Experimental Evaluation of Some Indoor Exploration Strategies

Jens Wettach and Karsten Berns

Robotics Research Lab, University of Kaiserslautern, Gottlieb-Daimler Straße, Kaiserslautern, Germany

Keywords:

Robotics, Exploration, Indoor Environments, Experiments, Evaluation.

Abstract:

A key capability of any indoor service robot is to explore arbitrary, unknown environments in order to record

a complete and correct map in minimal time. Such a map is a prerequisite of common tasks like surveillance,

transportation as well as search and rescue. In recent years a series of solutions has been proposed by the

authors: a dynamic enhancement of the frontier-based approach, ground plan-based exploration and a hybrid

combination of both. This paper evaluates the performance of each of these strategies within an everyday

ofﬁce scenario in simulation and reality and discusses their pros and cons.

1 INTRODUCTION

In order to perform a service task like surveillance

or transportation efﬁciently, a robot needs an accurate

map of its working space. Ideally, the robot should be

able to derive this map autonomously, so it can adapt

to a priori unknown environments without user inter-

vention. Feasible SLAM solutions exist for map con-

struction. But calculating a trajectory, that allows to

record a complete and correct map in minimal time,

is still an open problem. In the past, three solutions

have been developed by the authors of this paper, all

following the cyclic next best viewing pose (NBV) ap-

proach of (Amigoni, 2005). Each iteration consists of

three steps: map update, NBV calculation and target

approach. The strategies mainly differ in the type of

map and the sophistication of the NBV computation.

Subject of this paper is to evaluate and compare

the power of these approaches in a daily ofﬁce sce-

nario, both under well-deﬁned conditions in simu-

lation and in a real ofﬁce scenario with dynamic

changes (doors, people). Nota bene, the focus is on

the performance of the proposed solutions in a typical

application environment, especially on the differences

between simulation and reality as well as between the

three strategies. In contrast, an analysis of their ef-

fectiveness in different environmental layouts and a

comparison to competitive state of the art approaches

is subject of ongoing research.

Section 2 presents important state of the art solu-

tions regarding indoor exploration. The sophistication

of the respective approaches is condensed for putting

the strategies under study into context. Since each of

them has been published previously, only their main

ideas and features are introduced in section 3. The

experimental evaluation is described in section 4, and

the ﬁndings are condensed in section 5.

2 STATE OF THE ART

(Yamauchi, 1997) has developed one of the most fa-

mous approaches: NBVs are derived from the frontiers

between unknown space and regions already known

as free in an occupancy grid map. The one that maxi-

mizes expected information gain, i. e. frontier length,

is selected in Greedy manner. (Oriolo, 2004) de-

scribes a far simpler strategy based on random mo-

tions. (Freda, 2005) biases the random motion to-

wards the free frontiers of the known space.

(Gonz

alez-Ba

nos, 1999) and (Gonz

alez-Ba

nos,

2002) create polygonal maps of the environment

based on the concept of the safe region, i. e. the areas

already known as traversable. NBVs are sampled

everywhere within these areas and scored by weight-

ing expected information gain and travel cost. (Sur-

mann, 2003) uses polygonal 2D maps on several hori-

zontal height levels in order to generate a 3D model of

the working space. Within each map, NBVs are calcu-

lated along the frontiers, and the most promising one

regarding information gain is selected greedily.

(Makarenko, 2002) has deﬁned the problem of

automatically mapping an unknown environment for-

mally as integrated exploration. Accordingly, three

tasks have to be performed at the same time: map-

ping, localization and navigation. Hence, three cri-

280

Wettach, J. and Berns, K.

Experimental Evaluation of Some Indoor Exploration Strategies.

DOI: 10.5220/0005977502800287

In Proceedings of the 13th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2016) - Volume 2, pages 280-287

ISBN: 978-989-758-198-4

teria are evaluated simultaneously as a weighted sum:

information gain, travel cost and localizability. NBVs

are either taken from the frontiers or from locations

with a high chance of relocalization. (Amigoni, 2004)

and (Amigoni, 2005) propose NBV exploration as a

common solution: ﬁrst the current map is updated;

then a series of viewpoints is generated and the most

promising one is selected via a scoring function; ﬁ-

nally this position is approached. In contrast to the

previously used ad hoc functions, where determining

feasible weights is an open problem, here the entropy

of information gain and of travel cost is calculated

based on information theory. Whereas (Amigoni,

2004) balances these two criteria, (Amigoni, 2005)

performs a multi-objective optimization using also lo-

calization (map overlap) for calculating the Pareto-

optimal viewpoint candidate as NBV. As a further

extension, (Basilico, 2011) applies multi-criteria de-

cision making (MCDM) and fuzzy functions for an

online adaptation of the exploration criteria, e. g. in-

formation gain, travel cost and either chance of estab-

lishing a wireless link to a ﬁxed base station or chance

of reaching a charging station.

Beside the sophistication of the NBV scoring, the

kind of mapped features and the strategy of map

evaluation inﬂuences the exploration performance.

(Wurm, 2008) proposes to segment the already known

map according to the structure of the environment,

e. g. into different rooms, and to perform a segment-

wise exploration. Assigning to each robot a different

segment yields a very efﬁcient coordination of multi-

ple robots. Similarly, (Schmidt, 2006) combines re-

active behaviors for exploring the current room and

deliberative behaviors for switching between rooms.

The trajectory is continuously adapted to the changes

of the map. (Maffei, 2014) also performs an online

adaptation by regarding exploration as a boundary

value problem in a potential ﬁeld: obstacles have high

potential, unknown regions a low value and already

visited areas a variable one. Revisiting known areas

supports relocalization. Distortions are added to bal-

ance relocalization and information gain.

Finally, several comparative studies analyze the

pros and cons of existing approaches in order to de-

rive hints for future improvements. (Amigoni, 2008)

reports that a Greedy selection of NBVs based on in-

formation gain performs better than a random choice,

and balancing utility and travel cost is even more ef-

ﬁcient. However, a signiﬁcant difference between an

ad hoc function (Gonz

alez-Ba

nos, 2002) and a scor-

ing measure based on information theory (Amigoni,

2004) could not be approved. In contrast to that,

(Holz, 2011) has demonstrated that balancing two

criteria (Gonz

alez-Ba

nos, 2002) is not always better

than a simple Greedy approach choosing the clos-

est frontier and that a sound design of the viewing

pose candidate (vp) evaluation function like MCDM

(Basilico, 2011) pays off. Furthermore, a continu-

ous re-evaluation of the selected NBV based on on-

line map updates and a segmentation of the working

space have been proposed for increasing exploration

efﬁciency. In this regard, (Amigoni, 2013) has ana-

lyzed the inﬂuence of the frequency of map updates

(perception) and of NBV selection (decision) on the

exploration performance. As a result, a frequency-

based approach with online updates outperforms the

classic event-based scheme (updates only at an NBV),

especially in unstructured environments, but the fre-

quencies have to be adapted to the computational ef-

fort of the continuous re-evaluation.

The strategies discussed in this paper rely on a

continuously updated 2.5D grid map, but on a dis-

crete NBV selection. Hence, the target is not changed

once it has been determined, and either it is reached

in a reasonable time span or marked as inaccessible.

The trajectory is adapted online to new map informa-

tion and dynamic changes (people walking around,

doors opened or closed) during NBV approach. Thus

it may leave the areas that are known at the mo-

ment of the target selection. This yields a more ﬂex-

ible approach than those of (Yamauchi, 1997) and

(Gonz

alez-Ba

nos, 2002). Vps are sampled along

frontiers or within polygonal maps of ﬂoor and ceil-

ing, derived from a 3D reconstruction of the working

space similar to (Surmann, 2003). The scoring func-

tion is designed as weighted sum as in (Makarenko,

2002) and balances information gain, travel cost and

dispersion of NBVs. Localization is delegated to a

continuously operating particle ﬁlter and therefore no

issue for the scoring. The focus is on examining how

the developed approaches perform in a real world ap-

plication, what the beneﬁt of evaluating 3D features

is, how the strategies cope with dynamic changes and

how a combination of a sophisticated strategy with a

fallback to a simple, but always working approach in-

ﬂuences the reliability of the exploration process.

3 EXPLORATION STRATEGIES

The three exploration approaches under observation

have been presented in preceding papers. For a bet-

ter understanding, the main ideas and key features are

summarized in the following. Common base is an in-

door robot, shown in ﬁgure 1. It is equipped with a

differential drive, a planar laser scanner at the front

and the back, a circumferential belt of ultrasonic sen-

sors and a 3D laser scanner.

Experimental Evaluation of Some Indoor Exploration Strategies

281

planar scanner

ultrasound

3D scanner

Figure 1: Robot and sensors used for exploration.

Figure 2 presents the common concept for all three

strategies. Hardware abstraction, SLAM and naviga-

tion are shared. On the lowest level, distance meas-

urements are fused into local obstacle memories and

overlaid sector maps as virtual sensors. Besides, ab-

stract motion commands are transformed into con-

trol values for the differential drive. Localization is

based on odometry plus continuous pose corrections

via DP-SLAM (Eliazar, 2003). Mapping constructs a

Exploration

Mapping

SLAM

Localization

Path Planning

Navigation

Obstacle Avoidance

Hardware Abstraction

map

NBV

sensor data sensor data

motor

control

map

Figure 2: Concept of integrated exploration.

2.5D occupancy grid map, ﬁlled with distance meas-

urements of all laser scanners. The SLAM unit facil-

itates online map updates, while the robot is moving.

Obstacle avoidance is realized by a network of anti-

collision, keep distance and evasion behaviors accord-

ing to the iB2C methodology (Proetzsch, 2010). The

behaviors evaluate the sector maps for analyzing the

free space around the robot and compute abstract mo-

tor control commands. The navigation unit guides the

robot to an NBV computed by the explorer. This goal

approach uses A

for planning an initial path. It is

transformed into an elastic band that performs a con-

tinuous free space analysis and adapts the path ac-

cording to the online map updates (Quinlan, 1993).

Consequently, the three approaches differ in the

strategy for calculating the series of NBVs. The dy-

namic frontier-based approach only evaluates the oc-

cupancy grid map, whereas the ground plan-based

and the hybrid approach record the layout of ﬂoor and

ceiling for deriving better vp candidates.

SCORE(vp

) = λ

· information gain(vp

)

+ λ

· unknown cells(vp

)

− λ

· cost(vp

)

− λ

· occupied cells(vp

)

− λ

d p

· distance to past nbv(vp

)

− λ

· distance to unreachable nbv(vp

)

(1)

In order to determine the NBV at a particular explo-

ration step, a set V P = {vp

} of viewing pose candi-

dates is generated, and the best one is calculated via

the scoring function (1). This function is the same for

all three approaches regarding its layout as weighted

sum and its parameters λ

λ, but the contributing factors

are calculated differently for each strategy based on

the features that are recorded in the respective map.

Advantages of a vp are expected information gain

according to the particular strategy and number of un-

known cells at vp in the occupancy grid map. Dis-

advantages are path cost, number of occupied grid

cells at vp as well as distance to past and unreach-

able NBVs. The last two factors guarantee an efﬁcient

dispersion of NBVs. First, the robot has to keep off

previously visited areas, at least as long as there are

promising vps in unknown regions. Even more im-

portant is to stay away from NBVs that could not be

reached in the past, because these give a hint for an

unknown, but attractive area that cannot be mapped,

e. g. a neighboring room only visible through a win-

dow. Hence the last factor avoids deadlocks. All

values are calculated in relation to the respective op-

timum, that is minimum for dispersion and maxi-

mum for all other factors, i. e. they are normalized

to [0, 1]. The set of scoring weights λ

λ = (λ

,λ

d p

,λ

) deﬁnes the exploration “mood”, e. g. curios-

ity or guardedness. Reasonable values have been de-

termined via comprehensive experiments (sec. 4.1).

An important issue of any exploration strategy is

the termination criterion. The challenge is to decide

whether the accessible areas have been mapped ex-

haustively. Especially in real world scenarios, the ro-

bot may not be able to approach some visible places

closely enough due to narrow space or a door closed

casually. For this purpose, any NBV, that cannot be

reached in a certain period of time, is recorded as un-

reachable, and new vp candidates, that are close to

such one, get a high penalty. On the other hand, the

expected information gain decreases over time since

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

282

the remaining unknown areas shrink. Hence, the al-

gorithm stops when the score of the vp candidates at

a certain exploration step indicates, that there are no

rewarding targets left.

Dynamic Frontier-based Exploration. (Wettach,

2010) has extended the frontier-based strategy of (Ya-

mauchi, 1997) by a continuous localization and the

option to leave the safe region while approaching the

NBV. This facilitates online map updates even in un-

known areas and reduces detours. Vp candidates are

derived from the center of the free frontiers. Expected

information gain is calculated as length of the respec-

tive frontier. Unknown and occupied cells are calcu-

lated within a circular area with a predeﬁned radius

around the vp. Path cost is given by the A

planner.

The distances to past and unreachable NBVs are com-

puted via the Euclidean norm.

Ground Plan-based Exploration. (Wettach, 2012)

calculates NBVs from the differences between ﬂoor

and ceiling: space visible at the ﬂoor, but hidden at

the ceiling is usually caused by open doors; free space

in the ceiling, that is blocked on the ﬂoor, marks inac-

cessible regions that deserve a closer look, e. g. due to

furniture or a closed door with skylight. At each NBV

a 3D point cloud is collected via a 360

◦

panorama

scan, and the main structures of the environment are

extracted by RANSAC-based plane ﬁtting. This way

the ground plan of ﬂoor and ceiling is reconstructed

and recorded as polygonal maps. Vps are derived as

center of the difference polygons, and expected infor-

mation gain is calculated as polygon area. All other

scoring factors are computed from the grid map as for

the frontier-based approach.

Hybrid Exploration. In extreme situations the

ground plan approach may fail: unobstructed areas

like corridors do not provide signiﬁcant differences

between ﬂoor and ceiling and thus no valid vp candi-

dates; in narrow places like door frames a panorama

scan may not be performed safely due to the protrud-

ing 3D scanner. (Wettach, 2014) has developed a hy-

brid strategy, that evaluates the ground plans when-

ever possible and that uses the frontier information as

a fallback. Expected information gain is either com-

puted from the difference polygons or from the fron-

tier length, whereas all other contributing factors of

the scoring function (1) are derived from the grid map.

4 EXPERIMENTAL EVALUATION

The exploration strategies have been evaluated in a

simulated ofﬁce scenario, shown in ﬁgure 3, as well

as in its real world counterpart. It consists of an

meeting room

chair

tables

kitchen

stairway

open doors

pillars

hallway

robot

entrance

hall

start pose

Figure 3: Simulated ofﬁce scene (from (Wettach, 2014)).

Open doors are marked with labels a, b, c.

L-shaped hallway with an open entrance hall in be-

tween and adjacent ofﬁces on each side. Since the

tests took place at common business hours, the ofﬁce

rooms have been made inaccessible order to limit the

complexity of the experiments. Challenges are repre-

sented by a meeting room, equipped with tables and

chairs, as well as by a small kitchen. The respective

doors have been kept statically open. Figure 1 shows

the robot passing the door (b) from meeting room to

kitchen. Obviously, changing the room is a critical

operation due to the dimensions of doors and robot.

Besides, both rooms can be left via opposite doors,

yielding a potential loop as trial for the DP-SLAM

unit. The descending stairway at the south-east end

of the corridor represents a negative obstacle in real-

ity and is modeled by half-height walls in simulation.

The virtual model serves as ground truth, because it

has been built based on the construction plan of the

building. Naturally, the real scenario contains addi-

tional difﬁculties: the kitchen is equipped with a ta-

ble and a dresser; there are environmental dynamics,

e. g. people walking around, doors casually opened

and closed; sometimes doors closed in simulation stay

partially open, so the robot can look into but not enter

the neighboring room, which yields a big challenge

for the deadlock prevention (see ﬁgures 8, 9, and 10).

4.1 Impact of the Scoring Weights

In order to show the impact of the weights λ

λ in func-

tion (1), two different sets λ

= (5, 1,15, 1,8,50) and

= (20,10, 12,5,8, 50) are compared via a test run

of the dynamic frontier-based exploration in simula-

tion. In both sets, deadlock prevention is most im-

Experimental Evaluation of Some Indoor Exploration Strategies

283

portant in order to cope with traps of attractive, but

inaccessible areas like the stairway in the south corri-

dor. λ

then focuses on the path cost in order to avoid

oscillations between opposite borders of the known

space, i. e. to exploit locality. The distance to past

NBVs serves as a counterpart in order to achieve a

reasonable dispersion of NBVs. Expected information

gain is slightly less important. For putting the focus

on these four criteria, the unknown and the occupied

grid cells in the vicinity of a vp have least inﬂuence.

Apart from deadlock prevention, λ

emphasizes

information gain, followed by path cost in order to ex-

ploit locality. Number of unknown grid cells around

the target is a second measure of information gain and

has slightly less inﬂuence, similar to distance to past

NBVs as antagonist to locality. Estimated accessibil-

ity of the target is least important.

Figure 4: Simulation result of dynamic frontier-based ex-

ploration using scoring weights λ

= (5,1, 15,1,8, 50), with

occupied grid cells (red), NBVs (blue), trajectory (green)

and map orientation (cross hairs).

Table 1: Summary of experimental results.

Run Comp Corr T

Exp

∑

NBVs

% % min w/o scan m

ﬁg. 4 6

9 2.79 56.5 29/8 232

ﬁg. 5 14.4 1.87 3

0 22/2 160

ﬁg. 6 14.5 1

4 37.5 1

6 9

ﬁg. 7 11.5 2.38 53.8 24/5 139.9

ﬁg. 8 5

1 2

5/7 2

ﬁg. 9 94 33 273.2

ﬁg. 10 83.8 30/5 247.4

Figure 4 shows the test run with λ

, and ﬁgure 5

the one with λ

. The benchmark values are summa-

rized in table 1 (see sec. 4.2 for a description). λ

yields signiﬁcantly more detours than λ

, i. e. oscilla-

tions between meeting room, kitchen and corridor,

which leads to a much longer trajectory (232 m vs.

160 m) and exploration time (56.5 min vs. 30 min).

At least, λ

is competitive regarding map complete-

ness and correctness. In sum, λ

seems more promis-

ing due to the reduced execution time. Since here the

focus is on evaluating the different strategies depend-

ing on environmental conditions, λ

is used for all

subsequent experiments.

4.2 Comparison of the Three Strategies

In the following a test run with the robot starting in the

meeting room facing to the east is performed for all

three exploration strategies, ﬁrst in simulation, then in

the real environment. System parameters like scoring

weights, preﬁltering of vp candidates and number of

replanning operations per NBV are ﬁxed for all runs.

The maximum velocity of the robot has been limited

to 0.35 m/s. In simulation, the DP-SLAM unit and

corresponding pose correction have been switched off

in order to examine the power of the approaches with-

out odometry errors.

For comparing the results, a set of benchmarks has

been deﬁned. Non-functional criteria are: robustness

regarding collision avoidance: the robot must not hit

any object; deadlock prevention: the algorithm has to

terminate in ﬁnite time; a priori knowledge: a grid

cell may either be free or occupied, ﬂoor and ceiling

can be described by a set of polygons; genuineness of

application: the test has to be performed in an every-

day scene. These criteria have been fulﬁlled equally

by each experiment. Functional criteria are map com-

pleteness (Comp), map correctness (Corr) and explo-

ration time (T

Exp

), because the goal is to record a

complete and correct map in minimal time. Number

of NBVs with and without 3D scan and length of the

exploration trajectory

are subordinated param-

eters, indicating how the needed time has been spent.

All these values are summarized in table 1.

Exp

∑

NBVs and

are measured directly.

Comp is the number of cells missing in the recorded

map wrt. a ground truth map (derived from a con-

struction plan) versus total number of cells in the true

map (false negatives). Corr is calculated as number

of cells in the recorded map that are not present in

the true map versus total number of cells in this map

(false positives). For reasons of comparability, Comp

and Corr are always calculated for the grid map. In

reality, these two criteria can only be qualitatively in-

spected due to signiﬁcant map inaccuracies arising

from SLAM errors.

4.3 Simulation Experiments

Figure 5 shows the result of the dynamic frontier-

based exploration in simulation. First, meeting room

and kitchen are explored (vp 0-2), followed by the

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

284

south corridor (3-9). At positions 10, 11 time con-

suming oscillations back to distant, previously visited

areas occur. Eventually, the robot explores the en-

trance area and north corridor (12-20), before return-

ing to the last open frontier at the staircase (21). A

panorama scan took 30 s on average, i. e. 10 min in

total for 20 NBVs (no scan at vp 1, 21 due to close ob-

stacles). T

Exp

= 30min is the minimum over all test

runs.

= 160m yields a minimum travel time of

7 min 37 s (at maximum velocity). The time penalty

of 12 min arises from slow-downs in narrow passages.

Obviously, the robot never passed door (c) between

meeting room and corridor (cf. ﬁg. 3), because this

passage is most challenging due to tables and a chair

opposite to the door leaf. Besides, the meeting room

could not be recorded completely, leading to a signiﬁ-

cant Comp errors. At least Corr is quite competitive.

Figure 5: Simulation result of dynamic frontier-based ex-

ploration using scoring weights λ

= (20, 10,12,5, 8,50),

with occupied grid cells (red), NBVs (blue), trajectory

(green) and map orientation (cross hairs).

The test run for the ground plan-based explo-

ration is given in ﬁgure 6. Apart from some difﬁcul-

ties at the beginning (0, 1), the trajectory is a straight-

forward loop through the whole working space, close

to how a person would explore the environment. Ac-

cordingly,

= 99.4 m and 16 NBVs are minimal

over all test runs. Each scan took about 60 s, due to

the time needed for updating the polygonal maps, re-

sulting in a slight increase of T

Exp

= 37.5 min. This

yields 16 min for the scans, a minimum travel time

of 4 min 44 s and a time penalty of 17 min for slow-

downs. Comp and Corr are similar to the previous

run. However, meeting room and staircase area are

now almost completely covered.

Figure 7 reveals that the hybrid strategy performs

signiﬁcantly worse. Meeting room, kitchen and south

corridor are explored quickly (0-7). But then a couple

of oscillations occurs by moving back to the meet-

ing room (8, 9), to the staircase (10, 11) and to the

kitchen (12-14). Afterwards, the remainder of the

Figure 6: Simulation result of ground plan-based explo-

ration, with ground plan of ﬂoor (orange) and ceiling (blue).

8,9

3,12

Figure 7: Simulation result of hybrid exploration.

south corridor, entrance area and north hallway are

explored (15-22), before the meeting room is reen-

tered from the west (23). Apart from the oscillations,

the trajectory represents the same basic loop as in the

previous run. T

Exp

= 53.8 min is fairly poor and the

number of NBVs is signiﬁcantly increased to 24 (no

scan at vp 1, 4, 9, 14, 23, due to close obstacles). At

least

= 139.9 m is better than for the frontier

approach. This yields a minimum travel time of 6 min

40 s, 19 min for the scans and 28 min for slow-downs,

i. e. more than 50 %. As indicated by the accumula-

tion of NBVs, this time penalty is caused by narrow

passages in the meeting room and kitchen. However,

Comp and Corr are competitive, i. e. errors are only

slightly increased, and completeness is signiﬁcantly

better.

4.4 Real World Experiments

In the real ofﬁce scenario, the experiments are af-

fected by environmental dynamics beyond control,

e. g. actuated doors and people walking around. Be-

sides, the SLAM unit does not provide a common ref-

erence frame, i. e. the pose of the robot while record-

Experimental Evaluation of Some Indoor Exploration Strategies

285

ing the initial map deﬁnes its orientation. Therefore,

it is indicated by cross hairs for better comparability.

23, 24

14, 15

half-open doors

trajectory loop

Figure 8: Real world result of dynamic frontier-based ex-

ploration.

Figure 8 shows how the dynamic frontier-based

strategy guides the robot from the meeting room (vp

0-2) through the west door (c) into the corridor (3-

4), entrance area (5-9) and north hallway (10-12). On

the way back to the south corridor, it is attracted by

some half-open doors (13-15), resulting in a local tra-

jectory loop F, until the deadlock prevention inter-

venes. While exploring the remainder of the corridor

and the kitchen (16-20), small navigation problems

occur (H). Finally, the robot tries to get back to the

meeting room at NBVs 23, 24 via intermediate scans

(21, 22). Here, the planner searches for an alterna-

tive to door (c), as the loop to the north corridor N

indicates. Despite the challenges of the real world,

Exp

= 51 min and 25 NBVs are quite competitive, at

least compared to the hybrid exploration in simula-

tion. The average time needed for a scan and map up-

date was 45 s, which yields 13.5 min in total (no scan

at vp 2, 9, 13, 18, 19, 22 and 23).

= 208 m is

signiﬁcantly increased compared to the simulation re-

sults. This means a minimum travel time of 9 min 54 s

and a time penalty of about 27.5 min for slow-downs.

During the ground plan-based exploration,

shown in ﬁgure 9, the robot needs 5 NBVs (0, 1, 3-5)

plus a side trip to the kitchen (2), before it manages to

leave the meeting room via the west door. The accu-

mulation of NBVs 1, 4 and 3, 5 arises from the inabil-

ity to reach the original targets. Afterwards, entrance

hall (6-9), north (10-12) and south corridor (14, 15,

17-21) are explored quickly, only disturbed by an ad-

ditional scan (13) due to the adjacent half-open door

and an oscillation (16). After recording the rest of the

kitchen (22, 23), the map is rather complete. From

5,3

22,24,27

half-open doors

half-open door

door with

glass inset

Figure 9: Real world result of ground plan-based explo-

ration.

the following 9 vps, only the last one (32) provides

new information, whereas the intermediate oscilla-

tions reveal severe navigation problems due to envi-

ronmental dynamics. At least the strategy is robust

enough to prevent deadlocks and to terminate in ﬁnite

time. Correspondingly, 33 NBVs, T

Exp

= 94 min and

= 273.2 m are the worst of all runs. Each scan

took about 81 s, i. e. 44 min 33 s in total, the minimum

travel time is 13 min and the penalty for slow-downs

36.5 min.

19,24

half-open doors

half-open door

door with

glass inset

Figure 10: Real world result of hybrid exploration.

As shown in ﬁgure 10, the hybrid strategy man-

ages to explore the whole meeting room rapidly via

NBVs 0-3, followed by the entrance (4-6), north (8-

11) and south corridor (12-18), with only one interme-

diate scan (7) due to the adjacent half-open door. The

oscillations between door (c) (19, 23, 24), south cor-

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

286

ridor (20, 21) and entrance (22) are caused by the in-

ability to enter the kitchen via the west door (25). Fi-

nally, the robot drives back to door (a) and completes

the map (26-29). T

Exp

= 83.8min,

= 247.4m

and 30 NBVs represent a slightly better performance

than in the previous run. The total scan time is 33 min

45 s (no scan at vp 2, 16, 19, 25, 29), the minimum

travel time is 12 min and the time penalty for slow-

downs 38 min.

In sum, table 1 shows that the dynamic frontier-

based approach is always the fastest due to the time

needed to construct the ground plans for the other two

strategies. However, these two produce signiﬁcantly

more complete maps, especially regarding the meet-

ing room, both in simulation and reality. All three

procedures need notably more time, more NBVs and

a longer trajectory for the real world scenario. They

all manage to avoid deadlocks and to create a suit-

able map within a reasonable amount of time. All

in all, in simulation the ground plan-based approach

provides the best trade-off between needed time and

map quality, whereas in reality the hybrid approach is

most promising.

5 CONCLUSION

The paper has analyzed the performance of three

NBV exploration strategies, based on representative

test runs in a simulated ofﬁce scene and in its real

world counterpart. They all use the same weighted

sum of inﬂuence factors for scoring vps, online map

updates and adaptation of the exploration trajectory.

They differ in the kind of mapped features (2.5D

grid map vs. ground plan of ﬂoor and ceiling) and

corresponding estimation of information gain (fron-

tier length vs. area of difference polygons). The

dynamic frontier-based approach yields always the

fastest, but incomplete results, whereas the best trade-

off between needed time and map quality is provided

by the ground plan-based procedure in simulation and

by the hybrid strategy in reality.

Future work concentrates on a comparison with

state of the art approaches and on analyzing the in-

ﬂuence of the layout of the environment on the re-

sults, in order to check that the strategies work in gen-

eral. Other topics are to ﬁnd a generally optimal set

of scoring weights and to improve and evaluate the

room-wise exploration approach of (Schmidt, 2006)

in obstructed real world scenarios.

REFERENCES

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

ploration strategies for mobile robots. ICRA.

Amigoni, F., Caglioti, V., and Galtarossa, U. (2004). A mo-

bile robot mapping system with an information-based

exploration strategy. ICINCO.

Amigoni, F. and Gallo, A. (2005). A multi-objective explo-

ration strategy for mobile robots. ICRA.

Amigoni, F., Li, A. Q., and Holz, D. (2013). Evaluating the

impact of perception and decision timing on autono-

mous robotic exploration. ECMR.

Basilico, N. and Amigoni, F. (2011). Exploration strategies

based on multi-criteria decision making for search and

rescue autonomous robots. AAMAS.

Eliazar, A. and Parr, R. (2003). DP-SLAM: Fast, robust

simultaneous localization and mapping without pre-

determined landmarks. IJCAI.

Freda, L. and Oriolo, G. (2005). Frontier-based probabilis-

tic strategies for sensor-based exploration. ICRA.

Gonz

alez-Ba

nos, H., Mao, E., Latombe, J.-C., Murali, T.

M., and Efrat, A. Planning robot motion strategies for

efﬁcient model construction. ISRR.

Gonz

alez-Ba

nos, H. and Latombe, J.-C. (2002). Navigation

strategies for exploring indoor environments. IJRR

21(10-11).

Holz, D., Basilico, N., Amigoni, F., and Behnke, S. (2011).

A comparative evaluation of exploration strategies and

heuristics to improve them. ECMR.

Maffei, R., Jorge, V. A. M., Prestes, E., and Kolberg, M.

(2014). Integrated exploration using time-based po-

tential rails. ICRA.

Makarenko, A., Williams, S., Bourgault, F., and Durrant-

Whyte, H. (2002). An experiment in integrated explo-

ration. IROS.

Oriolo, G., Vendittelli, M., Freda, L., and Troso, G. (2004).

The SRT method: randomized strategies for explo-

ration. ICRA.

Proetzsch, M., Luksch, T., and Berns, K. (2010). Develop-

ment of complex robotic systems using the behavior-

based control architecture iB2C. RAS 58(1).

Quinlan, S. and Khatib, O. (1993). Elastic bands: connect-

ing path planning and control. ICRA.

Schmidt, D., Luksch, T., Wettach, J., and Berns, K. (2006).

Autonomous behavior-based exploration of ofﬁce en-

vironments. ICINCO.

Surmann, H., N

uchter, A., and Hertzberg, J. (2003). An

autonomous mobile robot with a 3D laser range ﬁnder

for 3D exploration and digitalization of indoor envi-

ronments. RAS 45(3-4).

Wettach, J. and Berns, K. (2010). Dynamic fron-

tier based exploration with a mobile indoor robot.

ISR/ROBOTIK.

Wettach, J. and Berns, K. (2012). Ground plan based explo-

ration with a mobile indoor robot. ROBOTIK.

Wettach, J. and Berns, K. (2014). Combining dynamic fron-

tier based and ground plan based exploration: a hybrid

approach. ISR/ROBOTIK.

Wurm, K. M., Stachniss, C., and Burgard, W. (2008). Coor-

dinated multi-robot exploration using a segmentation

of the environment. IROS.

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

mous exploration. CIRA.

Experimental Evaluation of Some Indoor Exploration Strategies

287