Sei sulla pagina 1di 6

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/265117465

Infrastructure Monitoring With Multi-Robot Teams

Conference Paper · September 2010

CITATIONS READS
11 69

4 authors, including:

Gonçalo Cabrita Lino Marques


University of Coimbra University of Coimbra
15 PUBLICATIONS   177 CITATIONS    157 PUBLICATIONS   2,147 CITATIONS   

SEE PROFILE SEE PROFILE

Aníbal T. de Almeida
University of Coimbra
411 PUBLICATIONS   5,832 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

EETT - Energy Efficiency Training of Trainers View project

ENERsip View project

All content following this page was uploaded by Lino Marques on 13 January 2015.

The user has requested enhancement of the downloaded file.


Infrastructure Monitoring With Multi-Robot Teams
Gonçalo Cabrita, Pedro Sousa, Lino Marques and Anı́bal T. de Almeida

Abstract— This article presents the first steps toward an aerial vehicles (UAVs) or even underwater unmanned vehi-
automatic indoor environmental monitoring system through the cles (UUVs), allowing for their deployment in almost any
use of a group of mobile robots. A metric map is provided a scenario. A mobile WSN will ultimately perform the same
priori to the robots, ensuring the navigation, localization and
extraction of points of interests for patrolling. The communica- task a static WSN would, however a small number of mobile
tion layer is robust and allows any robot to enter or leave the sensors is able to achieve a similar spatial resolution to that
patrolling task. The monitoring of the environment is achieved of a static WSN installed over a larger area. Furthermore a
by acquiring environmental information during the patrolling. team of robots can be deployed virtually anywhere in a short
If an abnormal condition is detected, the system should react, amount of time, hence being a far more flexible solution [5].
providing a fine coverage of the suspicious area and an accurate
identification of its source. Experiments were conducted inside Finally some applications can benefit from the use of both
a building and data was gathered and post processed. The yet static and mobile WSNs. Mobile robots deployed within
preliminary results demonstrate the effectiveness of the method an environment equipped with a static WSN can tap into
employed to control the system, showing good area coverage the existing network to access the environmental data of
from the patrolling algorithm and acceptable representations of the covered area. This allows the robot to make decisions
the collected variables (alcohol concentration and temperature).
based on this data and get more detailed readings, thus
improving the coverage and spatial resolution of the complete
I. INTRODUCTION
system [5].
Environmental monitoring can be described as the process Patrolling can be defined as the task of repeatedly visiting
of collecting the necessary data to characterize the quality a desired location with the purpose of assessing certain
of the environment. This notion can be applied to countless aspects of its environmental state. Since it is not possible to
indoor or outdoor applications. A wide range of Man-built cover all space at all times each point is visited once every
structures accommodate chemicals potentially dangerous to T seconds, thus the frequency of patrolling is defined as 1/T
both its occupants and Nature. Furthermore public buildings Hz [6]. Also known as sweeping or repetitive coverage, the
such as airports or train-stations have become preferential tar- task of patrolling has had considerable attention from the
gets for terrorist attacks. The timely detection of environmen- mobile robot community in the past few years. Solutions
tal anomalies in this scenarios can prevent potential disasters for multi-robot scenarios usually approach the problem by
and the consequent life losses and property damage [1]. dividing the area to patrol into sub-areas which are then
Other structures might hold animal or plant life that needs assigned to the available robots. Once this is done each
certain environmental conditions in order to be properly robot will patrol its own sub-area by means of a single-
sustained. Activities like agriculture can benefit from such robot patrolling algorithm [6]. In 2002 Machado presented
technologies. In museums all over the world variables like a discussion of multi-agent patrolling task issues. Several
temperature and humidity must be kept constant at all times architectures were then compared, from which the best
in order to help preserve the art pieces which continuously strategy was considered to be Conscientious Reactive, which
struggle against time [2]. is a local and reactive strategy with no communication, based
Environmental monitoring is usually achieved by means of on individual idleness and without central coordination [7].
a sensor network. The network nodes can be static, mobile Later in 2004, Chevaleyre carried out a study focused on two
or a combination of both. Static wireless sensor networks graph-theory centralized planning strategies: cyclic strate-
(WSN) consist of small nodes equipped with sensors capable gies and partitioning strategies [8]. Elmaliach studied the
of measuring the desired phenomena. The available solutions problem of generating patrolling paths for a team of robots
are usually cheap and easy to deploy, even over large areas, and presented it in [6]. The patrolling algorithm presented
both indoors and outdoors. Static WSN have found their guarantees that each point in the target area is covered at the
way into many monitoring applications, from museums [3] same optimal frequency through the use of Hamilton cycles.
to large glaciers [4]. Mobile robots equipped with multiple More recently, Portugal and Rocha developed the Multilevel
sensors can create a mobile WSN. Mobile robot platforms Subgraph Patrolling (MSP) algorithm [9]. This algorithm was
come in many shapes, from small ground robots to unmanned proven to be superior to existing patrolling algorithms.
Chemical variables can be monitored by a static sensor
G. Cabrita, P. Sousa, L. Marques and A. T. de Almeida are with Dept. network or by mobile robots equipped with monitoring
of Electrical and Computer Engineering, Institute of Systems and Robotics, devices. The most common types of chemical variables
University of Coimbra, 3030-290 Coimbra, Portugal.
{goncabrita, pvsousa, lino, monitored are: volatile organic compounds (VOCs), air con-
adealmeida}@isr.uc.pt taminants, and other type of toxic or hazardous gases [10],
[11], [12]. This type of chemical monitoring can have III. PROPOSED METHODOLOGY
multiple objectives from simple air quality supervision to
A. Map Representation
contamination level tracking or even to trigger alarms in the
case of a serious abnormal occurrence (such as a poisoning A task such as multi-robot patrolling, with the exception
gas leak or other accident). of simpler reactive algorithms, inherently needs a map. Maps
In the past few years, several researchers have studied usually come in two flavors, metric and topological. Metric
the ability of a mobile robot to react to chemical cues and maps provide a realistic representation of the environment,
track those cues until their source. This problem, known on the down side they are heavy and complex to process
as the odor plume tracking problem, has been addressed by most algorithms. Topological maps store less information
with several different approaches, either in indoor or in by comparison and as a consequence can be used much
outdoor environments. Marques [13] compared the perfor- more efficiently by most algorithms. For this project both
mance of a gradient-based strategy with two bio-inspired types of maps were used. A metric map of the infrastructure
strategies (moth surge/zig-zagging/spiraling, and bacteria to be patrolled is used for navigation and localization.
biased random walks) tracking two different odor sources in The same map is converted into a topological map using
a turbulent environment. In a recent work, Lochmatter [14] OpenCV. The discrete Voronoi diagram is extracted from the
compared the tracking ability of three bio-inspired algorithms metric map, from which the points of interest are obtained.
in a laminar wind field. Other researchers employ dense The topological map is necessary to perform the patrolling
coverage approaches [15] or nave physics rules to estimate algorithm.
the localization of odor sources indoor [16].
The rest of this paper is organized in the following B. Communication
order: Section II states the problem presented in this paper Although it is not mandatory, communication is useful
followed by Section III where the proposed methodology for when using a group of robots. Communication can be
solving the problem is explained. The experiments performed centralized or distributed. Distributed communication is more
through out the project are described in Section IV while robust since it does not depend on a server to operate
the results and discussion can be found in Section V. Finally properly. This allows for any part of the system to be dis-
Section VI presents both the conclusions and future work. connected without compromising the integrity of the group.
A communication node was developed by the researchers
II. PROBLEM STATEMENT involved in this project. It relies on OLSR1 to monitor robots
The goal of this work is to be able to monitor a known connecting and disconnecting from the network. Olsrd is
indoor environment. It is assumed that a given number of the implementation of the Optimized Link State Routing
robots, equipped with sensors able to measure the variables protocol (OLSR) which provides mesh routing for network
of interest and to communicate with each other are available. equipment. Olsrd is an ongoing open source project, fast,
Each robot should be able to navigate inside the facility, highly portable (runs on multiple platforms) and scalable
patrolling and detecting anomalous environmental conditions (runs on community wireless mesh networks with several
(e.g., a high concentration of a chemical vapor). When this nodes) routing protocol for mobile ad-hoc networks [17]. The
happens, the source of the problem should be identified, developed communication node also allows for the exchange
either by the detecting robot or jointly with the support of of messages between different robots without the need of a
other robots. In the current work only chemical releases were server, allowing for any part of the system to malfunction
considered and the problem can be divided in the following without compromising the integrity of the whole system.
sub-problems: (1) patrolling and (2) odor source localization.
The patrolling stage begins with the assignment of patrol C. Patrolling Algorithm
sub-areas to each robot, after which each robot will patrol its The MSP algorithm presented in [9] was chosen for
designated area. It is assumed that a metric map of the facility this project. The MSP Algorithm assumes that robots are
is provided to the robots. For a flexible and robust system, endowed with the environment map and the ability for
robots joining or leaving the communication network should self-localization and navigation. The algorithm is based on
lead to a redistribution of the patrolling areas thus providing multilevel partitioning of the environment map, assigning
a higher patrolling frequency given a higher number of robots different regions to each mobile agent. Each region corre-
or not allowing a certain area to remain un-patrolled in the sponds to a sub-graph extracted from the existing topological
event of a robot malfunction. representation. The algorithm deals, then, with effectively
The odor source localization stage is triggered when a patrolling each region by computing paths for every robot
robot detects an abnormal level of any of the chemicals it in the assigned sub-graph. To accomplish this, it searches
is able to sense in the environment. In order to perform a each sub-graph using a classical algorithm for Euler cycles
multi-robot odor source localization algorithm the necessary (visit each edge exactly once) and various heuristics for
number of robots should be mobilized. Upon the detection Hamiltonian cycles (being a closed loop and visit each
of the odor source a robot should remain to mark the spot. region exactly once), non-Hamiltonian Cycles (do not satisfy
The following section provides an overview of the pro-
posed methodology for solving the problem stated earlier. 1 http://www.olsr.org/
Fig. 1. Odor source (cross) and the chemical plume in a junction.

Hamiltonian conditions) and longest paths (a path with the


maximum length in the given sub-graph).

D. Navigation and Localization Fig. 2. The Roomba robots.


As stated earlier, a metric map was used for navigation
and localization. This is accomplished through the navigation
stack from ROS2 (further addressed in Section IV). The
navigation stack takes information from the robot odometry
and from a laser range finder, and outputs to the robot
platform safe velocity commands that allow reaching a
target positions [18]. The adaptive Monte Carlo localization
approach [19], which uses a particle filter to track the pose of
a robot against an existing map is present in the navigation
stack, providing robot localization in the known environment.
This allows the robot to visit the points of interest provided
by the patrolling algorithm described previously.

E. Monitoring the Environment


Fig. 3. Chemical sensor used in the experiments.
Monitoring the environment parameters is a complex and
difficult task. In this work, mobile robots are programmed to
patrol a certain area while acquiring various environmental F. Odor Source Localization
parameters. Each node of a given path can be visited with a
homogeneous frequency or with a frequency dependant with In the current implementation the robots are not measuring
its expected interest (e.g., the expected probability of ob- airflow intensity and direction. This constraints the odor
serving an accident or abnormal occurrence). This frequency plume tracking algorithms that can be employed, since the
depends from the robot path, so the path of each robot should most effective way to find an odor source consists in moving
be optimized in order to obtain complete area coverage with upwind through the plume. Given this constraint and the type
a frequency of visits as close as possible to the desired of environment that was used in the experiments (mostly
one. The robot can detect several items, namely: general corridors, junctions and rooms), a discrete search across the
pollutants (e.g., CO2 , CO, N Ox , SO2 , VOCs) or specific nodes and some heuristics was employed to move across the
chemical substances (e.g., hydrocarbons, ammonia, or other corridors containing the chemical vapor of interest and esti-
poisoning gases), particles, temperature or humidity, among mate its source. As can be seen in Figure 1, when a chemical
others. Basically, it tries to create a site characterization trace is detected by the robots, there are a limited number of
where its localization along with the acquired parameters branches from where it can come up and therefore a small
(alcohol and temperature in our experiments) can be shown. number of alternatives for the odor source localization. With
The spatio-temporal data acquired by each robot is sent a directional anemometer, it would be possible to explore
to a central station where all the relevant variables are immediately the most probable branch. The team is currently
fused in order to construct real-time risk maps of the whole working in the integration of directional anemometers and
environment. While performing the patrolling task, whenever in the implementation of efficient odor plume tracking and
an abnormal reading occurs, an alarming event should be localization.
triggered (to assess and address the information received)
IV. EXPERIMENTS
and a source tracking behavior should be started.
For this project a small group of Roomba robots depicted
2 Robot Operating System - http://www.ros.org/wiki/ in Figure 2 was used. The Roomba is manufactured by
93.75

81.25

68.75

56.25

43.75
Fig. 4. Nodes in the map (numbers) and the release point (A).
31.25

18.75

6.25

% of detection

Fig. 6. Robot paths and the chemical data acquired by the robots.

75

65

Fig. 5. The chemical release mechanism.


55

45

iRobot as an autonomous vacuum cleaner, however it is


possible to access and control the robot. This provides a 35

cheap yet robust platform, ideal for multi-robot experiments. 25

Each robot is equipped with a small Eee PC running ROS, an


15
Hokuyo laser range finder, a small artificial nose (shown in
Figure 3) and a thermopile array - TPA81 - from Devantech 5

(used for temperature acquisition). temperature (ºC)

ROS is an open-source operating system for robots. It


provides the services expected to be found in an operating
system such as hardware abstraction, low-level device control Fig. 7. Robot paths and the temperature measurements acquired by the
or message-passing between processes [20]. robots.
Real world experiments were carried out in the corridors of
the Institute of Systems and Robotics (ISR) in the University
of Coimbra. The initial experiments did not include real concluding a patrol cycle the robots can either go back to
chemical releases, instead PlumeSim was used to simulate their first node or perform the inverse path. The inverse path
a plume in the corridors of ISR. PlumeSim is an odor was chosen for this project.
plume simulator designed by the researchers involved in The chemical and temperature data collected during the
this project [21]. This allowed to fine tune the algorithm patrolling with real chemical release is here represented in
before releasing actual chemicals into the somewhat crowded the form of spatial-temporal graphics. Figure 6 shows the
corridors of ISR (Figure 4) for the final experiments. results for the chemical data while Figure 7 shows the results
The final set of experiments included a heat source and for the temperature data. Axis x and y represent the position
the setup of Figure 5 capable of releasing alcohol into of the robot throughout the experiment in meters, the z-
the environment. The location of the heat source and the axis represents time in seconds and the color represents the
chemical release mechanism is represented as point A in chemical concentration or temperature.
Figure 4. Each of the three robots was deployed at their Both the chemical (Figure 6) and temperature (Figure 7)
starting locations, and the corridors were patrolled for 30 data indicate that after the robot has passed point A in
minutes during each experiment, during which time both the Figure 4 the values read by the sensors take a considerable
chemical and temperature data were monitored. travelled distance to come back to clean air concentration and
room temperature. This can be explained by the relatively
V. RESULTS AND DISCUSSION fast speed imposed on the robot to assure that a satisfactory
The MSP algorithm provided the nodes shown in Figure 4. patrolling frequency is achieved and by the time that the
The resulting patrolling paths were: robot one 1-2-3-2-4- sensors take to settle back after being exposed to high
2-5; robot two 6-5-7-5-8; robot three 9-8-10-8-11. After chemical concentrations or high temperatures. This is also
R EFERENCES
[1] L. Marques, A. Martins, and A. T. de Almeida, “Environmental
monitoring with mobile robots,” in IEEE/RSJ Int. Conf. on Intelligent
Robots and Systems, Edmonton, Canada, 2005.
[2] D. Camuffo, R. Grieken, H. Busse, G. Sturaro, A. Valentino,
A. Bernardi, N. Blades, D. Shooter, K. Gysels, F. Deutsch et al.,
“Environmental monitoring in four European museums,” Atmospheric
environment, vol. 35, pp. 127–140, 2001.
[3] D. Abruzzese, M. Angelaccio, B. Buttarazzi, R. Giuliano, L. Miccoli,
and A. Vari, “Long life monitoring of historical monuments via
wireless sensors network,” in Proc. of the 6th Int. Conf. on Symposium
on Wireless Communication Systems. IEEE Press, 2009, pp. 570–574.
[4] K. Rose, J. Hart, and K. Martinez, “Seasonal changes in basal
conditions at Briksdalsbreen, Norway: the winter–spring transition,”
Boreas, vol. 38, no. 3, pp. 579–590, 2009.
[5] R. Stranders, “Coordinating Teams of Mobile Sensors for Monitoring
Environmental Phenomena,” School of Electronics and Computer
Fig. 8. The Roomba robots during patrolling task. Science, University of Southampton, Tech. Rep., May 2009. [Online].
Available: http://eprints.ecs.soton.ac.uk/17535/
[6] Y. Elmaliach, N. Agmon, and G. Kaminka, “Multi-robot area patrol
under frequency constraints,” in IEEE Int. Conf. on Robotics and
clear by the fact that for example on Figure 6 none of the Automation (ICRA2007), 2007, pp. 385–390.
[7] A. Machado, “Patrulha Multiagente: Uma Análise Empı́rica e Sis-
other robots sense the chemical even when close to nodes 5 temática,” Master’s thesis, Universidade Federal de Pernambuco
(robot 1) and 8 (robot 2) show in Figure 4. Nonetheless the (UFPE), Recife, Brasil, 2002, (In Portuguese).
robot is capable of pinpointing the point of chemical release, [8] Y. Chevaleyre, “Theoretical analysis of the multi-agent patrolling
problem,” in Proc. of IEEE/WIC/ACM Int. Conf. on Intelligent Agent
a maximum of chemical concentration is perfectly visible in Technology (IAT 2004). IEEE Computer Society, 2004, pp. 302–308.
Figure 6 coincident with point A in Figure 4. On the other [9] D. Portugal and R. Rocha, “MSP algorithm: multi-robot patrolling
hand the robot was not able to pinpoint the same point for based on territory allocation using balanced graph partitioning,” in
Proc. of the 2010 ACM Symposium on Applied Computing. ACM,
the heat-source, indicating that a more suitable temperature 2010, pp. 1271–1276.
sensor should be used for the purpose of this experiment. [10] R. Shepherd, S. Beirne, K. Lau, B. Corcoran, and D. Diamond,
“Monitoring chemical plumes in an environmental sensing chamber
with a wireless chemical sensor network,” Sensors and Actuators B:
chemical, vol. 121, no. 1, pp. 142–149, 2007.
VI. CONCLUSIONS AND FUTURE WORK [11] M. A. Pillai, S. Veerasingam, and Y. S. D., “Implementation of Sensor
Network for Indoor Air Quality Monitoring Using CAN Interface,” in
IEEE Int. Conf. on Advances in Computer Engineering, Los Alamitos,
The presented work represents another step towards practi- CA, USA, 2010, pp. 366–370.
cal real-life implementation of recent advances in the field of [12] J. Hayes, S. Beirne, K. Lau, and D. Diamond, “Evaluation of a low
cost wireless chemical sensor network for environmental monitoring,”
mobile robotics in the area of environmental monitoring. The IEEE Sensors, pp. 530–533, 2008.
initial results are satisfactory as the group of robots is able [13] L. Marques, U. Nunes, and A. de Almeida, “Olfaction-based mobile
to perform the patrolling and monitoring tasks as expected. robot navigation,” Thin Solid Films, vol. 418, no. 1, pp. 51–58, 2002,
selected from 1st Int. School on Gas Sensors.
The focus of this work will now go to the multi-robot odor [14] T. Lochmatter and A. Martinoli, “Theoretical Analysis of Three Bio-
source localization algorithm. Further experiments will also Inspired Plume Tracking Algorithms,” in Proc. of the 2009 IEEE Int.
be performed with a larger group of robots to determine Conf. on Robotics and Automation (ICRA 2009), May 2009.
[15] A. Loutfi, S. Coradeschi, A. Lilienthal, and J. Gonzalez, “Gas dis-
the robustness and flexibility of the system as a whole, tribution mapping of multiple odour sources using a mobile robot,”
determining how the group of robot reacts to new robots or Robotica, vol. 27, no. 02, pp. 311–319, 2008.
malfunctioning robots and how the patrolling routes are re- [16] G. Kowadlo and R. Russell, “Improving the robustness of naı̄ve
physics airflow mapping, using Bayesian reasoning on a multiple
arranged as needed. Moreover the temperature sensor proved hypothesis tree,” Robotics and Autonomous Systems, vol. 57, no. 6-7,
to be inadequate for the purpose of this project, as a result pp. 723–737, 2009.
a new temperature sensor will be procured for the future, [17] A. Tnnesen, T. Lopatic, H. Gredler, B. Petrovitsch, A. Kaplan, S. Tcke
et al., “Olsrd,” http://www.olsr.org/.
along with an humidity sensor (or even a combination of [18] E. Marder-Eppstein, E. Berger, T. Foote, B. Gerkey, and
both), relevant for many of the applications described early K. Konolige, “The Office Marathon: Robust Navigation in an
in the paper. Indoor Office Environment,” in IEEE Int. Conf. on Robotics
and Automation (ICRA2010), May 2010. [Online]. Available:
Since the developed work is modular it is possible to easily http://www.ros.org/wiki/Papers/ICRA2010 Marder-Eppstein
integrate future advances in the fields of patrolling and odor [19] D. Fox, “KLD-Sampling: Adaptive Particle Filters and Mobile Robot
Localization,” in In Advances in Neural Information Processing Sys-
source localization, hence improving the final result of this tems (NIPS, 2001.
project. Future work could also include an initial stage of [20] M. Quigley, K. Conley, B. Gerkey, J. Faust, T. Foote, J. Leibs,
autonomous mapping, which would allow the robots to be R. Wheeler, and A. Y. Ng, “ROS: an open-source Robot Operating
System,” in ICRA Workshop on Open Source Software, 2009.
deployed in an unknown environment. If this was the case [21] G. Cabrita, P. Sousa, and L. Marques, “PlumeSim - Player/Stage Plume
providing the robots with a map of the facility would not be Simulator ,” in ICRA Workshop on Networked and Mobile Robot
necessary. Furthermore, if the robots were able to recharge by Olfaction in Natural, Dynamic Environments, 2010.
means of charging stations it would be possible to maintain
robots patrolling the target facility indefinitely.

View publication stats

Potrebbero piacerti anche