z goal

point. Some of the related work applied RRT* as a

core algorithm. Additionally, we limit a max amount

of tree nodes and stop expansion after reaching a cer-

tain number of nodes. A typical result of RRT* sam-

pling is visualized in Fig. 4.

In our algorithm there are two exit strategies that

involve RRT* path-planning algorithm with NBV

gain. The ﬁrst is for exploring empty environment

without visible obstacles in range of a depth sensor

and for exiting dead ends. The second is used in rare

occasions when the robot may get stuck in a dead-end

with only a narrow corridor for returning. To avoid

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

412

Figure 4: RRT* produces a tree with nodes as NBV view-

points. The process stops when the tree reaches the end

point (red dot) from the start point (red circle).

this we added another strategy with a different us-

age of RRT* without NBV gain: if the robot does

not receive information gain from RRT* planner for a

few steps then it ﬁnds a closest empty frontier and in-

creases the limit of nodes by a large number to ensure

the robot will ﬁnd a way out.

3.2.2 Next-best-view Gain

Next-best-view is a class of problems that target for

optimizing an environment coverage path to minimize

a number of required camera shots. In addition next-

best-view term is widely used as a description for

views with high gain of new information about en-

vironment. Consecutively, every conﬁguration ξ is a

point of view, which represents some part of an envi-

ronment by a set of visible and unmapped voxels. The

most common method for gathering this data is to cast

a grid of rays from a point of view with a given ﬁeld

of view and limited distance. For this action we use a

raycast algorithm implementation for 3D voxel-based

grid (Williams et al., 2005). After all rays are pro-

jected, the next-best-view gain describes a proportion

between amount of occluded and not-occluded rays.

NBV raycasting concept is visualized in Fig. 5.

3.3 Random 2.5D Dense Clutter

Environment Generation

We propose an algorithm for generation of 2.5D clut-

tered environment for 3-DoF UGVs robot conﬁgura-

tion. In this paper 2.5D restricts 3D space by pro-

hibiting elevation in free space that is a typical struc-

ture inside buildings. Algorithm consists of two steps.

Firstly, we sample onto 2D plane a random set of

points and lines between them with a random width in

a way that there is no single point without a connec-

tion to some other point. The result deﬁnes collision-

free space. Secondly, we sample random voxels onto

Figure 5: Next-best-view gain is deﬁned by a set of raycasts.

The density of rays is decreased for better visualisation.

3D space outside collision-free space with addition of

much smaller amount of random cuboids increasing

variety. For visualization we use a voxel plotting tool

that is made by (Shabat and Fischer, 2015). Figure 6

presents an example of a generated environment.

Figure 6: This ﬁgure shows a labeled generated environ-

ment. Observable voxels are green, non-observable voxels

are red, ”hard to observe” voxels are yellow (V

hto

). Walls

and voxels above the robot head are omitted for better visu-

alization.

4 VIRTUAL EXPERIMENTS

For this paper, we selected a 3-DoF conﬁguration

UGV robot with position and yaw represented by

ξ = (x, y, ψ)

. All path-planning is done inside a

collision-free space. A depth sensor is located on a

ﬁxed height from a ground plane, has a limited ﬁeld

of view (90 degrees in horizontal axis and 70 degrees

in vertical axis) and a limited range (R = 15 voxels).

Towards Total Coverage in Autonomous Exploration for UGV in 2.5D Dense Clutter Environment

413

The depth sensor pitch is lowered to make sure the

robot can see a path in front, parts of environment

that are located above the robot head are not visible.

It is preferable to place the camera that way in order to

avoid an extensive search of a highest reachable voxel

rows, but for complete scanning results we plan to uti-

lize a high-DoF mobile robot with vertically mobile

camera as a part of our future work. The robot size

is 3x3x6 voxels. Generated environments are repre-

sented inside a 70x70x6 box. The limit for maximum

number of RRT* nodes is set relatively high for 2D

free space in order to increase a chance of obtaining

more V

hto

at the end of the coverage process. Convex

detection range is set higher than the sensing range,

but HTO scanning range is set lower to speed up com-

putations.

The algorithm was implemented in MATLAB,

except the Woo raycasting algorithm library (C++

code), which was compiled as a mexa64 ﬁle in order

to speed up the computations. We assume an ideal

3D reconstruction in our setup to achieve persistent

estimations for equivalent experiment settings. The

algorithm performance was evaluated with regard to

a RRT* planner (Bircher et al., 2016), although we

dismissed the idea of cutting off a current best branch

on the ﬁrst node as it leads to worse performance in

our dense clutter environment than without it. Other

frontier algorithms were considered, but they lack the

ability to determine unreachable frontiers or can ig-

nore actually reachable frontiers while randomly sam-

pling viewpoints.

We took 6 generated maps (Fig. 7) from our ran-

dom environment generator. The maps were chosen

on the variety of quantity of V

hto

they have and topo-

logical differences. In every map the robot starts at

the left bottom corner. Due to stochastic nature of

the RRT* algorithm, we additionally included 3 ex-

tra trials for both planners on every map. Termination

condition for both planners in these simulation experi-

ments was set as a coverage of 97 % and usually HTO

planner explored all walls up to this point, so after 97

% both algorithms worked in the same way. The al-

gorithms parameters are presented in Table 1.

5 RESULTS

We evaluated performance of both algorithms com-

paring odometrical distance, computation time and

coverage (both V

hto

and V

occ

). Rotation was also in-

cluded in the distance metric. For evaluation we as-

sumed voxel’s size of 0.1m. Possible coverage for a

map was measured with the next-best-view method

with 360 degree angle from free space grid with dou-

Table 1: Parameters for algorithms. ’v’ stands for voxels.

Parameter Value

Sensor range 15 v

Robot’s height 6 v

Robot’s size 3x3 v

RRT* tree edge 6 v

λ for RRT* 0.4

RRT* max nodes 25

HTO scanning area 20x20 v

Convex detection area 40x40 v

Figure 7: All 6 generated maps after full or partial explo-

ration. Blue voxels are both unobserved and unobservable

voxels from ground truth and green voxels are explored vox-

els. Blue lines are taken paths in exploration.

bled density. V

hto

threshold was set to 9 which leads

in selected maps to 5-10 % of total coverage.

We run total coverage tests on several maps and it

took our planner to reach 96-97 % coverage with fully

explored obstacles (stated by our planner) in 600-900

s computation time, while for RRT* planner it took

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

414

1100-1600 s in average to obtain the same coverage

level (except for the most simple map that was once

completed in 601 s) The measurements did not take

into account huge distance penalties, which resulted

in extra thousands of meters (above 2000 m), while

our algorithm rarely went above 1000 m. An example

of the resulting paths by both planners are presented

in Fig. 8.

Figure 8: An example of the resulting paths of our planner

(blue lines, on the left) and RRT* planner (red lines, on the

right).Both planners started at the left-bottom corner of the

map.

Total coverage measure of 100 % was not

achieved because we only check visibility of surface

voxel’s central point, while next-best-view method

uses raycasting visibility check for voxel’s edges. Re-

maining 3 % in most cases also contains more than

8 % of V

hto

. After checking total coverage we tried

employing our V

hto

metric as a prediction tool. Fig-

ure 9 shows that in most cases it is useful to deﬁne an

overall success of a selected algorithm rather than an

overall coverage. The plot on the right demonstrates

that our planner outperforms RRT* planner by V

occ

and V

hto

coverage rate and we link this success to to-

tal coverage success.

Figure 9: V

hto

coverage (dashed line) of our planner (red)

and RRT* planner (blue). The ﬁgure shows the worst case

scenario for our planner. RRT* planner’s V

hto

coverage is

generally worse and its coverage process is uneven.

6 FUTURE WORK

Our next goal is to further improve the proposed algo-

rithm. Firstly, currently we decided to accept a huge

rotation penalty it is much less than distance penalty.

Secondly, the shortcoming of limiting convex regions

is that this step ignores non-convex regions, which are

also important. Thirdly, although our planner out-

performs RRT* planner in minimizing traveled dis-

tance, it is more computationally demanding and re-

quires extensions for robots with movable sensor in

3D space (UAVs or High-DoF UGVs). Fourthly, there

is a lack of reliable parameters in order to deal with

random obstacles. We attempted to ﬁnd more parame-

ters to improve quality of the resulting path-planning

in addition to restricting the robot within convex re-

gions, but the obtained solutions did not seem to work

better in general cases for dense clutter environment.

Therefore, an omnidirectional scanner would better

suite for our solution, but we do not limit the appli-

cation of our algorithm solely to this solution. Fifthly,

the random environment generator needs to be im-

proved towards a real 3D environment or (as a start-

ing point) should include obstacles above the robot’s

head, which would require major changes in our al-

gorithm.

Another goal, which is more technical that the

above mentioned improvements, is to port the both

planning and random environment generation algo-

rithms to C++ and to integrate them into Robotic Op-

erating System framework and to enable its function-

ality in Gazebo simulator (Afanasyev et al., 2015).

The random environment generator will be modiﬁed

to export generated models as 3D models for further

simulation experiments in Gazebo, which provides

actual physical simulation. Then we will re-evaluate

current restrictions that were taken for MATLAB sim-

ulation to be more suitable for real experiments. In

addition, we are interested to implement an instru-

ment of automatic map generation (similar to a con-

venient automatic map generation tool (Lavrenov and

Zakiev, 2017)) of dense clutter environments for fur-

ther autonomous exploration by a UGV in order to

provide a convenient benchmark for various coverage

algorithms comparison.

We believe that V

hto

metric could be a useful tool

for real world experiments, where it is possible to

check for hard to observe locations of an environment

and then check their representation in reconstruction

manually by scanning only those parts instead of tak-

ing an entire scan of the environment. After all, we

consider expanding the scope of our current work to

face more difﬁcult problems,e.g., such as dynamic en-

vironment and multi-robot exploration.

Towards Total Coverage in Autonomous Exploration for UGV in 2.5D Dense Clutter Environment

415

7 CONCLUSIONS

In this paper, we presented a novel algorithm for

autonomous exploration and coverage problem for

UGVs in 3D dense clutter environment. The algo-

rithm plans the exploration path towards ﬁnding spe-

ciﬁc voxels that are hard to observe. Our approach

provided a solution that is more robust for 2.5D dense

clutter environment with highly non-linear contact-

space comparatively to currently known solutions, as

in practice they largely ignore such features as con-

tact space and narrow passages, while our algorithm

is opportunistically looking for such locations.

The MATLAB code is available for academic

community for research and educational purposes in

our Gitlab repository

ACKNOWLEDGEMENTS

The reported study was funded by the Russian Foun-

dation for Basic Research (RFBR) according to the

research project No. 19-58-70002. Part of the work

was performed according to the Russian Government

Program of Competitive Growth of Kazan Federal

University.

REFERENCES

an, A., Quintana, B., V

azquez, A. S., Olivares, A., Parra,

E., and Prieto, S. (2015). Towards the automatic scan-

ning of indoors with robots. In Sensors (Basel). 2015

May; 15(5): 11551–11574.

Afanasyev, I., Sagitov, A., and Magid, E. (2015). ROS-

based SLAM for a Gazebo-simulated mobile robot in

image-based 3D model of indoor environment. In In-

ternational Conference on Advanced Concepts for In-

telligent Vision Systems, pages 273–283. Springer.

Bircher, A., Kamel, M., Alexis, K., Oleynikova, H., and

Siegwart, R. (2016). Receding horizon ”next-best-

view” planner for 3d exploration. In IEEE In-

ternational Conference on Robotics and Automation

(ICRA), 2016.

Bircher, A., Kamel, M., Alexis, K., Oleynikova, H., and

Siegwart, R. (2018). Receding horizon path plan-

ning for 3d exploration and surface inspection. In Au-

tonomous Robots, February 2018, Volume 42, Issue 2,

pp 291–306.

Dang, T., Parachristos, C., and Alexis, K. (2018). Visual

saliency-aware receding horizon autonomous explo-

ration with application to aerial robotics. In IEEE In-

ternational Conference on Robotics and Automation

(ICRA), 2018.

https://gitlab.com/LIRS Projects/Simulation-3d-reconstru

ction/tree/master/autonomous exploration and coverage

Dornhege, C., Kleiner, A., and Kolling, A. (2013). Cover-

age search in 3d. In IEEE International Conference

on Robotics and Automation (ICRA), 2013.

Engel, J., Sch

ops, T., and Cremers, D. (2014). Lsd-slam:

Large-scale direct monocular slam. In IEEE Euro-

pean Conference on Computer Vision (ECCV), 2014,

pp 834-849.

Gonz

alez-Banos, H. H. and Latombe, J.-C. (2002). Naviga-

tion strategies for exploring indoor environments. In

The International Journal of Robotics Research, vol.

21, no. 10-11, pp. 829–848, 2002.

Heng, L., Gotovos, A., Krause, A., and Pollefeys, M.

(2015). Efﬁcient visual exploration and coverage with

a micro aerial vehicle in unknown environments. In

IEEE International Conference on Robotics and Au-

tomation (ICRA), 2015.

LaValle, S. M. (1998). Rapidly-exploring random trees a

new tool for path planning. Technical report.

Lavrenov, R., Matsuno, F., and Magid, E. (2017). Mod-

iﬁed spline-based navigation: guaranteed safety for

obstacle avoidance. In International Conference on

Interactive Collaborative Robotics, pages 123–133.

Springer.

Lavrenov, R. and Zakiev, A. (2017). Tool for 3d gazebo map

construction from arbitrary images and laser scans.

In 2017 10th International Conference on Develop-

ments in eSystems Engineering (DeSE), pages 256–

261. IEEE.

Mendez, O., Hadﬁeld, S., Pugeault, N., and Bowden, R.

(2017). Taking the scenic route to 3d: Optimising re-

construction from moving cameras. In IEEE Interna-

tional Conference on Computer Vision (ICCV), 2017.

Meng, Z., Qin, H., Chen, Z., Chen, X., Sun, H., Lin, F.,

and Ang Jr., M. H. (2017). A 2-stage optimized next

view planning framework for 3-d unknown environ-

ment exploration and structural reconstruction. In

IEEE Robotics and Automation Letters, Volume 2, Is-

sue 3, July 2017.

Senarathne, P. G. C. N. and Wang, D. (2016). Towards au-

tonomous 3d exploration using surface frontiers. In

IEEE International Symposium on Safety, Security,

and Rescue Robotics (SSRR), 2016.

Shabat, Y. B. and Fischer, A. (2015). Design of adaptive

porous micro-structures using curvature analysis for

additive manufacturing. In the 25th CIRP Design con-

ference. 2015, Haifa, Israel.

Williams, A., Barrus, S., Morley, R. K., and Shirley, P.

(2005). An efﬁcient and robust ray-box intersection

algorithm. In SIGGRAPH ’05 ACM SIGGRAPH 2005

Courses, Article No. 9.

Zhang, G., Shang, B., Chen, Y., and Moyes, H. (2017).

Smartcavedrone: 3d cave mapping using uavs as

robotic co-archaeologists. In IEEE International

Conference on Unmanned Aircraft Systems (ICUAS),

2017.

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

416