Ubiquitous Computing and Communication Journal

ISSN 1992-8424

AUTONOMOUS MULTI-ROBOT BEHAVIOURS FOR SAFETY

INSPECTION UNDER THE CONSTRAINTS OF UNDERGROUND
MINE TERRAINS

Chika Yinka-Banjo, Antoine Bagula

ISAT Laboratory, Department of Computer Science, Faculty of Science, University of Cape Town, Cape Town,
South Africa([email protected]; [email protected])
Isaac O. Osunmakinde
Semantic Computing Group, School of Computing, College of Science, Engineering and Technology,
University of South Africa, P.O. Box 392, UNISA 0003, Pretoria, South Africa. ([email protected])

ABSTRACT
The wide adoption of underground mining has contributed immensely to the rapid
growth of the countries economy, job creation, gross domestic product, fixed capital
formation, gross investment and merchandise exports on an enormous global scale.
At the same time, the tedious and dangerous nature of underground mining makes it
vulnerable to mine accidents and disasters, such as the fall of loose rocks, trapping of
miners and miners death. Building upon a comprehensive survey of safety
inspection issues in South Africa, this paper proposes an autonomous multi-robot
cooperative behavioural model that can help to guide multi-robots in pre-entry safety
inspection of the underground mines. We discuss in detail the different safety fears
in underground mines and categorise these issues into mine-related and roboticrelated disasters, the latter being classified according to random component failures,
human errors, software failures and systematic hardware faults. We present an
empirical analysis of how our proposed model can be used or implemented through
two real-life scenarios: a) the inspection of rock falls and b) the detection of gas
levels in underground mines. This survey can be used as a reference guide to
understand safety management and cooperative behavioural models for underground
mines and to conduct further research on the existing models to make them more
efficient, reliable and scalable, which can promote their use in larger mines and
applications.
Keywords: Safety inspection, underground mine, multi-robot system (MRS), cooperative
behaviour
1

INTRODUCTION

Mining industries are significant economic sector

for many countries, including the Republic of South
Africa, and they incorporate the use of coal, metal and
non-metal minerals. The usefulness of the minerals
extracted from mining cannot be overemphasised: they
are used in electrical generation, production of steel,
commercial and residential building products, and
asphalt, and in medicine, household, electronic and
other manufactured products [1].
Over the past 130 years, the mining sector in South
Africa has contributed immensely to the development
of various industries through the use of mining
products [31]. World-class industries such as energy,
financial services, water services, engineering

Volume 7 Number 5

services,
specialist
seismic,
geological
and
metallurgical services, etc., owe their existence to the
mining sector [31]. In 2009 South Africas mining
industry contributed about 19% of gross domestic
product , over 50% of merchandise exports, close to
1million jobs, approximately 18% of gross investment
[31] etc. Fig. 1 from South Africas Mineral Industry
report (SAMI)1, displays mining contributions
from1999 2008. It shows that mining is a huge
source of revenue to South Africa.

Data in all figures and tables, unless otherwise stated, refer to

South African mines and sources are: MHSI, MHSC, DME, SAMI,
F&F Annual Reports and some online resources from SAHO and
NUM

Page 1316

www.ubicc.org

Ubiquitous Computing and Communication Journal

Figure 1: Contributions of Mining and Quarrying to

Gross Domestic Product, Fixed Capital Formation of
South Africa, and Total Exports from 1999 2008
Underground mining remains the most hazardous
occupation in the world because of its disaster rate.
Over the past five decades, South Africa has recorded
a high number of deaths in underground mining
accidents. The worlds fifth worst underground mining
disaster, in which 437 miners died, happened in South
Africa at Coalbrook2 mine in 1960. Several disasters
have happened since then. Between 1983 and 1987,
3000 miners died and 5000 sustained permanent
disabilities [2] in South African mines. In recent times,
there have been a large number of disasters in South
African underground mines and this is a major
motivation for our research work.
In this context it is worth noting that after daily
mining activities in the mines, miners currently do not
know the state of rocks, whether very weak, partially
weak, or stable for the next mining activities and are
consequently ignorant of the potential for rock falls.
Similarly, old and abandoned mines are very
dangerous to anyone who attempts to explore them
without proper knowledge and safety training 3. These
mines often contain deadly gases, standing water that
may hide deep pits, weak rocks and cracked roofs.
Miners are often not aware of the state of the mine
before they enter. The most important question in this
context is how miners can be aware of the situation in
mines before they engage in operations. Can miners
know the state of rock formations, whether very weak,
partially weak, or stable for mining activities and
consequently gauge the potential of rock falls?
Understanding these concerns through robotic
automation of multi-robot cooperative behaviour will
help to prevent disasters in underground mines.
Mining companies and researchers have been
pursuing the goal of mitigating some of these mine
disasters. A holistic five-point4 ground management
strategy to address the incidence of accidents and
fatalities related to fall of ground was initiated in 2002
by a local South African mining company. This has
been implemented in an attempt to deal with adverse
2

http://worldnews.about.com/od/disasters/tp/Worlds-Worst-MiningDisasters.htm
3
http://en.wikipedia.org/wiki/Mining

ISSN 1992-8424

rock conditions by looking at mine design and mine

layout. Numerical modelling systems were used in
considering the size of the support pillar, stope
dimension and extraction sequence. These measures
are taken to reduce damage to the rock during mining
because mining involves fracturing rock structures or
working with already fractured rock. In 2003 the mine
design and layout phase conditions were controlled.
All types of mine support standards were replaced by
electronic-based systems, which could be incorporated
into computer aided design plans and made accessible
by all mines. Molina et al. contributed to safety in
underground mine by using wireless networks as
monitoring systems for the detection of dangerous
gases and collapses [23]. Their system is used for early
warning signs to prevent injuries and significant
economic losses. This contribution is the preparedness
phase in the prevention of disaster, as shown in Fig. 4.
On the other hand, Stormont et al. proposed using
autonomous robots to manage risk in disaster
scenarios. One of the scenarios presented in this paper
is the robots deployed for rescue at the World
Trade Centre disaster site in 2001. The authors also
discussed the semi-autonomous exploration rover that
explored the planet Mars and the computer model used
to investigate issues of trust and the impact of
reliability in a fire-fighting scenario [22]. In the firefighting scenario, the robots are called upon for
assistance when the human fire-fighters are not
making enough progress putting out the fires. The
model was simulated and created using a Net Logo
agent modelling language. There has been also ongoing research focusing on reducing rock fall and rock
burst accidents in underground mines since 2006. A
local South African mining company is co-sponsoring
a project with a South African research institute called
Integrated Damage Rheology Model (IDRM), which is
focusing on numerical modelling of mining and
seismic data [30].
It is commonly said that prevention is better than
cure. Rather than spend millions of US dollars on
deploying a solution in a disaster rescue, it is better to
spend that money on prevention (safety).

Figure 2: The four Phases of a Disaster

According to the U.S. Federal Emergency
Management Agency (FEMA), disaster has four
phases, as depicted in Fig. 2, which are associated by
the public with immediate life-saving or mitigation
efforts, leading to disaster response or disaster
management5. We are using this research to contribute
to the first phase of a disaster, which is prevention.
This will be achieved by sending interacting
autonomous multi-robots to inspect the safety of the

http://www.anglogold.com/subwebs/informationforinvestors/reportt
osociety05/values_bus_principles/safety_health/sh_cs_sa_5_9.htm

Volume 7 Number 5

http://cra.org/ccc/docs/init/computingfordisasters.pdf

Page 1317

www.ubicc.org

Ubiquitous Computing and Communication Journal

mine by assessing the status of the rocks, roofs, gases

and water level before miners enter to perform
operations. Mine inspections before miners enter the
underground area to work have not really been
explored. Such inspections are part of what we are
proposing as preventive safety measures.
The objective of this paper is to present a
comprehensive survey of safety inspection issues and
propose an autonomous multi-robot cooperative
behavioural model for underground mines. The main
contributions in this paper are as follows:
Development of a newly proposed cooperative
behavioural model for autonomous multi-robot
system (MRS) deployed for preventive underground
mine safety inspections.
Knowledge generation as a reference guide to
understand mine safety management in general and
cooperative behavioural models in particular detail.
A systematic survey and representation of research
challenges and open problems in the design of a
cooperative behavioural model.
The remainder of this paper is arranged as follows:
In section 2, we provide a paradigm of cooperative
behavioural models. In section 3, some insights on
robotics and mine safety disasters are provided. In
section 4, our proposed system model for MRS
behaviours in underground mines is presented. We did
some empirical analysis in section 5.In section 6,we
outline some open research issues in behavioural
modelling for autonomous MRS. We conclude this
paper in section 7.
2 COOPERATIVE BEHAVIOURAL
MODELLING PARADIGM
2.1 What is Safety?
Before we proceed to findings, it is important to
mention that the information technology, robotics and
mining disciplines have different perceptions of safety.
However, the definitions of the term safety that are
provided broadly underpin its dictionary meaning (see
Fig. 3). From the point of view of cooperative
behavioural models, the definition of safety appears to
be closer to those used in the robotics and mining
sphere. In robotics, self-driving cars are likened to a
safety system by Seth Teller6. Teller claims that the
number of accidents will decrease substantially with
self-driving cars. In the area of smarter transportation
systems, Jessica Richeri7 also developed an
autonomous vehicle able to recognise and avoid
obstacles, thus, it is hoped, preventing accidents.
6

http://www.popsci.com/cars/article/2012-04/who-blame-whenrobotic-car-crashes
7
http://theinstitute.ieee.org/career-and-education/preuniversityeducation/autonomous-robotic-vehicle-earnsstudent-ieeescholarship

Volume 7 Number 5

ISSN 1992-8424

Collision avoidance is another related safety area that

enables robots to navigate freely to their destinations
without colliding with any obstacle.
The mining definition of safety includes all mining
activities ranging from slips and falls to protection
from electric shock, machinery, eye injury and
drowning. Safety in mining involves training of
miners, which helps in maintaining the relationship
between working safely and productivity. Fig. 3 show
different safety definitions.
The researchers views and acceptance of a
cooperative behavioural model will be better
understood as a safety paradigm towards the end of the
paper
IT defines safety as a device
preventing unintentional
operation

Dictionary defines safety as

protection from or not being
exposed to the risk of harm or
injury

Robotics defines safety in terms

of robot safety as safeguarding
industrial robots from hazardous
conditions by programming and
maintaining the robots
Mining definition of safety
includes all mining accidents
ranging from slips and falls to
drowning etc.

Figure 3: Safety Definitions

2.2

Cooperative Behavioural Modelling

Frameworks
The cooperative behaviour of MRS has been studied
by some researchers; but its practicality and
implementation have not been finalized. Much
research is still required in this area to achieve
cooperation in real life. Generally, two types of
cooperation exist [5]:
Passive Cooperation: In this case, cooperation
exists when the whole environment is observed,
which is called emergent behaviour. Robots that
sense one another as obstacles and navigate away
from one another are examples of this type.
Active Cooperation: Robots in this group
cooperate through any of the means of
communication (wired, radio, wireless, etc.). In this
type of cooperation, details of the action to be
performed by the robots needs to be communicated
to all the robots.
One can use unmanned ground vehicles (UGVs) as an
example for the distinction of active and passive
cooperation. Active cooperation exists between
vehicles with sufficient bandwidth on the
communication channel for negotiation of actions,
while passive cooperation exist in situations where
there is limited bandwidth and so no negotiation with
one another occurs. MRS has been classified using
different definitions by different authors. Dudek et al.
present the taxonomy for MRSs in [6].

Page 1318

www.ubicc.org

Ubiquitous Computing and Communication Journal

ISSN 1992-8424

2.2.1 Task-based classification: This explains the type

of task a group of robots can undertake. For instance,
some: (i) require multi-agents, (b) are traditionally
multi-agents (transportation, industrial, agricultural
and fishing-related tasks), (iii) are inherently single
agents, and (iv) may benefit from the use of multiple
agents.
2.2.2 Size of robot collectives: This explains the
number of robot that can be in a group. Dudek
summarised this as: (i) the minimal collective is made
up of one robot (SIZE-AlONE), (ii) the simplest group
is made up of two robots (SIZE-PAIR), (iii) there
could be a limited number of multiple robots (SIZELIM) and (iv) an infinite number of robots (SIZEINF).
2.2.3 Communication: This is a class that shows the
different ways robots interact with other robots and the
environment. In this class, there are three subclasses:
(i) a situation where there are no direct
communications between robots (COM-NONE), (ii)
robots communicating locally (COM-NEAR) and (iii)
communication of robots in a wider range (COMINF).
2.2.4 Reconfiguration: This shows the rate at which
the group can re-organise itself and move with respect
to one another. The arrangement can be: (i) static
(ARR-STATIC), which means that the topology is
fixed (ii) a coordinated rearrangement (ARR-COMM);
here the topology can be alternated and (iii) dynamic
(ARR-DYN); behaviours here can change randomly.
2.2.5 Composition: This explains the different units
from which the group is made up. The collective can
be made up of different units: (i) identical (CMPIDENT), (ii) homogeneous (CMP-HOM) and (iii)
heterogeneous (CMP-HET).
2.2.6 Control: The existing architectures are explained
here. They are: (i) centralised (CTL-CEN), (ii)
decentralised (CTL-DEC) and (iii) hybrid (CTL-HYB)
[5].

2.3.1
Synchronous approach in MRS navigation:
The synchronous approach in MRS navigation
involves the use of an attached learning capability to
individual robots for navigation to their destinations at
the same time. It also involves real-time
communication and collaboration between robots as
they simultaneously navigate to achieve a given task.
2.3.2
Asynchronous approach in MRS navigation:
In this case, robots start navigation at different times
but connect according to their own schedule. The
interaction between the robots is easily captured,
shared and distributed in this approach. The ultimate
goal for a team of robots is to complete their mission
as quickly as possible without wasting energy. This
goal is worth investigating using both the synchronous
and asynchronous modes of navigation.

2.3 Deploying Cooperative Behavioural Models in

MRS Navigation
Our proposed software architecture for a
cooperative behavioural model is a fully distributed
architecture that uses adaptive action selection to
achieve cooperation. Robots in this system are
designed using behaviour-based approaches, where a
number of task achieving behaviours are active
simultaneously. The robots in this system react
quickly to the actions of other robots, by performing a
task differently from the other robots task and turning
in a direction different from the direction of the other
robots. To detect and interpret the actions of other
robots, there exists a communication network in the
form of wireless communication between the robots.
The robots broadcast statements of their current
actions to one another. Based on communication
strategy, the use of cooperative behavioural models
can be categorised in this paper as synchronous and
asynchronous in MRS navigations.

3 ROBOTICS, MINES AND SAFETY FEARS

Volume 7 Number 5

2.4 Application of Cooperative Behavioural

Models
Researchers have investigated the applications of
having a group of robots that cooperate in order to
complete some required tasks. Some of the
applications are exploring an unknown planet [7],
capturing intruders[8], pushing objects [9,10],
selecting action in a robot soccer game [17, 18],
cooperative construction and transportation [13],
cleaning up toxic waste [14], decision-making [15],
joint transportation [16].
The new behaviour-based model proposed in this
work involves the use of behaviours as the basic
representation level for learning. The goal of the
behaviour is to achieve a cooperative safety inspection
for a MRS in a rough terrain such as an underground
mine. To the best of our knowledge, despite being
widely used for cooperation in other applications, the
cooperative behavioural models of MRS have not been
applied to safety inspection tasks in an underground
environment.

3.1 Underground Mine Disasters

Mine disasters are caused by explosions (methane,
coal dust or others), fires, rock and roof falls, toxic gas
outbursts and inundations/rushing water [3]. Figs. 4
and 5 show the recent causes of fatalities and injuries
in South African mines. At present, fall of ground
accidents occur mainly in the deep underground mines
as shown in Figs. 4 and 5.

Page 1319

www.ubicc.org

Ubiquitous Computing and Communication Journal

ISSN 1992-8424

3.1.2 Specific disasters in the worlds underground

Mines
In Table 2, some underground mine disasters from
various countries and the causes are explained. The
most frequent cause of these disasters was rock fall.
The fatality numbers in Table 2 showed that South
Africa had the highest number (435), followed by
India (375).
Table 2: Disasters in Underground Mines [3]
Figure 4: Fatality per Casualty Classification

Year
6 August
2007

Mine
Crandall Canyon
Mine, USA

25 April
2006

Beaconsfield
Tasmania Mine,
Australia
Chasnala Mine,
India

27
December
1975

Figure 5: Injury per Casualty Classification

3.1.1 Comparison of South African fatality rates in
coal mines with other countries.
Table 1 show a comparison between the South
African coal mine fatality rate and that of other
countries such as the United Kingdom (UK), the USA,
Western Germany, New South Wales and India from
1978 to 1988. According to the table, the average
South Africa fatality rate is over eight times that of the
UK, three times that of Indians and approximately
double the rate in the USA, Western German and New
South Wales [2].
Table 1: Coal Mine Underground Fatality for South
Africa and other Countries [2]
South
Africa

UK

USA

New
South
Wales
0.38
0.83
1.90

India

0.52
0.79
0.72

West.
Germ
any
0.56
0.56
0.63

1978
1979
1980

1.46
1.72
1.37

0.23
0.21
0.20

1981

1.58

0.14

0.92

0.55

1.24

0.47

1982
1983

1.58

0.17

0.70

0.38

0.39

0.48

2.46

0.15

0.49

0.32

0.52

0.41

1984

1.05

N/A

1.04

0.54

0.66

0.34

1985

0.71

0.17

0.51

0.34

0.58

0.46

1986

0.39

0.14

0.61

0.26

0.56

0.45

1987

2.26

0.10

0.48

0.22

0.56

1988

0.72

0.35

0.16

0.11

Average

1.89

0.17

0.65

0.41

0.77

0.45

Volume 7 Number 5

21 January
1960

Coalbrook North
Mine South
Africa

5h August
2010

San Jose Coopergold Mine, Chile

Causes
A roof collapse trapped 6
miners in an underground coal
mine
A small earthquake triggered
an underground rock fall,
trapping 17 miners
An explosion that was followed
by flooding and then roof
cave- in drowned 375 miners
trapped underground.
This disaster was caused by
the collapse of several walls
and a series of rock falls,
releasing large quantities of
methane.
A roof cave-in trapped 33
miners in a chamber of about
2,300 feet below the surface.

Fatalities
9 deaths
6 injuries
1 death

375
deaths

435
deaths

3.2 Robotics Related Disasters

Robotics accidents result from equipment
malfunction, poor operating practices and other
elements as depicted in Fig. 6[29]. Over the years
industrial robot safety has received extensive attention,
followed by robots outside factories, such as domestic
and hospital robots. Recently, however, hazardous
area robots have been employed in cases where safety
and reliability are critical. The types of breakdown that
affect robots reliability and safety, according to [29],
are human error, software failure, random component
failure and systematic hardware failure.
Systematic
Hardware Fault
Radom
Component
Failure

Breakdown
Types

Human Error

Software
Failure

Figure 6: Types of Breakdown that Affect Robots

Reliability and Safety
Software failures are the most important concern in
robot related accidents. This is because they determine
the flexibility and robustness of the robot and hence
are the focus in this research. Human error is also part
of total robot breakdown, which is caused by people

Page 1320

www.ubicc.org

Ubiquitous Computing and Communication Journal

who design, manufacture, test, operate and maintain

robots. Unpredicted breakdowns are regarded as
random component failures [29]. Unknown
mechanisms in robot designs are major causes of
systematic hardware faults.
Over the years, industrial robots have caused some
hazards. The following are some past robot related
accident data from [18].
An injury led to the death of a worker who violated
safety devices by entering a material handling robot
area. The worker was trapped between the robot and
a post anchored to the floor.
The first robot-induced fatal accident occurred in
Japan in 1978.
The first fatal robot-related accident in the USA
occurred in 1984.
Between 1978 and 1984, at least five fatal accidents
involving robots occurred, four of these accidents
occurred in Japan and one in the USA.
Line workers and personnel were at the greatest risk
of injury, according to a study of 32 robot-related
accidents in Japan and the USA in 1987.
To understand the rationale for this work better,
we conducted a safety survey by sampling articles on
behavioural model from IEEE and science direct
databases. Fig. 7 shows the findings from the
empirical survey. About twenty articles from 1990 to
the present date were reviewed for this survey. The
outcome shows that while research efforts have been
undertaken to mitigate some related safety issues,
more work still needs to be done to deal with safety
issues, especially in underground mines. The larger
percentage of the existing work needs improvements,
about five out of the twenty surveyed articles do not
need improvements and 5% of the surveyed articles
did not refer to safety issues on both robotics and
mining. For instance, in the mining sector, several
rules and regulations have been issued to miners to
ensure safety in the mine, but the more the rules, the
more the disasters and fatalities. [27] states that mine
managements and regulators should stop creating more
rules. Human beings cannot keep up with all these
rules for safety; they are prone to errors most of the
time, especially in confined environments such as
mines [3]. This has created urgency to build models
for MRSs to mitigate the danger in these hazardous
environments.
Unknown
5%

No
27%

No
Yes

Yes
68%

Unknown

Figure 7: Do Existing Behavioural Models Need

Improvement for Robotics and/or Mine Safety?
4

ISSN 1992-8424

4.1

Machine Learning Algorithms as a Safety

Measure
Learning is needed in applications where humans
are unable to explain their skills, for example in
speech recognition, where solutions change over time
as in routing on computer networks, and where a
solution needs to be adapted to a particular case, for
example user biometrics . There are three main types
of
machine
learning:
supervised
learning,
unsupervised learning and reinforcement learning. Fig.
8 displays the major subfields of machine learning. In
this work, we intend to use machine learning
algorithms to address the safety inspection problem.
The platform which is used in developing our model
consists of two autonomous robots that know the
actions of one another through collaborating. For
instance, before the first robot A could take an action,
it needs to know the action taken by the second robot
B and vice versa. Learning is required to address this
cooperative problem and this can be undertaken by
reinforcement learning algorithms and swarm
intelligence (SI) techniques. To develop cooperative
behaviours for these robots, collaboration is required.
Cooperation and collaboration are two fundamental
elements required in behaviour-based robotics [5].
These elements are adapted from biological
mechanisms. For instance, the study of cooperative
behaviour of bees and ants from SI show the
possibilities for simple robots to work together and
solve a very complex problem [4]. Here, the model is
designed to help the robots to learn, make decisions,
execute actions and in the end get a reward. The model
that is developed will be used to discover the best
behaviour to handle inspections in underground mines
by MRS.
4.1.1 Reinforcement learning algorithm and swarm
intelligence (SI)
SI is a component of computational intelligence that is
used in solving stochastic/probabilistic problems like
in the field of robotics/artificial intelligence (AI),
process
optimization,
telecommunication
and
entertainment [28]. In the field of robotic/AI, SI can be
employed to address the cooperative behaviour of
robots because of the emergent collective intelligence
of groups of simple agents. Group foraging of social
insects, cooperative transportation8 are some of the
examples of emergent collective intelligence. Ant
colony optimisation (ACO), particle swarm
optimisation (PSO) and bee colony optimisation
(BCO) are some of the techniques of SI, as depicted in
Fig. 8b. SI is interesting in solving the aforementioned
problems because of the following features: distributed
system of interacting autonomous agents, selforganised control and cooperation, indirect interaction,
performance optimisation and robustness. These
features are inspired by nature and modelled through

PROPOSED COOPERATIVE
BEHAVIOURAL MODEL FOR
AUTONOMOUS MRS
8

Volume 7 Number 5

http://staff.washington.edu/paymana/swarm/krink_01.pdf

Page 1321

www.ubicc.org

Ubiquitous Computing and Communication Journal

the enablement of biological species (bees, ants,

particles etc.)9.
Reinforcement learning (RL) is used to learn how to
behave in a new environment. Here, autonomous
agents learn to act optimally/best without human
intervention. Agents learn by randomly interacting
with their environment, and getting sporadic rewards.
The purpose is to maximise these rewards. Fig. 9
explains the repeated sequence of events and the cause
and effect of an RL cycle. An RL system has four subelements apart from the agent and the environment:
policy, a reward function, a value function and/or a
model of the environment.
Models imitate behaviours of the environment,
predict the next state and give the initial state and
action. Models are generally used in planning future
situations. Fig. 10 is an example of learning in
robotics, where the small round lemon objects
represent static navigational features, the shaded
rectangular objects represent obstacles, the wheeled
objects represent multi-robots and the green spherical
object represents the goal state.
Supervised

Classification

Regression
Associative
Rule
Learning

K-nearest
neighbor etc.
Regression
Analysis etc.

ISSN 1992-8424

In the scenario, if the agent hits an obstacle, it gets a

negative reward, if it gets to the goal, it gets a positive
reward and when it gets the goal faster, it gets a bigger
reward.

Figure 9: Reinforcement Learning Cycle with Cause

and Effect
The list of some of the RL algorithms are shown in
Fig. 8a with two different categories: the model-based
and the model-free. Model-based ones build a model
of the environment and use it to learn the environment.
Examples include adaptive dynamic programming and
Monte Carlo methods. The model-free ones learn the
policy without model and the advantage of this
method is that it requires limited memory for its
operation. The temporal difference method is a type of
model-free RL.
4.1.2
POMDP
POMDP, the Partially Observable Markov Decision
Process, is used in modelling many kinds of real-world
problems such as robot navigation problems, planning

Apriori
Algorithm
etc.

Unsupervised
Types of
Machine
Learning

Reinforcement

Hierarchical
Clustering

Conceptual
Clustering
etc.

Partitional
Clustering

K-means
Algorithm
etc.

ModelBased

Adaptive
Dynamic
Programming

ModelFree

Temporal
Difference
Methods

Q-Learning
etc.

Figure 8a: The Taxonomy of Machine Learning

Hill Climbing

Probability
/Stochastic
Algorithms

Evolutionary
Algorithm

Evolution
Computation
Swarm
Intelligence (SI)

Random
Optimization

Genetic
Algorithm (GA)

Particle Swarm
Optimisation
(PSO)

Ant Colony
Optimisation
(ACO)
Bee Colony
Optimisation
(BCO)

Figure 8b: The Taxonomy of Stochastic Algorithms

Figure 10: Reinforcement Learning in Robotics

under uncertainty etc. It is used where access to a state
is not available but information about it is obtained
through to observation model. It is a generalisation of
the Markov Decision Process (MDP). MDPs are
modelled mathematical frameworks used in making
decisions in cases where outcomes are partly random
and partly under the control of a decision-maker. We
use POMDP for modelling because it is a standard
framework for stochastic processes and it also captures
process and sensor uncertainty. POMDP is a special
case of MDP10. For MDP, the environment is fully
observable while it is partially observable for
POMDP. The agent in POMDP cannot execute the
10

http://en.wikipedia.org/wiki/Swarm_intelligence

Volume 7 Number 5

http://www.cs.brown.edu/research/ai/pomdp/tutorial/in
dex.html

Page 1322

www.ubicc.org

Ubiquitous Computing and Communication Journal

optimal policy, since the current state is not

necessarily known. Equation (1) defines POMDP.
Here, agents can represent the situation of the
environment through the belief state. A belief state (b)
is a probability distribution over s. b(s) is the
probability of being in state s when the belief state is
b. Subsequently b can be calculated from a previous b.
The following defines a POMDP:
Set of states S, set of actions A, set of observations O
Transition model T(s, a, s)
Reward model R(s)
Observation model O(s, o) probability of
observing observation s in state o.
Optimal action that depends on the agents current
belief state.
POMDP can be discrete or continuous. Given a belief
state, an agent can perform an action a and perceive
observation o, and the new belief state becomes

-b(s) = O(s, o) T(s, a, s) b(s)

(1)

Where the optimal policy *(s) maps from belief states

to action.
Many real world problems are not fully observable
and the Markov assumption is often effective. The
proposed problem of this research is not an exception.
Some research works done with machine learning
techniques was surveyed and we conducted a survey to
show whether the ML techniques are appropriate for
space and collision avoidance problems. Fig. 11
summarises the results obtained from the survey. Out
of 25 literatures surveyed from IEEE and science
direct databases, 17 suggested that ML systems are
safe enough to avoid collision in any environment;
five of the publications disagree while three are
indifferent about using an ML system for collision
avoidance. The work on a robot soccer player platform
done in [11] and [12] revealed that ML techniques are
safe to avoid collision. [11] applied coupled agents on
modular Q-learning to enable the robot soccer system
to avoid colliding with one another. Furthermore, [12]
used the support vector machine and multilayer
perceptron techniques to develop a teamserver
program that support collision avoidance in MRS.
Not Safe
22%

Unknown
11%
Safe
67%

ISSN 1992-8424

we intend to focus scalability on the number of robots

(size) that can lead to safety inspection of underground
mines. This is expedient because teams of robots tend
to achieve tasks more quickly and effectively than
single robots [20]. Rosenfeld et al. surveyed a good
number of studies on how performance scales with an
increasing number of robots in a group and their
corresponding changes in productivity.
We intend to contribute to scalability in
underground mines by monitoring how the memory of
the system handles the addition of more robots to a
team using the memory management scheme called
paging (virtual memory). Again, we conducted a
survey on 25 publications from IEEE and science
direct databases on scalability and how it can
contribute to safety in underground mines. Fig. 12
shows that 62% of the work done in scalability will
support safety, 20% will not and 18% is undecided.
This was supported by [26], who states that using
multiple robots to assist in dangerous real life
scenarios such as in mines, security patrolling jobs and
rescue operations helps in saving human lives. He
stated that it is better for humans to monitor tasks done
in hazardous environments from a safe location as
robots get the work done. However, [21] demonstrated
safety in scalability by adding and deleting robots in
the system. Damaged robots are deleted from the
system to create room for new robots to be added to
enhance productivity and safety.
Unknown
18%
Not
Contribute
20%

Contribute
Contribute
62%

Not Contribute
Unknown

Figure 12: Does Scalability Contribute to Safety?

4.3

Packbot and Open Source Robots as a

Safety Measure
Fig. 13 show a picture of a real life robot called
Packbot. Packbot is a rugged robot manufactured by
iRobot. Packbots were the first robots to enter the
damaged Fukushima nuclear plant after the 2011
Japan Tohoku earthquake and tsunami11.

Safe
Not Safe
Unknown

Figure 11: Are Machine Learning Systems Safe

Enough to Avoid Collisions?
4.2
Scalability as a Safety Measure
There are different definitions of scalability
depending on the background of the person defining it,
the technology being considered, and the operational
use of the technology [19]. In robotics, there are
different types of scalability; (i) size (ii) memory (iii)
power (iv) information, etc. However, in this research,

Volume 7 Number 5

Figure 13: Packbot

Figure 14: Surveyor SRV-1 Blackfin

11

http://en.wikipedia.org/wiki/2011_T%C5%8Dhoku_earthquake_and
_tsunami

Page 1323

www.ubicc.org

Ubiquitous Computing and Communication Journal

In this research, simulations that determine the

safety of the model will be experimented using open
source robots designed for research, education and
exploration and similar underground environment such
as transportation tunnels. An open source robot that
could also be used for this work is the SRV-1 Blackfin
mobile robot, which is distributed by the inertial labs
[32]. SRV-1 in Fig. 14 can run on board interpreted C
programs or user-modified firmware, or it can be
managed remotely from windows, Mac OS/X or Linux
base station with python or Java-based console
software. Some of the amazing features of this

ISSN 1992-8424

platform are: an open design with full access code

(GPL), fully programmability for autonomous
operation, extensive software support through third
party applications, wireless remote control and the fact
that it can be controlled from a terminal or console for
easy testing.
4.4
The Proposed Framework
Fig. 15 is the proposed framework for this research.
The summary of how this framework works is
displayed in Figs. 16 and 17.

Figure 15: Framework of the Proposed Cooperative Behavioural Model

Figure 16: Layers of

the Framework
5

Figure 17: Processes of MultiRobot Behavioural System

EMPIRICAL ANALYSIS

5.1

Scenarios of the Cooperative Behavioural

Model

Volume 7 Number 5

Figure 18: Communicating

Robots Inspecting for Rock
Falls in an Underground
Mine

Figure 19: Communicating

Robots Inspecting for
Level of Gases in an
Underground Mine

Figs. 18 and 19 are scenarios of robots communicating

in an underground mine while inspecting for rock falls
and gas levels respectively and they are explained in
sections 5.1.1 and 5.1.2.
5.1.1 Scenario 1 MRS inspecting rock falls with
cooperative behaviour support
From the start of an inspection of roof tops in the
underground mine depicted in Fig. 18, robot A
broadcasts its south west (SW) movement direction to

Page 1324

www.ubicc.org

Ubiquitous Computing and Communication Journal

robots B and C. This triggers B and C to take

directions north east (NE) and south east (SE),
different from robot A behaviour. After inspecting
these aforementioned regions, robot A communicates
the finding obtained from the hanging bolts/materials,
robot B and C communicate their findings for cracks
in shot crete and loose rocks to the router and the
results are stored in the mine server through the inmine base gateway device. To locate the position of
the inspected region, radio frequency identification
(RFID) [33] tags are embedded on each of the
inspecting sensors on the robots. Following the
completion of the first inspection, robot A makes
another move in the north west (NW) direction and
broadcasts this to B and C as well. This again
stimulates B and C to navigate towards the East (E)
and South (S) directions respectively. The results of
this second inspection are also communicated to the
gateway through the wireless router. This process is
continued until the underground mine environment
(entrance, stope and gallery) has been fully inspected
and findings have been stored.
5.1.2
Scenario 2 MRS inspecting toxic gases with
cooperative behaviour support
Robot C starts inspection of the various toxic gas
levels (CH4, CO, SO2, NO2, NH3) in various
concentration levels in the underground mine as
depicted in Fig. 19, by transmitting its direction (SE)
to robots A and B. This stimulates robots B and A to
inspect the toxic gases in the NE and SW directions
respectively. This is expedient because no robot will
navigate to the directions already taken by the other
robots, which is one of the advantages of cooperative
behaviour for MRS. The findings obtained from their
inspection behaviours are communicated to the
gateway through the router and are stored in the mine
server. The inspection process is repeated again with
robot A broadcasting its new direction to B and C, and
robots B and C taking immediate action based on
robot As action, and findings from the actions being
stored. RFID tags are placed on each gas sensor for
location of the gas position.
5.2
Comparative Evaluations of Cooperative
Behavioural Models
Comparison between different behavioural models
is difficult since there is no common set of test data.
However, we are comparing the different
approaches/methods used in building the models,
different domains and platforms used in testing the
existing models and our proposed model. Table 9
compares some selected behavioural models with our
proposed model and also explains why our proposed
behavioural model seems better than some existing

Volume 7 Number 5

ISSN 1992-8424

models. A few of the features are presented in 5.2.1

5.2.3.
5.2.1 Behavioural models based on collision
avoidance success
Collision avoidance involves methods that attempt
to avoid simultaneous access to the same resource,
such as scheduling of timeslots, randomised access
times and carrier detection schemes 12. However, there
are different methods that have been used by
researchers to achieve collision avoidance in
behavioural based methods. A step-forward approach
built on an omni-directional vision system was
implemented to avoid static obstacles and dynamic
obstacles by [24]. This approach has helped in
decision-making on collision avoidance in MRS. A
sparse edge detection and reconstruction algorithm has
confirmed a more detailed view of the environment,
which aids in collision avoidance with obstacles for
unmanned aerial vehicles [25].
5.2.2 Behavioural models based on intelligence
Intelligence is a fundament tool for any successful
behavioural model. Cooperative model based
intelligence has been achieved in different domains
and environments. Some of the intelligence applied to
cooperative behavioural models are: An FSM principle
was used on cooperation capture for MRS [9]. The
author states that FSM was used because it simplifies
the design process of the system and also achieves
maximum results with little effort. Several solutions
for how robots cope with uncertainties along the path
from interpreting raw sensor inputs to behaviour
selection was achieved using autonomous colour
calibration, illumination invariance and autonomous
sensor and actuator modelling [15].
5.2.3 Behavioural models based on scalability
Scalability is the ability of a system to handle a
growing amount of work13. [21] listed one of the
benefits of MRS as the ability to add and remove a
robot from the system. A study of scalability
properties revealed that the ability to increase the
number of robots is needed in hazardous environments
and in certain domains where there are high
probabilities that robots may suffer damage [20]. It
was further stated that a group of robots is likely to
finish one task quickly and robustly. One can see that
the proposed model is sophisticated, with its socioeconomic impact, scalability, and cooperative
behaviour.
6

Open Research Issues in Behavioural

Modelling for Autonomous MRS
Researchers began investigating issues related to
multiple robot systems in the late 1980s when the field
of distributed robotics originated [17]. Prior to this
time, research did not involve robotic components but
12

http://en.wikipedia.org/wiki/Collision_avoidance_(networking)

13

Page 1325

http://en.wikipedia.org/wiki/Scalability

www.ubicc.org

Ubiquitous Computing and Communication Journal

was rather concentrated on a single robot system.

Parker identified eight primary research topics within
MRSs in 2003, which are still being studied today.
Fig. 20b summarises these issues. However, this
review paper contributed to discovering some open
research issues in cooperative MRS after critical
survey and investigations, as depicted in Fig. 20a. The
open research issues are scalability, hazardous
environment and real-life application.
Table 3: Comparing Different Behavioral Models with
our Proposed Model
S/N

Features

Action
Selection
Model[11]

InputOutput
Learning
Model[12]

Fault
Tolerant
Model[21]

Dynamic

static

Dynamic

Environme
nt
Domain

Underground
mine

Scalability

Scalable

Intelligence
Methods

Robot
soccer
platform
Not
scalable
SVM and
MLP14

On-line

Robot
soccer
platform
Not
scalable
Modular
Qlearning

Collision
Avoidance
Success

Yes, with
coupled
agent

Yes, with
teamserver

No

Autonomy

Not
autonomous

Yes

Navigation
Approach

Not
autonomo
us
Univector
field

Distributed
Memory
RL and
swarm
Intelligence
techniques
RL and
swarm
intelligence
principles
Yes

No
navigation
approach

Three
layered
feed
forward
neural
network

BRL15

Undergroun
d Mine
Safety
Inspection
Model
Dynamic

ISSN 1992-8424

empirical analysis of methods to use the proposed

model. The results of these surveys conformed to the
proposed mission of developing a safety inspection
model in a hazardous environment (underground
mine). Contributions to open research problems in
MRS cooperative behaviours were identified after a
critical review of this area and include: scalability,
hazardous environment and real life application. It is
believed that this proposed model for MRS will lead to
more safety applications in different hazardous
environments in the future while this survey can be
used as a reference guide to understand safety
management and cooperative behavioural models for
underground mines and to conduct further research on
the existing models to make them more efficient,
reliable and scalable, which can promote their use in
larger mines and applications.

Synchronous
and
asynchronous
approaches

Conclusion

This paper reviewed different causes of

underground mining disasters, proposed a behavioural
approach to safety inspection of underground mine
terrains by MRS and compared some behavioural
models with the proposed model. The proposed
behavioural model was explained using a three-layered
framework. The behavioural layer shows how two
robots cooperate while taking action and this concept
is achieved with ML algorithms and SI techniques.
The scalability layer corresponds to the situation of
increasing the number of robots to test how many
robots the system can handle without breaking down.
Achieving this involves using the memory
management scheme. Paging, a virtual address space
method, is applied to the scalability memory aspect of
this work.
The implementation with some open source
robots is suitable in the application layer. We
conducted an empirical survey and presented some
14
15

Support Vector Machine and Multi-Layer Perceptron

Bayesian-discrimination-function-based reinforcement learning

Volume 7 Number 5

Page 1326

www.ubicc.org

Ubiquitous Computing and Communication Journal

Figure 20b: MRS Research Topics and Their Key

Open Issues

Figure 20a: MRS Open Research Issues on

Cooperative Behavioural Modelling
ACKNOWLEDGEMENTS
The authors gratefully acknowledge the financial
support of the Organisation for Women in Science
for the Developing World (OWSDW) and the
resources made available by the University of Cape
Town (UCT) and University of South Africa
(UNISA), South Africa.
8 REFERENCES
8.1
Journals and Conferences
[1] W.A.Groves, V.J. Kecojevic, D. Komljenovic:
Analysis of Fatalities and Injuries involving
Mining Equipment, Journal of Safety
Research,Vol. 38, pp. 461 47(2007).
[2] J.P. Leger: Trends and causes of Fatalities in
South African Mines. Journal of Safety Science,
Vol. 14, PP. 169 185(1991).
[3] J.H. Saleh, A.M. Cummings: Safety in the
Mining Industry and the Unfinished Legacy of
Mining accidents: Safety Levers and Defense-indepth for Addressing Mining Hazards. Journal
of Safety Science,Vol. 49, pp. 764 777(2011).
[4] G. Beni, J. Wang: Swarm Intelligence in Cellular
Robotic Systems. In Proc. NATO Advanced
Workshop on Robotics and Biological Systems
1989.
[5] J. A. Leitner: Survey of Multi-Robot
Cooperation in Space. AT-EQUAL ECSIS
Symposium on Learning and Adaptive Behavior
in Robotics Systems, Vol. 2, pp.144
151(2009).
[6] D. Dudek, M. Jenkin, E. Milios, D. Wilkes: A
Taxonomy for Multi-agent Robotics. Journal of

Volume 7 Number 5

ISSN 1992-8424

Autonomous Robots, Vol.

3, pp. 375
397(1996).
[7] R. S. Aylett, D. P. Barnes: A Multi-robot
Architecture for Planetary Rovers. Proc. 5th ESA
Workshop on Space Robotics, ASTRA 1998.
[8] H. Yulan, Z. Qisong, X. Pengfei: Study on
Multi-Robot Cooperation Stalking Using Finite
State Machine. In International Workshop on
Information and Electronics Engineering
(IWIEE), Vol. 29, pp. 3502 3506(2012).
[9] M. Mataric, M. Nilsson, K.T. Simsarian:
Cooperative multi-robot box-pushing. In Proc. of
IEEE/RSJ IROS, Vol. 3, pp. 556 561(1995).
[10] D. Rus, B. Donald, J. Jennings: Moving
Furniture with Teams of Autonomous Robots. In
Proc. of IEEE/RSJ IROS, Vol. 1, pp. 235
242(1995).
[11] P. Kui-Hong, K. Yong-Jae, K. Jong-Hwan:
Modular Q-Learning Based Multi-Agent
Cooperation for Robot Soccer. Journal of
Robotics and Autonomous Systems, Vol. 35, pp.
109 122(2001).
[12] A. Sadegh: Machine Learning in a Multi-Robot
System. Department de Electique de lEcole
Polytechnique de Montreal, Canada
[13] F. W. Heger,; L. M. Hiatt, B. Sellner, R.
Simmons, S.
Singh: Results in Sliding
Autonomy for Multi-Robot Spatial Assembly. In
8th International Symposium on Articial
Intelligence, Robotics and Automation in Space,
58(2005).
[14] L. E. Parker: ALLIANCE: An Architecture for
Fault-tolerant Multi-robot Cooperation. IEEE
Trans. on Robotics. and Automation, Vol. 14,
220 -240(1998).

Page 1327

www.ubicc.org

Ubiquitous Computing and Communication Journal

[15] P. Stone, M. Sridharan, D. Stronger, G.

Kuhlmann, N. Kohl, P. Fidelman, N.K. Jong:
From Pixels to Multi-Robot Decision-Making: A
Study in Uncertainty. Journal of Robotics and
Autonomous Systems, Vol. 54, pp. 933943(2006).
[16] J. Fink, N. Michael, S. Kim, V. Kumar: Planning
and Control for Cooperative Manipulation and
Transportation with Aaerial Robots. In Int.
Symposium of Robotics Research 2009.
[17] L.E. Parker: Current Research in Multi-robot
Systems. In Seventh International Symposium
on Artificial Life and Robotics (ISAROB), Vol.
7, 1 5(2003).
[18] B.S. Dhilion: Robot Safety Analysis Methods.
Department of Mechanical Eng. University of
Ottawa, Canada.
[19]
E.S. Redden, R.A. Pettitt, C.B. Carstens,
L.R. Elliott: Scalability of Robotic Displays:
Display Size Investigation. In Army Research
Laboratory, 2008.
[20] A. Rosenfeld, G.A.; Kaminka, S. A.; Kraus:
Study of Scalability Properties in Robotic
Teams. Bar Ilan University, Ramat Gan, Israel.
[21] T. Yasuda, K. Ohkura, K..A. Ueda: A
Homogeneous Mobile Robot Team that is FaultTolerant. In Journal of Advanced Engineering
Informatics , Vol. 20, pp.301-311(2006).
[22] D.P. Stormont, V.H. Allan: Managing Risk in
Disaster Scenarios with Autonomous Robots. In
Journal of Systemics,
Cybernatics
and
Informatics (IIISCI), Vol. 7, pp. 66 71(2009).
[23] S. Molina, I. Soto, R. Carrasco: Detection of
Gases and Collapses in Underground Mines
using WSN. In Industrial Technology (ICIT) and
IEEE International Conference, pp. 212
225(2011).
[24] C. Cai, C. Yang, Q. Zhu, Y. Liang: Collision
Avoidance in Multi-Robot Systems. In
Proceedings of the IEEE International
Conference on Mechatronics & Automation,
2007.
[25] E. Hanna, P. Straznicky, R. Goubran: Obstacle
Detection for Low Flying Unmanned Aerial
Vehicles using Stereoscopic Imaging. In IEEE
International Instrumentation and Measurement
Technology Conference Victoria, 2008, Vacuver
Island, canada.
[26] D. Portugal, R.P. Rocha: On the Performance
and Scalability of Multi-Robot Patrolling
Algorithms. In proceedings of the IEEE
International Symposium on Safety, Security and
Rescue Robotics, pp. 50 55(2011).
[27] D. Laurence: Safety Rules and Regulations on
Mine Sites The Problem and a Solution.
Journal of Safety Research, Vol. 36, pp. 39
50(2005).
[28] Duch, W. What is Computational Intelligence
and Where is it Going?. Department of

Volume 7 Number 5

ISSN 1992-8424

Informatics, Nicolaus Copernicus University,

Poland.
8.2
Books
[29]
D.F. Tver, R.W. Bolz: Robotics Source
Book and Dictionary, Industrial Press, New
York, 1983.
[30] Mine Health and Safety Council Annual Report
2009-2010, 2003-2004.
8.3
[31]

Websites
Bullion web Services. South Africas
Mineral Industry Facts & Figures 2009, URL:
http://www.bullion.org.za/Publications/Facts&Fi
gures2009/F&F2009.pdf.
(Accessed
on
10/06/2012)
[32]
SRV-1 Blackfin Mobile Robot: Open
Source Robots designed for Research 2006
2009.
http://surveyorcorporation.stores.yahoo.net/srrowestkit.html
(Accessed on 10/06/2012)
[33]
RFID Technology:
http://www.12manage.com/methods_rfid_techno
logy.html (accessed on 14/09/2012)

Page 1328

www.ubicc.org