Spiralling Human-Inspired Exploration Algorithm with Doorway

Detection

Rasmus Borrisholt Schmidt, Andreas Sebastian Sørensen, Thor Beregaard and Michele Albano

Dept. of Computer Science, Aalborg Universitet, 9220 Aalborg Øst, Denmark

Keywords:

Minotaur, TNF, Online Exploration, Signal Degradation.

Abstract:

Exploration of unknown environments is an important task for autonomous robot swarm systems. The faster

they can fully explore an area, the faster a coordinated plan can be made, or points of interest found, to support

further tasks. Previous algorithms have often focused either on frontier based, or nature-inspired heuristics.

We present a human-inspired exploration algorithm, Minotaur, that enables simple and efﬁcient exploration

of buildings. We studied how Minotaur and a state-of-the-art algorithm, namely The Next Frontier (TNF),

perform. Minotaur follows walls to discover doorways, after which it coordinates with robots in the same room

to extend the exploration to rooms accessible through the discovered doorways. Most algorithms assume either

perfect communication, or line-of-sight (LOS) communication, which hinders the realism of the simulation

results. We then modiﬁed an existing simulator to take into account realistic communication technologies that

have limited penetration of materials through walls. Comparative experiments between Minotaur, TNF, and a

simple greedy algorithm show the superiority of Minotaur when multiple robots are exploring buildings-like

maps. However, when considering cave-like maps, Minotaur appears to have bad performance, but the greedy

algorithm outperforms TNF, particularly when the algorithms are limited in their communication capabilities.

1 INTRODUCTION

Robot exploration is a common problem in swarm

robotics, with applications in search & rescue, ﬁre-

ﬁghting, emergency response, and space and undersea

exploration (Queralta et al., 2020). Many exploration

algorithms have been proposed, some inspired by an-

imals (Nicola et al., 2018; Gul et al., 2021; Gul et al.,

2023b; Gul et al., 2023a), some using frontier explo-

ration (Burgard et al., 2005; Bautin et al., 2012; Co-

lares and Chaimowicz, 2016), and some using various

mathematical concepts (Mirjalili, 2016; Kennedy and

Eberhart, 1995).

This paper proposes Minotaur, a human-inspired

exploration algorithm. Each robot explores along the

nearest wall, at a little less than their maximum vi-

sion range, detecting doorways on the way. If there

are multiple robots in the current room, half of the

robots will enter the discovered doorway, and explore

from there, spreading throughout the map. When

an explored area is seen, robots follow the explored

area similarly to the initial wall-following strategy.

From time to time, robots communicate their maps

https://orcid.org/0000-0002-3777-9981

and marked doorways to limit redundant movements.

Inter-robot communication is useful to share the

Simultaneous Localization and Mapping (SLAM)

maps that the robots are building during exploration

and robots current paths, and in general to strategize

further exploration tasks. Most simulation solutions

assume perfect communication between robots. This

paper lays foundation for more realistic simulation

by implementing a novel communication mechanism

that takes into consideration signal loss through ma-

terials in an existing simulator, and uses it to analyze

the performance of Minotaur against a state-of-the-

art exploration algorithm, namely The Next Frontier

(TNF) (Colares and Chaimowicz, 2016) algorithm.

The rest of this paper describes related works,

namely simulators and exploration algorithms (Sec-

tion 2), presents our proposed algorithm (Section 3)

and discusses its implementation and the mechanisms

used to implement realistic communication

( Sec-

tion 4). Experimental results are described in Sec-

tion 5, followed by our conclusions on the topic at

hand and suggestions for future work (Section 6).

The developed code is released as open-source on

https://github.com/DEIS-Tools/MAES. A demo video can

be seen on https://youtu.be/gjnVm46Ezd0.

112

Schmidt, R. B., Sørensen, A. S., Beregaard, T. and Albano, M.

Spiralling Human-Inspired Exploration Algorithm with Doorway Detection.

DOI: 10.5220/0013237700003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 1, pages 112-122

ISBN: 978-989-758-737-5; ISSN: 2184-433X

2 RELATED WORK

This section discusses simulators for exploration al-

gorithms, and describes existing exploration algo-

rithms relevant for this work.

2.1 Exploration Simulators

When simulating robots, realism is paid in terms of

complexity and performance. In fact, on the one

hand many works used simple simulators that con-

sider robots moving on a grid, disregard collisions,

and allow for unrestrained communication (see for

example (Cheraghi et al., 2020; ?)). On the other

side of the spectrum there are high ﬁdelity simulators

that are very complex to set up and have bad perfor-

mance (Pitonakova et al., 2018).

We opted for Multi Agent Exploration Simulator

(MAES) (Andreasen et al., 2023), which provides a

trade-off between realism and ease of usage. MAES is

open source, it was used to compare exploration algo-

rithms (Andreasen. et al., 2022) and provides the code

to simulate them. Further, MAES supports the cre-

ation of transmission models dependent on distance

and walls passed through, such as attenuation based

models, and can generate maps that represent either

cave-like or building-like environments.

2.2 The next Frontier (TNF)

TNF (Colares and Chaimowicz, 2016) is an explo-

ration algorithm designed to use information gain and

distance cost as a basis for autonomous swarm explo-

ration. TNF uses the Near-Frontier Exploration (?),

which identiﬁes frontiers in the map, which are loca-

tions that are in-between explored and non-explored

areas. In the ﬁtness function that decides robots’ way-

points, TNF considers the information gain that going

to a given area would provide and the cost to get there.

Moreover, each robot stores the newest destina-

tion received from each other robot. When the robot

decides upon the next frontier to explore, a wavefront

will be created from each stored destination, reduc-

ing the score of each frontier near it. This means that

frontiers close to the target frontiers of other robots

will be lower in the score, causing the robots to spread

out more in the map, exploring more disparate sec-

tions and reducing redundant work, speeding up the

exploration.

2.3 Spiraling and Selective

Backtracking (SSB)

SSB (Gautam et al., 2018) is a coverage algorithm

that uses spiraling patterns and expects global com-

munication and discrete grid-based movement. The

robot spirals through known areas and updates other

robots of the covered space and the nearby uncovered

tiles it reserves for backtracking.

The backtracking is used to ﬁnd new uncovered

areas, and it uses an auction mechanism to decide

which robot should go to each backtracking point. If

no backtracking point is available, or if the robot did

not win any auction for a backtracking point, an auc-

tion is instead done for the nearest unvisited point. As

the robot spirals and creates backtracking points, it re-

serves the points of the spiral and backtracking, and

broadcasts them to other robots, limiting their move-

ment in an attempt to reduce redundant movement.

Our proposed solution uses concepts from SSB,

while improving it in terms of robots’ coordination

and resilience to communication impairments.

2.4 Human Heuristics

Most previous algorithms are generally focused on

mathematical formulas, or get inspired by animal be-

havior. An alternative approach is to explore in a way

that makes sense to humans.

Human heuristics are used in various tasks in

robotics (?; ?). An example of human heuristics used

for exploration is in (Liu et al., 2023), where the au-

thors utilize computer vision with human heuristics to

detect where doors are to separate rooms apart. Fur-

thermore, it prioritizes ﬁnishing exploring the current

room before proceeding to the next one.

3 THE MINOTAUR

EXPLORATION ALGORITHM

The general idea behind our proposed exploration al-

gorithm is to use two human-inspired heuristics: (i)

robots will spiral inside a room to explore it; (ii) when

a robot identiﬁes a door in a room, half of the robots

in the room will move out through the door to cover

as many rooms as possible. Part of this approach is

inspired by SSB (Gautam et al., 2018), particularly

the spiraling movement and the auction mechanism

to decide which robot goes to important locations, in

our case doors.

To explain the algorithm, we will initially go over

a constrained version for one single robot, then ex-

pand it to show the full multi robot behavior, with

Spiralling Human-Inspired Exploration Algorithm with Doorway Detection

113

smaller black-box around doorways functionalities

that are then described in their own sections, all to-

gether creating the full algorithm.

3.1 Single Robot

In the single robot algorithm, whose pseudocode is

shown in Alg. 1, the robot starts by exploring along

the wall nearest to its initial location, moving in a

straight direction looking for a wall if it cannot see

any. After this, the robot follows the wall, identifying

(see Section 3.3) and marking the doorways it ﬁnds

without entering them until current room is fully ex-

plored.

After covering all walls in a room, it will explore

the center of the room in a spiral pattern, until the

whole room is explored. After that, the robot will

enter the nearest unexplored doorway in the room,

and continue the exploration from there, returning to

unexplored doorways in previous rooms if no unex-

plored door can be found in the new room.

Data: Maximum range of robot sensors

while No walls or explored areas are in range do

Move forward ;

end

while Map is not explored do

while Unexplored area is reachable without

passing doorways do

if Doorway found (Alg. 3) then

Store the doorways, stay in the room

and continue exploring along

nearest wall ;

end

if Wall or explored area in visual range

&& unexplored area ahead then

Follow wall or explored area at max

visual range counter-clockwise ;

else

Go to the nearest unexplored area

without passing a doorway ;

end

if Unexplored doorway accessible without

passing other doorways then

Move through the nearest unexplored

doorway within room ;

else

Move through the nearest unexplored

doorway ;

end

Algorithm 1: Minotaur Algorithm (Single Robot) .

Data: Maximum range of robot sensors

while No walls or explored areas are in range do

Each robot moves forward ;

end

while Map not explored do

Each robot broadcasts map, including walls,

explored area, doorways and known robots

location ;

For half of robots in current room spin ←

counter-clockwise, for the rest spin ←

clockwise ;

while Unexplored area is reachable without

passing doorways do

if Doorway found (Alg. 3) then

Store & handle doorway (Alg.4) ;

end

if Wall or explored area in visual range

&& unexplored area ahead then

Follow wall or explored area at max

visual range in spin direction ;

else

Go to the nearest unexplored area

without passing a doorway ;

end

if Unexplored doorway accessible without

passing other doorways then

Move through the nearest unexplored

doorway within room ;

else

Move through the nearest unexplored

doorway ;

end

Algorithm 2: Minotaur Algorithm.

3.2 Multiple Robots

In the strategy for multiple robots, see Alg. 2, half of

the robots explore around the rooms in one direction,

the rest proceed in the other direction, to identify all

doorways quickly.

As they move, the robots broadcast their SLAM

map and any found doorways for other robots to add

to their own maps. On ﬁnding a doorway, the robot

starts an auction with all the robots in the same room,

which respond with a bid equal to their distance to

that doorway, so long as they can get there without

passing another doorway, i.e. they are in the same

room. This ensures that no robot spends time mov-

ing to a doorway far away, and that all room that have

started being explored also get ﬁnished. This is fur-

ther described in Section 3.4.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

114

3.3 Doorway Detection

When a robot has one or multiple walls in their line of

sight, it will check if it is possible that it discovered a

door, with a set of conditions that are different for one

or multiple walls, as described in Alg. 3.

If there is a single wall, then when the robot ﬁnds

a gap in the wall, it will continue along the wall as far

as the door width parameter, and mark the gap as a

door if it ﬁnds the continuation of the wall. In Fig. 1,

in the ﬁrst case there is a door in the middle of a wall;

in the second one the door is located in the room’s

corner; in the third case there is no door in a corner;

the fourth case represents a hallway, i.e. the distance

between the interruptions of the wall is larger than the

maximum door width.

If there are multiple walls, then the robot will cre-

ate intersection points between the walls as inﬁnite

lines. If these intersection points have been seen, are

non-solid, and close enough to make a doorway, then

it is a door (last case in Fig. 1).

Data: Maximum range of robot sensors

Data: Maximum width of doors

if Wall ends/stops then

move forward doorway width ;

if Wall continues after doorway width then

Mark doorway and direction it is seen

from ;

end

if Two walls in vision area then

Store the wall points;

Continue vision range forward;

Store the new wall points;

Create lines along wall direction between the

points;

if Lines intersect && distance between wall

points greater than robot size then

Find intersection point of lines;

Move towards intersection point until it is

explored;

Create line segments between wall points

and intersection point;

if Ensemble of line segments is open then

Mark doorway and direction doorway

it is seen from;

end

if any doorways ”seen” from both sides or

doorway was passed through then

Mark doorway as explored;

end

Algorithm 3: Doorway detection, as function of door max-

imum width.

Communicate doorway location to all robots

(comprising itself);

if Receiving robot can reach the doorway on the

same side without passing any doorway then

Receiving robots give ACK and distance;

end

out ← half of received ACK (rounded down);

Command the out robots to exit the doorway;

The rest of the robots will stay in the room ;

Algorithm 4: Doorway auction.

3.4 Doorway Auction

When a robot, let us call it Alice, ﬁnds a door, it starts

an auction by sending out a ”doorway found” mes-

sage. When a robot in communication range receives

the message, comprising Alice, if it is able to reach

that door without passing through another doorway,

i.e. it is in the same room, it will respond with an

ACK and a bid equal to its distance to that door. Al-

ice too will respond to the auction with an ACK and

its own bid. Afterward, Alice tells to the half of the re-

spondents having the lowest bids to proceed through

that door (Alice will probably be in this set). The rest

of the robots will stay in the room to ﬁnalize its ex-

ploration. This is formalized in Alg. 4.

The approach is also visually explained on Fig. 2,

where green areas are unexplored, the colored dots are

robots, and the dashed lines are communication lines,

with their color signifying whether they respond with

an ACK (green) or not (red). On Fig. 2, Pink ﬁnds a

door and communicates to the other robots in range.

Cyan will be able to go to the room. Yellow cannot

get to the door without passing another doorway, and

therefore will not reply with an ACK. While Red tech-

nically can get to the door, it does not know this, as the

area in-between is unexplored. From its perspective,

it would have to backtrack through multiple doorways

to reach Pink, and so it will not reply with an ACK.

Figure 1: Base cases for the doorways.

Spiralling Human-Inspired Exploration Algorithm with Doorway Detection

115

Figure 2: Communication scenario for Alg. 4.

3.5 Greed Algorithm

When the Minotaur has no nearby unseen areas, or

unexplored doorways to go through, it will move to

the nearest unexplored tile anywhere on the map and

continue exploration from there.

This feature has been extracted into its own al-

gorithm, creating a greedy algorithm called Greed,

which always moves to the nearest unexplored tile,

and still communicates and shares its map with other

robots.

Compared to Minotaur, Greed has the advantage

of being simpler and not needing to follow walls. On

the other hand, Minotaur can better divide the robots

between the rooms when it identiﬁes doors.

4 DESIGN & IMPLEMENTATION

This section describes the support for realistic com-

munication through walls in the selected simulator

(MAES (Andreasen. et al., 2022)), and the spiral-

ing and doorway detection mechanisms in Minotaur,

since its implementation was surprisingly more com-

plex than expected.

4.1 Materials in MAES

The MAES simulator allowed to simulate either

global (unrestrained) and line-of-sight (LOS) only

communication between robots. However, we

deemed both of them to be not realistic enough and

we decided to add to each section of the walls a Ma-

terial property, and then use it to check if two robots

are able to communicate with each other at a given

time.

The physics of this problem has already been well

studied, and multiple signal attenuation tables can be

found for various materials at various frequencies, see

for example (Ikpehai et al., 2019; Rappaport, 1996;

?).

Signal loss can be described using the general

path-loss algorithm, see for example Eq. 1, where α,

β, and γ are empirically found ”magic” numbers that

describe how much the signal is degraded for line of

sight and non line of sight communication (?).

(d, f ) = 10αlog

(d) + β + 10γ log

( f ) (1)

In terms of implementation, we modiﬁed the mesh

generation used to generate a map to add a Material

property, which contains the α, β, and γ from Eq. 1, to

each triangular section that makes up a wall in MAES.

In terms of the Unity Engine, this was done using sub-

meshes, which are subsets of a larger mesh and can

contain separate materials, textures and animations.

This was preferable in comparison to splitting the dif-

ferent meshes into separate game objects for perfor-

mance sake, as it allows for the Unity renderer to per-

form draw call batching more efﬁciently.

Whenever MAES checks if two robots are able to

communicate, it now draws a line between the robots,

it splits it into segments each time the material be-

tween the robots change, and it calls a function with

a list of tuples (Material, length), with one element in

the list for each segment. The function will then ap-

ply Eq. 1 for each element in the list to compute how

much the signal was attenuated. Equation 2 formal-

izes this process, where S

is true when the transmis-

sion can get received successfully for robot i, R

is the

receiver sensitivity of robot i, T

is the transmission

strength of robot j, m is the number of walls (or open

spaces) that have been crossed with a message, and

n, f

are the attenuation values.

= R

≤ (T

−

∑

n=1

n, f

) (2)

The attenuation values for the different materials

used in a map can be provided by the end user as a

part of the conﬁguration along with the transmission

strength, frequency of the signal, and receiver sensi-

tivity of the robots. MAES provides the default val-

ues in Table 1 and, if the end user does not specify

their own, the transmission strength is set to 15 dBm,

the receiver sensitivity is set to -82 dBm, and the fre-

quency is set to 2.4 GHz. These values were chosen in

accordance to 802.11ax also known as Wi-Fi 6 (IEEE,

2021).

4.2 Spiraling Implementation

This subsection focuses on the spiraling strategies that

the robots follow while exploring a room. The imple-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

116

Table 1: Default values for signal attenuation at different

frequencies and materials (Ikpehai et al., 2019).

Frequency 1300 MHz 2400 MHz 5200MHz

Air 0 dB 0 dB 0 dB

Concrete 13 dB 15 dB 23 dB

Wood 5.1 dB 6.7 dB 14 dB

Brick 4.5 dB 5.5 dB 15 dB

mentation of spiraling relies on the overlaid grids on

the SLAM map that the robot creates while navigat-

ing. Each tile on this grid can have one of three states,

Open, Unseen and Solid, as can be seen in Fig. 3.

While following the wall on the initial exploration

of a room, a robot looks for solid tiles within its vi-

sion range. The robot decides which wall to follow by

checking if there are any unseen tiles in front of it. If

there are none, the robot is done following that wall.

When the robot has multiple candidate walls, it

prioritizes moving in the same direction that it is al-

ready moving. The robot computes the perpendicu-

lar vector from the furthest points it can see on the

walls, to decide where to move to next, as illustrated

on Fig. 4. If the wall ends or turns away from the

robot without having a doorway, the robot will ini-

tially lose track of it, but then follow it around the

bend. When reaching the end of a conﬁned area, like

a dead-end in a hallway, the robot will return to the

last seen area with unexplored tiles and continue the

spiraling pattern from there.

On completing the ﬁrst lap around the room, the

robot will start to spiral around the edge of the previ-

ously explored area. This is tricky, since opposite to

walls, edges between explored and unexplored areas

disappear when unexplored area in question gets ex-

plored. Therefore, the robot must use the fact that the

explored area is continuously fed into its SLAM map,

to look ahead beyond its own vision range and follow

the edge there.

If the spiraling movement cannot reach unseen

tiles, the robot will move greedily towards the near-

est unseen tile in the room. When a room is fully

explored, its robots enter the nearest unexplored door-

ways and continue the exploration.

4.3 Doorway Detection Implementation

While the robot is moving to any destination, it checks

for doorways. For a doorway with a single wall, as

Figure 3: Representations of the different tile statuses.

Figure 4: A valid wall due to having an unseen area where

there presumably is a wall with edges and movement vector

based upon those edges.

seen on Fig. 5, the robot sees that the next tile of the

wall is an open tile. When this happens, the robot

will move forward for a distance equal to the doorway

width parameter. The robot will then be standing at

the end of the movement vector and will be able to see

if the wall continues, and if so, creates an intersection

point between the walls. Then it checks if that is a

valid door, according to the parameter maximum door

width.

Figure 5: Doorway detection when seeing a single wall.

Spiralling Human-Inspired Exploration Algorithm with Doorway Detection

117

Figure 6: Doorway detection when seeing multiple walls. The left shows the state on detection of a potential door, and the

right shows the state after moving to check intersection points.

If there are two or more walls, the robot looks for

an intersection point between those, as can be seen

on Fig. 6. When the intersection point is in an un-

seen area, as it is likely to happen, the robot will go

to explore the intersection point. After the robot has

moved to see the point, it takes the closest points to

the intersection from the two walls that created it, as

can be seen on the two edge points on the right state of

Fig. 6. With those two points, the robot checks if they

create a valid doorway, by once again considering the

doorway maximum width parameter.

5 EXPERIMENTS AND RESULTS

In this section, we go through the experimental setup

and the results of our experiments. These experiments

are meant to understand the effect of more realistic

communication on the exploration algorithms, and to

compare the performance of Minotaur with TNF, as

well as Greed (see Section 3.5).

5.1 Experimental Setup

To perform a fair comparison of the algorithms and

the simulation settings, we initialized a random num-

ber generator with the seed 123456, then we gener-

ated 100 random building maps and 100 random cave

maps, we selected random initial locations for the

robots, and later we performed all simulations with

those same initial conditions. An experiment was

considered complete when 98% of the area was ex-

plored. On all the graphs in Section 5.2, we present

the percentage of map that was explored at a given

time, averaged over the 100 maps.

The simulation campaign we performed consid-

ered a map of 50x50, 100x100 or 150x150 tiles, 1,

2, 4, 8 or 16 robots, robots spawned either close to-

gether or at random over the whole map, maps were

either representing caves or buildings (see Section 3.1

in (Andreasen et al., 2023)), the algorithm was ei-

ther TNF, Minotaur or Greed, and the communication

technology had either global coverage, line-of-sight

limitation or we simulated communication through

Table 2: Settings for the exploration scenarios.

Robot random seed 123456

Maps size 50x50, 100x100,

150x150

Number of robots 1, 2, 4, 8, 16

Initial position of robots Close together, random

Map type cave, building

Exploration algorithm TNF, Minotaur, Greed

Communication Global, LOS, materials

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

118

Figure 7: On the left, comparison of communication technology for the Greed algorithm, 16 robots, 150x150 building maps.

On the right, comparison of algorithms for 2 robots in 150x150 tiles building maps.

Figure 8: Comparison of algorithms with 16 robots on building maps (size is 50x50 tiles on the left, 150x150 tiles on the

right).

materials as presented in Section 4.1. Table 2 pro-

vides a summary of the parameters used for the simu-

lation campaign.

5.2 Results

We have compared the performance of the exploration

algorithms when different communication capabili-

ties were considered. In all cases, as expected, per-

formance with LOS limitations was worse than with

global communication, with communication through

materials being in the middle. The performance gap

was more pronounced with larger maps and when

more robots were part of the picture. However, the ef-

fect was always quite limited. The left graph in Fig. 7

shows one of the most pronounced effects, for exper-

iments with the Greed algorithm on large (150x150

tiles) building maps, 16 robots spawned close to each

other. The rest of the results we are presenting will

consider communication through materials.

One of our hypothesis was that Minotaur requires

a sizable number of robots to beneﬁt from its capac-

ity to coordinate its robots efﬁciently. Results from 2

robots exploration conﬁrmed it, but also hinted that in

large building maps (150x150 tiles) both Greed and

Minotaur outperform TNF, with Greed being slightly

better than Minotaur. The right graph in Fig. 7 shows

one example of this, for 2 robots over 150x150 tiles

building maps.

When many robots are part of the picture, we ex-

pected that Minotaur had an edge over the other algo-

rithms, since it pushes robots to cover the most rooms

as fast as possible. The results conﬁrmed it, even

though the performance of the Greed algorithm was

quite close to Minotaur’s. Fig. 8 reports on a com-

parison of the algorithms using 16 robots on building

maps of different sizes (50x50 tiles vs 150x150 tiles).

Finally, we decided to compare the algorithms on

cave maps. The performance of Minotaur was very

bad with respects to the other algorithms. This is

explainable by the fact that Minotaur depends on its

mechanism to recognize doors as opening on straight

walls. However, cave maps do not have straight walls,

invalidating the beneﬁts of Minotaur, which still suf-

fers from overheads. Fig. 9 reports on a compari-

son of the algorithms on cave maps, with 2 robots on

50x50 tiles maps on the left and 16 robots on 150x150

maps on the right, where TNF outperforms Greed on

Spiralling Human-Inspired Exploration Algorithm with Doorway Detection

119

Figure 9: Comparison of algorithms with on cave maps (2 robots on 50x50 tiles on the left, 16 robots on 150x150 tiles on the

right).

smaller maps with less robots, while Greed takes the

lead on larger maps with more robots.

After the collection of the results, some thoughts

were given to the causes of Minotaur slowdowns, for

example on cave maps; and why on building maps

Greed was often able to compete with Minotaur while

both outperformed TNF. When a corner is encoun-

tered on a map, see Fig. 10, Greed approaches the

corner, then it moves out looking for unexplored area.

Minotaur instead moves closer to the corner and,

when it sees the opposing wall of the corner, needs

to check if there is a doorway. This causes it to move

much closer to the wall than Greed.

A different situation is represented in Fig. 11,

where Greed has a more direct movement through the

door, while Minotaur spends time more to identify it,

and then it decide to disregard the doorway and to

continue the exploration of the room. These wasted

movements all cause slowdown, providing Greed a

relative advantage.

6 CONCLUSIONS AND FUTURE

WORK

This paper presents an efﬁcient exploration algorithm

based on human-inspired heuristics. The algorithm

uses the simple ideas of wall following, fully explor-

ing rooms before leaving, looking for doorways, and

splitting a group of robots as much as possible to con-

tinue the exploration of the current room while still

crossing the newly found doors.

Comparisons have been made with The Next

Frontier (TNF), and a greedy algorithm called Greed.

When maps of buildings are explored, the more robots

are added, the better Minotaur performes compared

to Greed, and both algorithms perform far better than

TNF. In case of maps of caves, Minotaur has bad per-

formance since it loses its capability of door identiﬁ-

Figure 10: Minotaur (top) and Greed (bottom) exploring a

corner without a doorway.

cation while still suffering from the overheads to exe-

cute its mechanisms.

To perform a more realistic comparison, we also

enhanced the simulator MAES to support the associa-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

120

Figure 11: Minotaur (top) and Greed (bottom) exploring a

corner featuring a doorway.

tion of materials to walls, which then are used to com-

pute signal loss for communication between robots.

Future work on the topic can comprise the deﬁ-

nition of an alternative door identiﬁcation algorithm

that can work also on cave maps.

More far fetched improvements can consider the

extension of the Minotaur algorithm to the 3rd dimen-

sion. This would allow the algorithms to be used with

aerial drones, mapping out far more complex spaces.

This would likely require a different understanding of

the SLAM map.

While algorithms can show promising results in

the MAES simulator, implementing them on real

robots, and testing them on areas created in the real

world, would likely present unique challenges and

different results.

ACKNOWLEDGEMENTS

This work was partially supported by the S4OS Vil-

lum Investigator Grant (37819) from Villum Fonden,

and by the Validation WP of the ERC Advanced Grant

LASSO (Learning, Analysis, SynthesiS and Opti-

mization of Cyber-Physical Systems).

REFERENCES

Andreasen., M. Z., Holler., P. I., Jensen., M. K., and Al-

bano., M. (2022). Comparison of online exploration

and coverage algorithms in continuous space. In

Proceedings of the 14th International Conference on

Agents and Artiﬁcial Intelligence - Volume 1: SDMIS,

pages 527–537. INSTICC, SciTePress.

Andreasen, M. Z., Holler, P. I., Jensen, M. K., and Albano,

M. (2023). MAES: a ROS 2-compatible simulation

tool for exploration and coverage algorithms. Artif.

Life Robotics, 28(4):757–770.

Bautin, A., Simonin, O., and Charpillet, F. (2012). Minpos:

A novel frontier allocation algorithm for multi-robot

exploration. In Intelligent Robotics and Applications:

5th International Conference, ICIRA 2012, Montreal,

Canada, October 3-5, 2012, Proceedings, Part II 5,

pages 496–508. Springer.

Burgard, W., Moors, M., Stachniss, C., and Schneider, F.

(2005). Coordinated multi-robot exploration. IEEE

Transactions on Robotics, 21(3):376–386.

Cheraghi, A. R., Abdelgalil, A., and Grafﬁ, K. (2020). Uni-

versal 2-dimensional terrain marking for autonomous

robot swarms. In 2020 5th Asia-Paciﬁc Conference

on Intelligent Robot Systems (ACIRS), pages 24–32.

IEEE.

Colares, R. G. and Chaimowicz, L. (2016). The next fron-

tier: combining information gain and distance cost for

decentralized multi-robot exploration. In Ossowski,

S., editor, Proceedings of the 31st Annual ACM Sym-

posium on Applied Computing, Pisa, Italy, April 4-8,

2016, pages 268–274. ACM.

Gautam, A., Richhariya, A., Shekhawat, V. S., and Mohan,

S. (2018). Experimental evaluation of multi-robot on-

line terrain coverage approach. In 2018 IEEE Interna-

tional Conference on Robotics and Biomimetics (RO-

BIO), pages 1183–1189.

Gul, F., Mir, I., and Mir, S. (2023a). Aquila optimizer with

parallel computing strategy for efﬁcient environment

exploration. Journal of Ambient Intelligence and Hu-

manized Computing, 14(4):4175–4190.

Gul, F., Mir, I., Mir, S., and Abualigah, L. (2023b). Multi-

agent robotics system with whale optimizer as a multi-

objective problem. Journal of Ambient Intelligence

and Humanized Computing, 14(7):9637–9649.

Gul, F., Mir, I., Rahiman, W., and Islam, T. U. (2021). Novel

implementation of multi-robot space exploration uti-

lizing coordinated multi-robot exploration and fre-

quency modiﬁed whale optimization algorithm. IEEE

Access, 9:22774–22787.

Spiralling Human-Inspired Exploration Algorithm with Doorway Detection

121

IEEE (2021). Ieee standard for information technology–

telecommunications and information exchange be-

tween systems local and metropolitan area networks.

IEEE Std 802.11ax-2021 (Amendment to IEEE Std

802.11-2020), pages 1–767.

Ikpehai, A., Adebisi, B., Rabie, K. M., Anoh, K.

O. O., Ande, R., Hammoudeh, M., Gacanin, H., and

Mbanaso, U. M. (2019). Low-power wide area net-

work technologies for internet-of-things: A compara-

tive review. IEEE Internet Things J., 6(2):2225–2240.

Kennedy, J. and Eberhart, R. (1995). Particle swarm opti-

mization. In Proceedings of ICNN’95 - International

Conference on Neural Networks, volume 4, pages

1942–1948 vol.4.

Liu, J., Lv, Y., Yuan, Y., Chi, W., Chen, G., and Sun, L.

(2023). An efﬁcient robot exploration method based

on heuristics biased sampling. IEEE Transactions on

Industrial Electronics, 70(7):7102–7112.

Mirjalili, S. (2016). Sca: a sine cosine algorithm for solving

optimization problems. Knowledge-based systems,

96:120–133.

Nicola, G., Pedrocchi, N., Mutti, S., Magnoni, P., and

Beschi, M. (2018). Optimal task positioning in multi-

robot cells, using nested meta-heuristic swarm algo-

rithms. Procedia CIRP, 72:386–391.

Pitonakova, L., Giuliani, M., Pipe, A., and Winﬁeld, A.

(2018). Feature and performance comparison of the

v-rep, gazebo and argos robot simulators. In Towards

Autonomous Robotic Systems: 19th Annual Confer-

ence, TAROS 2018, Bristol, UK July 25-27, 2018, Pro-

ceedings 19, pages 357–368. Springer.

Queralta, J. P., Taipalmaa, J., Can Pullinen, B., Sarker,

V. K., Nguyen Gia, T., Tenhunen, H., Gabbouj, M.,

Raitoharju, J., and Westerlund, T. (2020). Collabo-

rative multi-robot search and rescue: Planning, coor-

dination, perception, and active vision. IEEE access,

8:191617–191643.

Rappaport, T. S. (1996). Wireless communications - princi-

ples and practice. Prentice Hall.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

122