The COMRADE System for Multi-Robot Autonomous Landmine

Detection in Post-Conflict Regions

Prithviraj Dasgupta1,Jose Baca Garcia1, K. R. Guruprasad2,

Angelica Munoz-Melendez3, Janyl Jumadinova4

1Computer Science Department, University of Nebraska, Omaha, USA2Mechanical Engineering Department, NIT, Karnataka, India

3Computer Science Department, INAOE, Mexico4Computer Science Department, Allegheny College, PA, USA

Abstract

We consider the problem of autonomous landminedetection using a team of mobile robots. Previousresearch on robotic landmine detection mostlyemploys a single robot equipped with a landminedetection sensor to detect landmines. We envisagethat the quality of landmine detection can be signif-icantly improved if multiple robots are coordinatedto detect landmines in a cooperative manner byincrementally fusing the landmine-related sensorinformation they collect and to then use that in-formation to visit locations of potential landmines.Towards this objective, we describe a multi-robotsystem called COMRADES to address differentaspects of the autonomous landmine detection prob-lem including distributed area coverage to detectand locate landmines, information aggregation tofuse the sensor information obtained by differentrobots, and, multi-robot task allocation (MRTA)to enable different robots determine a suitablesequence to visit locations of potential landmineswhile reducing the time required and batteryexpended. We have used commercially availableall-terrain robots called Coroware Explorer that arecustomized with a metal detector to detect metallicobjects including landmines, as well as indoorCorobot robots, both in simulation and in physicalexperiments, to test the different techniques inCOMRADES.

Keywords: robotic landmine detection: coverageand exploration; sensor information fusion; multi-robot task allocation

1 Introduction

Humanitarian demining is a crucial effort for thesafety and sustenance of human lives in post-conflictregions. Unfortunately, recent surveys on landminemonitoring report that humanitarian demining ef-forts are considerably lagging behind anti-personnellandmine planting activities due to several techno-logical and economic reasons [27]. This results inenormous loss to human lives; e.g., in 2010 alone,explosions of landmines and similar devices resultedin 4, 191 casualties, with civilians accounting for 70%of the casualties. One the major technological chal-lenges in humanitarian demining is to detect land-mines rapidly and with reasonable accuracy, whilereducing the number of false positives. We envis-age that automating landmine detection operationsusing multiple, off-the-shelf autonomous robots willprovide a reasonably accurate yet economical solu-tion to the problem of detecting landmines. Towardsthis objective, we describe a multi-robot systemcalled COMRADE (COoperative Multi-Robot Au-tomated DEtection) System for humanitarian dem-ining. The central objective of COMRADES is todevelop novel coordination techniques between mul-tiple low-cost, mobile robots, which enable them toautonomously and collaboratively detect landmineswith high accuracy in post-conflict regions. COM-RADES includes techniques that allows each robotto explore an initially unknown region while search-ing for landmines, recognize landmine-like objects onits sensors, share and fuse the landmine-related sen-sor information with other robots and coordinate itsactions with other robots, so that multiple robotscan converge on the object to analyze and confirm

it as a landmine. In this paper, we present thedescription and experimental results from differenttechniques for coverage, task allocation, and, multi-sensor information aggregation and sensor schedul-ing using multiple robots, that we have developed aspart of COMRADES. Specifically, we describe thefollowing aspects of multi-robot autonomous land-mine detection in COMRADES:

• A distributed area coverage technique that al-low a set of robots to dynamically partition aninitially unknown environment into a set of non-overlapping regions and search for landmineswithin each region. The techniques are robustto individual robot failures and are able to scalewith the number of robots and size of the envi-ronment.

• A distributed information fusion technique toaggregate landmine-related sensor informationfrom different robots using a prediction marketbased technique and a decision making tech-nique that uses the fused information to allo-cate additional robots (sensors) to rapidly clas-sify the object.

• A multi-robot task allocation (MRTA) tech-nique using a spatial queueing model that en-ables a set of robots to determine a suitable or-der or performing a set of landmine detectionrelated tasks while reducing the time and en-ergy spent in performing the tasks.

To realize the above techniques, we have cus-tomized commercially available all-terrain robotscalled Coroware Explorer with a metal detector toenable them to detect metallic objects includinglandmines. We have also developed a user inter-face that allows shared autonomy between robotsand humans. Humans can visualize informationabout the health and status of the robots and theirprogress in the landmine detection operation on acontrol station, as well as selectively supersede theirautonomous operations by remotely controlling themovement and some operations of the robots. Wehave verified the operation of the robots in differ-ent types of outdoor terrain and different operationalconditions. We have also used indoor robots calledCoroware Corobots, which have very similar featuresto the outdoor Explorer robots, both in simulationand within an indoor arena to test the differenttechniques used in COMRADES. Our results show

that our proposed techniques offer suitable meansto rapidly perform autonomous landmine detectionwith inexpensive robots.

The rest of our paper is structured as follows: Inthe next section we provide an overview of existingresearch on robotic landmine detection. In Section3, we describe the main features of our proposedsystem, the robots and the landmine detector used,and the user interface. The specific algorithms, tech-niques and results related to the three main techni-cal aspects of COMRADES - distributed area cover-age, distributed task allocation and information fu-sion are addressed in Section 4 and finally we discussfuture directions of our work and conclude.

2 Related Work

Autonomous landmine detection using robotic sys-tems has been an active research topic over the pastdecade. Excellent reviews of the state-of-the-arttechniques in robotic landmine detection are avail-able in [4, 24]. The research in this topic can be di-vided into three major directions - designing robotsattached with suitable sensor devices to detect andpossibly extract landmines, developing data and in-formation fusion techniques to improve the accuracyof detecting landmines, and, computational tech-niques to coordinate multiple robots and present theinformation collected by the robots in a structuredand visualizable format to a human supervisor.

Much of the recent research on autonomous land-mine detection has been concentrated on develop-ing robotic systems for detecting landmines; mostof these systems consist of a single robot attachedwith appropriate sensors for landmine detection. Forexample, some of these robots include a mecha-nism mounted on small robot platforms to flail theground and detonate landmines along with vegeta-tion clearing tools [26]. Many deployed robotic sys-tems for landmine detection rely on tele-operationrather than autonomy. Examples of such systemsinclude a remote operated vehicle called MR-2, theenhanced tele-operated ordnance disposal system(ETODS), TEMPEST robot, etc. These robots ben-efit from the improved precision in detecting andneutralizing landmines due to a human’s presence inthe loop, but they require humans to be in the vicin-ity of landmines to tele-operate the vehicle. Laterimprovements to some of these systems such as theMR-2 have added partial autonomy in navigation

2

and increased the tele-operation range to 5 km us-ing feedback sensors. In contrast to these largerrobots, researchers have also investigated smallerrobots that are highly agile, have a small footprintand low weight, to reduce the risk of accidental ex-plosion of a landmine. Examples of such robotsare the Ares, Shrimps, Pemexs, Dervishs and Tri-dem robots and legged robots such as the AMRU,Shadow Deminer and COMET [24]. Despite theiragility, smaller robots are limited in the weight ofsensors that they can carry on-board and are notsuited for heavier, more robust sensors like groundpenetrating radar (GPR) or large-coil metal detec-tors. To accommodate such sensors, robots such asthe Titan [13], Gryphon [19], mine detection robotand mine hunter vehicle (MHV) [24] have been de-veloped. Most of the robots discussed above are fordetecting and locating landmines. In contrast, thePEACE robot [41] and a mechanical hand called theMinehand [19] have been developed mainly to ex-cavate detected landmines. Researchers have alsoproposed unmanned aerial vehicles (UAVs) in land-mine detecting robotic systems to aid in terrain map-ping before deploying ground robots [4] or to unob-trusively detect landmines using a conceptual sys-tem of sensors attached to cables suspended fromUAVs [25].

For detecting landmines, a wide variety of sensorsincluding metal detectors (MD) [39], ground pene-trating radar (GPR) [21, 48], infra-red cameras [14]and chemical sensors for detecting plumes emanatingfrom landmines [6] have been proposed. For most re-search as well as commercially deployed applications,metal detectors, either individually or in combina-tion with GPR [19, 18] are the sensors of choice, asthey provide reasonably accurate source localizationand are relatively straightforward to acquire as off-the-shelf components and integrate on robots. Forthis reason, we have used metal detectors on therobots for detecting landmines in our system.

Yet another important aspect of landmine detec-tion is fusing sensor readings obtained by multiplesensors. The Joint Multi-Sensor Mine-Signaturesproject was one of the earliest research efforts tocollect landmine detection data using multiple sen-sors [51]. Milisavljevic et al. [8] have proposedseveral sensor fusion techniques based on Dempster-Shafer Theory including techniques for incorporatinghuman confidence values with the data collected bysensors [40]. A multi-sensor demining robot con-

sisting of metal detector, an infra-red sensor and achemical sensor is described in [47]. In [48], theauthors report that Bayesian inference techniquesfor fusing data using multiple sensors - metal de-tector, GPR, infra-red camera and magnetometer,can significantly improve the detection rate of land-mines. The Advanced Landmine Imaging System(ALIS) uses signatures from metal detectors andGPR to more accurately locate deep mines, althoughthe sensors are operated manually and their signa-tures are inspected manually as well. This idea hasbeen extended to robotic landmine detection by col-lecting data from a metal detector array and GPRmounted on a single robot and using a combinationof Bayesian inferencing and clustering algorithmsdepending upon the context of the collected data,to get improved detection rate of landmines [20].Across the world, several recent projects for human-itarian demining using robots are also utilizing mul-tiple sensors to perform landmine detection more ac-curately. The ongoing TIRAMISU project in the Eu-ropean Union [1] proposes to use multi-sensor datafusion techniques for combining information from ametal detector and a chemical sensor [46] to improvethe location and detection accuracy of landmines. In[30], authors describe field tests with the ALIS andGryphon systems while using only metal detectors,and metal detectors along with GPR, for landminesburied in different types of soil in test mine-fields inCroatia. Similarly, in [17], the author describes amechanical system equipped with nuclear detectorsfor measuring gamma radiation and backscatteredthermal neutrons, which is planned to be deployedin Libya. The challenges reported in these projectsinclude accurately localizing a landmine’s depth andmitigating the false alarm rate subject to the prob-ability of detection of the sensor used to detect thelandmine. Soil composition and clutter in the soilare also important factors affecting the accuracy oflandmine detection [30].

Recently, some researchers have proposed usingmulti-robot systems for landmine detection[30, 50,36]. The clear advantage afforded by multiple robotsis the ability to include sensors of different types ondifferent robot platforms, and, making the systemrobust against the failure of single or multiple robots.In this direction, the distributed field robot architec-ture (DFRA) [36] proposed a software framework forthe coordinated operation of an aerial and a wheeledground robot with visual and thermal sensors to de-

3

tect landmines, while in [50], the authors includedwheeled, legged and aerial robots using a combina-tion of GPR, metal detectors and vapour sensors forlandmine detection. Both these works are mentionedto be preliminary research, and mainly focus on asuitable software architecture for integrating multi-ple robots into a single system for landmine detec-tion. Our work in this paper is mainly along thisdirection of coordinating multiple robots to performlandmine detection with a focus on specific tech-niques and algorithms that can be used by the robotsfor performing various aspects of autonomous land-mine detection.

3 Description of the COMRADE

System

The central objective of the COMRADE (COop-erative Multi-Robot Automated DEtection) systemfor humanitarian demining is to develop novel coor-dination techniques between multiple low-cost, au-tonomous, mobile robots which enable them to col-laboratively detect landmines in post-conflict re-gions. The robots used in the COMRADE systemare off-the-shelf, relatively inexpensive, autonomousrobots that are equipped with appropriate sensorsfor detecting landmines. We consider three maincandidates for sensors - metal detectors (MD), IR-based multi-locator device and ground penetratingradar (GPR). The costs, accuracy and capabilitiesof the different sensors are given in Table 1. Be-cause of the differences between sensors across thesethree factors, it is important to deploy the sensorsin a region based on the possibility of existenceof landmines and risks to human lives in that re-gion. For example, in low-risk, low-incidence ar-eas, more robots with low cost/low accuracy sen-sors (e.g., MD only) can be deployed using looselycoordinated robot teams that offer very coarse guar-antees on the time required to confirm a detectedobject as a landmine. On the other hand, in a high-risk, high-incidence area, it would make sense to in-clude more accurate and more expensive sensors, us-ing tightly coupled robot teams so that a potentiallandmine could be confirmed rapidly. To achievethis in the COMRADE system, the area of interest(AOI) is classified by human experts into sub-areasbased on landmine incidence possibility and risks tohuman life, and robots with appropriate sensors aredeployed in each sub-area. An example classification

of an AOI is given in Table 2 and a diagram show-ing the deployment of robot teams based on the AOIclassification is shown in Figure 1. As shown in Fig-ure 1, for high-risk areas, the entire AOI is coveredusing all sensor types. For moderate risk areas, as atrade-off between landmine detection costs (time andenergy expended in detection) and accuracy, higheraccuracy sensors are deployed at a certain locationonly when a lower accuracy sensor has detected asuspicious object at that location. If 100% detectionis required, the entire AOI can be classified as high-risk area to ensure that it is covered at least once byevery sensor. For the sake of legibility, in the rest ofthe paper, we refer to the sub-area in which a teamof robots is deployed as their environment and con-sider algorithms for coordinating the robots that areonly within that sub-area.

We consider an initially unknown environmentwhose boundaries are known, but the locations ofobstacles within the environment are initially notknown or known with inaccuracies. The environ-ment contains landmines as well as non-landmineobjects that are buried underground and can bedetected by the landmine detection sensors on therobots. These objects are together referred to as ob-jects of interest and their locations are not knowna priori. When a robot is in the vicinity of an ob-ject of interest, we assume that the robot is appro-priately positioned, so that the object’s signaturecan be registered on its detection sensor. We re-fer to the operations performed by a robot to de-termine the signature of a buried object of interestusing its landmine detection sensor as a task. Theobjective of the robots is to search for landminesusing their landmine detection sensors. When anobject of interest is detected on a robot’s sensor, itcalls other robots to the location at which the ob-ject was detected to analyse the object’s signatureusing the other robots’ sensors. Finally, the datacollected from the object of interest by the robotshas to be fused so that the object can be classifiedas a landmine or non-landmine. To realize these ac-tivities, the robots in COMRADES have to performthree major functions: (1) Distributedly cover thefree space in the environment to search for objectsof interest. (2) For each robot, determine a sched-ule or order to perform tasks it is aware of, so thatthe cost of the overall schedule in terms of energy ex-pended by the robots is reduced. The set of tasks canchange dynamically as robots discover new objects

4

of interest. (3) Fuse the information at an object ofinterest obtained by different robots to classify it asa landmine or non-landmine. The algorithms for re-alizing these functionalities within COMRADES aredescribed in Section 4.

Figure 1: General schematic of the COMRADEsystem for landmine detection using multiple, au-tonomous, mobile robots

Table 1: Costs of different types of landmine detection

sensors in the COMRADE system.

3.1 COMRADES Robots

We have used two robots called the Explorer andCorobot, manufactured by Coroware Inc. Bothrobots are four-wheeled and use a skid-steer mech-anism for maneuvering. The Explorer robot hasrugged construction with a higher clearance and issuitable for outdoor navigation over rough terrains,while the Corobot robot is a lighter, scaled-downversion of the Explorer robot and more suited for in-door usage and testing purposes. Photographs of theExplorer and Corobot robots used in COMRADESare shown in Figure 2. Each robot is equipped withan on-board computer, an AMD Athlon Dual CoreProcessor 5050e running at 2.6 GHz with 1.87 GB

Table 2: Classification of different types of AOI in COM-

RADES depending on risk to human lives.

of RAM under Windows XP. Both robots’ on-boardsensors include a color VGA Webcam capable of 2MP resolution, Wi-Fi to communicate with the con-trol station, and a Hokuyo URG-04LX-UG01 laserrangefinder with a detection range of 2 cm to 4 mand a sweep angle of 240◦.

The Corobot robot is equipped with a HagisonicStargazer localization device that uses IR-based po-sitioning using overhead markers to determine thelocation and heading of the robot with an accuracyof ±2 cm. To avoid obstacles at close proximity, theCorobot robot is also fitted with two cross-beam IRsensors mounted on the front bumper, one IR sensoron each side, oriented sixty degrees from the front ofthe vehicle; the IR sensors provide proximity mea-surements within a range of 10 − 80 cm. For theCorobot robots, we used their Webcams to detectspecific objects or marks on the ground correspond-ing to virtual landmines due to the complicationsin using landmine detectors such as metal detectorsin indoor environments (e.g., inside buildings withmetal frames).

Additional on-board sensors on the Explorer robotinclude an inertial measurement unit (IMU) and aGarmin GPS16x LVS differential GPS that provideslocalization with an accuracy of ±3m. To localizethe robot, a Kalman filtering technique was used tocombine the GPS, IMU and encoder readings, re-sulting in a localization accuracy of ±1 m.1 TheExplorer robots were also customized with a metaldetector attached with a fixed arm to the front ofthe robot, as described below.

1Techniques to fuse localization data from multiple sen-sors [20] can be used to further improve the localization accu-racy of detected landmines.

5

(a)

(b)

Figure 2: Robots used in COMRADES along withtheir sensors. (a) Explorer robot, (b) Corobot robot.

3.2 Metal Detector System

The metal detector (MD) system used in COM-RADES consists of an Infinium LS Metal Detectormanufactured by Garrett Inc. attached to a fixed,forward facing arm on the Explorer robot. The MDis designed to work in moist and heavily mineralizedenvironments. The device has been customized andintegrated on the Explorer Robot via a USB inter-face to detect landmines over different environmentssuch as grass, snow and rock based surfaces, as showin Figure 3. The metal detector implements PulseInduction (PI) technology which works by sendingshort (50µs), high current (20A) pulses to a searchcoil. It then listens for the “echo” from a metallicobject. This echo is actually the residual magneticfield induced in an object near the coil and changesthe decay rate from the natural decay of the coil.If there is no metal near the coil this pulse will de-cay rapidly and predictably. When metal is near thecoil it will decay at a different rate than the natural

decay. The difference between those two times indi-cates the presence of metal. The slope rate of thecurve will indicate how strong the signal is from themetal. An advantage of the pulse induction technol-ogy over other comparable technologies is that it isaffected very little by mineralization in soil and wa-ter which means it can be used in a broader rangeof soil types and locations.

To verify the operation of the MD, we conducteda simple experiment - we used a mockup landminewith a very little amount of metal in its construc-tion, shown in Figure 4 (a). Figure 4 (b) shows themetal detector coil placed on the landmine when it isplaced on the surface and underground. Figures 4 (c)and (d) show the signal strength from the metal de-tector and the standard deviation error when whenthe mockup landmine is placed at different depthsbelow ground ranging from 0−100 mm, directly un-der the coil of the MD. The strength of the signalsshow that the MD used is able to detect the landmineand isolate it from non-metallic material despite itslow metal content.

Controllers for a repertoire of low-level behaviorssuch as obstacle avoidance, wandering or randomwalk, boundary following (both physical and virtualboundaries) and mine avoidance were programmedin C++ on each robot. These behaviors are utilizedfor implementing more complex coordinated opera-tions for coverage and task allocation on multiplerobots, which are described later.

Figure 3: The Explorer robot searches for landminesin different types of environments. It is possible toscan areas covered by grass, snow and rocks due tothe characteristics of the MDS.

3.3 Graphical User Interface

A graphical user interface (GUI) has been designedusing Java to interact with a team of physical robotsduring a mission of collective detection of landmines.The primary goal of the GUI is to visualize in real-time the state of a mission including the positionand current status of the robots, their available bat-tery power and the location of potential landminesdetected by robots, etc., as shown in Figure 5(a).

6

(a)

(b)

0 20 40 60 80 1000

200

400

600

800

1000

1200

Depth (mm)

Meta

l D

ete

cto

r D

ata

(R

AW

data

) 996.80

387.73

182.41 160.78122.41 104.45

(c)

0 20 40 60 80 1000

0.5

1

1.5

2

2.5

3

3.5

4

Depth (mm)

Sta

ndard

Devia

tion E

rror

3.26

3.03

3.51

3.233.06

3.28

(d)

Figure 4: (a) A real action paintball mine is usedto simulate mines made with few metallic parts, (b)Landmine has been buried 100mm in soil with thepurpose of mine detection analysis, (c) Metal detec-tor raw data for different buried depths of the paint-ball mine, (d) Standard deviation error for differentdepths at which the paintball mine is buried.

The operator can also command robots to do spe-cific actions through the GUI such as to abort its

current operation and navigate to a specific locationin the environment, recall robots to the base sta-tion, and stop and restart the robots. Figure 5(a)shows a snapshot of an experiment involving twophysical robots deployed in an indoor environment of3.5×5 square meters and the corresponding state ofthe GUI. The paths followed by robots are dynam-ically recorded and highlighted, and the landmine-like objects detected by robots are represented withred spots in the GUI.

(a)

(b)

Figure 5: (a) Snapshot of the COMRADES graphi-cal user interface. The image illustrates a panoramicof the experimental environment using two Corobotrobots. After 5 minutes of exploration the robotshave detected 4 landmines shown by red circles. Thepath followed by the robots is shown by the greentrail. (b) Communication architecture of the COM-RADE system. Dotted lines indicate wireless com-munication via Wi-Fi between the robots and therobot server at the control station.

Every two seconds each robot transmits data suchas its pose and battery state to the control sta-tion. When a landmine-like object is detected therobot sends the estimated position of such object.Landmine-like objects are represented in indoor ex-periments by red pieces of paper detectable by therobots’ camera. The communication between therobots’ server and the GUI is established through

7

TCP ports. The robot server periodically updatesthe state of robots to the GUI. Each robot receivesand transmits data to the robot server located at thecontrol station via Wi-Fi, as shown in Figure 5(b).Videos showing the operation of COMRADES alongwith its GUI are available at [2].

4 Robot Planning Techniques in

COMRADES

Planning techniques form a central part of COM-RADES to enable the deployed robots to navigateautonomously and reach objects of interests (poten-tial landmines) while avoiding obstacles as well asother robots. We consider two categories of planningin COMRADES - coverage path planning to ensurethat the robots cover the entire free space within theenvironment using their landmine detection sensorswhile searching for landmines, and, task planningor task allocation, to determine the order in whicha robot will visit the locations at which objects ofinterests have been discovered by other robots.

4.1 Distributed Terrain Coverage in

COMRADES

Coverage path planning [11] techniques enablerobots to plan their paths so that they can coverthe entire free space of their environment using theircoverage sensors. Distributed coverage with multiplerobots offers several advantages over using a singlerobot to perform coverage, such as reduced time tocomplete coverage and improved robustness againstsingle or multiple robot failures. However, a chal-lenging problem in distributed, multi-robot coverageis to ensure that different robots do not impede eachothers’ movement, or, robots with the same set ofsensors repeatedly cover the same region while leav-ing portions of the environment uncovered. Previ-ous approaches to multi-robot distributed coverageassume that the environment is decomposed into acellular or grid-like structure before deploying therobots [43, 49, 29, 53]. They then use graph traver-sal techniques to completely cover the environment.Robots send messages to each other with their cov-erage information to ensure that they cover disjointregions. In our research, we have used Voronoi par-titioning [3] to divide the free space in the environ-ment into disjoint cells or regions. The robots’ initialpositions are used as the sites for generating the par-

tition. After partitioning, each robot is responsiblefor covering the Voronoi region it is situated in. Thisremoves the overhead for avoiding repeated coverageand collision between robots. One issue with Voronoipartitioning is that the size of the Voronoi regions aredependent on the initial positions of the robots anda bad initial spatial distribution of the robots (e.g.,robots being very close to each other) might resultin disproportionate regions. To alleviate this situa-tion, we propose to use simple dispersion strategies[7] between robots to achieve a well-spaced initialspatial distribution.

4.1.1 Voronoi Partition Coverage (VPC) Al-gorithm

The main contribution of our research on cover-age in COMRADES is an algorithm called VoronoiPartition-based Coverage (VPC). The detailed de-scription of the algorithm is avaialable in [32], herewe provide an overview of the key aspects of thealgorithm. In the VPC algorithm, each robot firstpartitions the environment into a set of disjoint re-gions given by the Voronoi partition of the environ-ment, using the robots’ initial positions [12, 16]. Therobots in the adjacent Voronoi regions of a robot arecalled its Voronoi neighbors. In [23], we have de-scribed a fully distributed technique for computingthe Voronoi partition using communication betweena set of robots. Once the partitions are determined,each robot proceeds to cover the region in which itis situated. For this, a robot decomposes the freespace of its region using a grid-like cellular decompo-sition; each cell of the grid corresponds to four timesfootprint size of the robot’s coverage tool (landminedetection sensor).

The robot then uses a cellular coverage techniquecalled Spanning Tree Coverage (STC) [54] to coverthe cells2. The STC algorithm allows the robotto cover successive cells in its direction of motion.When an obstacle is encountered, the robot selectsa previously uncovered cell from the neighbors ofits current cell, while checking the neighbors in aclockwise direction from its current cell. If no freeneighboring cell is found (e.g., robot is in a cave),the robot backtracks to the previous cell from whichit had arrived at the current cell. The STC algo-rithm terminates when the robot reaches its start

2Although we have used STC for implementing our algo-rithm, any other single-robot coverage algorithm can be usedin conjunction with VPC

8

(a) (b) (c)

Figure 6: 4 robots marked by blue circles cover a) a 20m × 20m square environment, and b) a X-shapedregion within this environment using VPC algorithm. (c) One robot fails in the same setting as in scenario(a), and an operational robot takes over coverage of the failed robot’s Voronoi region.

cell by backtracking. When the robot completes cov-erage of its Voronoi region, it broadcasts a coveragecompletion message to its Voronoi neighbors. Weverified the operation of the VPC algorithm usingCorobot robots on the Webots simulator. Figures6(a) and 6(b) show snapshots from the simulationusing 4 robots in a 20× 20 m2 environment with noobstacles and an X-shaped obstacle respectively; thecoverage paths followed by the different robots aremarked with colored trails.

The VPC algorithm is also robust to failure of in-dividual or a few robots. To ensure robustness, eachrobot periodically exchanges alive messages with itsVoronoi neighbors. If a non-responsive neighbor thathas not completed its coverage is discovered, it ismarked as a failed robot. The Voronoi regions arerecomputed while discarding the failed robot and theVoronoi neighbors of the failed robot then proceedto cover their new Voronoi regions after excludingalready covered portions in the region. Figure 6(c)shows when the robot with the purple trail fails, oneof its Voronoi neighbors (robot with red trail) takesover and finishes coverage of its unfinished Voronoiregion. This process ensures that the entire environ-ment will be covered as long as at least one robotremains operational. We have also shown analyti-cally that the VPC algorithm ensures complete, non-overlapping coverage provided the single-robot cov-erage algorithm achieves complete non-overlappingcoverage [32].

4.1.2 Repartitioning Coverage Algorithm

In COMRADES, because the environment is initiallyunknown, we assume that the robots are not aware

(a) (b)

Figure 7: (a) The Voronoi cells of two robots are partiallyinaccessible due to obstacles. The blue solid arrows show

the path taken by a robot to reach the inaccessible por-

tions of its cell using a bug-like path planning algorithm.

(b) Robots coordinate with each other to repartition the

initial Voronoi cells so that each robot has a contiguous

region to cover.

of the location and geometry of obstacles. A po-tential problem while using the VPC algorithm insuch a scenario is that a robot might discover thata portion of its Voronoi region is occluded by obsta-cles, as shown in Figure 7(a). The robot then has touse path planning techniques to find a path to reachthe inaccessible portions of its Voronoi region; suchpath planning technique can involve complex com-putations [11] and increase the robots’ time and en-ergy requirements. To avoid excessive path planningcosts, we investigated an approach that adaptivelyrepartitions regions where coverage is impeded byobstacles and reallocates the repartitioned portionsto other robots that can complete the coverage, as

9

(a) (b) (c)

Figure 8: Snapshots from Webots showing repartition coverage by 7 robots in a 3×6 m2 environment with different

obstacle features, (a) initial Voronoi partition, (b) robots performing boundary coverage on Voronoi cell, black/light

blue boundaries show inaccesible regions. (c) repartitioned cells and robots completing coverage of entire environment.

shown in Figure 7(b).At the core of our repartitioning approach is an al-

gorithm called Repartition Coverage [28]. The maininsight of this algorithm is that even if the Voronoiregion that a robot is covering gets disconnected dueto obstacles, because the free space is connected, theinaccessible portion of the region must be adjacentto at least one of the neighboring regions and acces-sible to the robot in that region. Consequently, therobot performing coverage in the adjacent neighbor-ing region could be requested to augment its cover-age with the inaccessible portion of the disconnectedregion, as shown in Figure 7(b). While using theVPC algorithm, if a robot determines that it cannotreach portions of its coverage region due to obsta-cles, it uses an auction-based protocol to systemat-ically repartition and reallocate the inaccessible re-gions to other robots. We have shown analyticallythat the Repartition Coverage algorithm guaranteescomplete, non-overlapping coverage and that it con-verges to termination within a finite number of steps,determined by the number of robots in the environ-ment. The performance of the algorithm was alsoverified in Webots for different environments anddifferent obstacle geometries to verify its complete-ness and coverage times for environments of differentsizes and different obstacle geometries. Some of theresults are shown in Figure 8.

4.2 Multi-Robot Task Allocation in

COMRADES

In COMRADES, when robots have detected objectsof interest they request other robots, possibly withother types of sensors, to visit the location of theobject and inspect it with their sensors. Conse-quently, each robot might receive requests to visitobjects of interest at different locations from multi-ple robots. To avoid expending excessive energy and

time to visit these locations and analyze the object,the robots need to determine a suitable itinerary forvisiting the locations. Multi-robot task allocation(MRTA) techniques provide a structured method tosolve this problem - how to find a suitable assignmentof tasks to robots so that the tasks performed by therobots can be completed in an efficient manner interms of time and energy expended by the robots.We consider a category of MRTA problems calledST-MR-TA (single task robot, multi-robot tasks,time extended assignment) [5], where ST stands forsingle-task robots, i.e., each robot is able to executeas most one task at a time, MR means multi-robottasks, tasks that require multiple robots to be com-pleted, and TA means time-extended assignment,problems where the information to allocate tasks torobots arrives over time. A task in the COMRADESscenario corresponds to a robot visiting the locationof a potential landmine (not necessarily at the sametime as other robots) to analyze the object using therobots’ sensors. The location of potential landminesarrives dynamically as they are detected using thecoverage techniques described in Section 4.1. MRTAin such a scenario corresponds to the multiple travel-ing salesman problem (mTSP) that has been shownto be NP-hard [44]. Previous work in MRTA forST-MR-TA problem considers local or market-basedheuristics [22, 15, 35]. In COMRADES, we have useda stochastic queueing-based technique to address theMRTA problem [52, 10]. Using spatial queueing isattractive for our ST-MR-TA MRTA problem as itprovides a formal framework for distributed decisionmaking by the robots so that they can respond effi-ciently to dynamic changes in the task distribution.In the next section, we give an overview of a spa-tial queuing based MRTA algorithm used in COM-RADES; details of the algorithm along with exten-sive simulation results are available in [42, 34].

10

4.2.1 Spatial Queueing for Multi-RobotTask Allocation

To motivate our MRTA problem we consider an au-tomated landmine detection scenario where a setof robots are deployed within a bounded 2D envi-ronment with potential landmines. The location ofthe landmines is not known a priori. Robots areequipped with sensors that are capable of detectinglandmine-like objects, albeit within a certain levelof uncertainty due to sensor noise. Robots initiallyexplore the environment and when a robot finds anobject of interest that could potentially be a land-mine, it requests other robots, possibly with differ-ent sensor types to visit the location of the detectionand confirm the object on their sensors. Within thissetting, a task corresponds to a set of robots visit-ing the location of an object of interest and recordingthe object’s signature on their sensors. For legibility,we have referred to each robot’s visit to the object’slocation and taking its reading, as the robot per-forming its portion of the task. Robots can performa task asynchronously by performing their portion ofthe task at different times. A task is considered tobe complete when the desired number of robots haveperformed their portion of the task. Finally, taskscan arrive dynamically as robots explore the envi-ronment and find objects of interest. Within thiscontext, the MRTA problem corresponds to findinga suitable allocation of tasks to robots so that thetotal time required to complete the tasks is reduced.

The MRTA problem described above correspondsto the MR-ST-TA setting [22], where MR (multi-robot task) denotes that multiple robots are requiredto complete a task, ST (single task robot) denoteseach robot can perform a single task at a time andTA (time-extended assignment) denotes that eachrobot can determine and update its schedule or or-der of tasks to perform over a finite time window, asopposed to determining the task schedule instanta-neously.

We consider a set of mobile robots R = {ri : i =1, 2, ...,m} that are deployed within a bounded en-vironment E ⊂ ℜ2. We assume that each robot iscapable of localizing itself with respect to the en-vironment and its pose at any instant is given byρri . The environment also contains a set of tasksT = {τi : i = 1, 2, ..., n} that are distributed arbi-trarily within the environment; the location of taskτi is denoted by ρτi . Robot and task positions areinitially shared between the robots and assumed to

be common knowledge. The objective of the robotsis to visit the location of each task and performcertain operations related to the task. We assumethat task τi requires operations to be performed bydτi ≤ |R| distinct robots to be completed. Becausethe focus of this paper is on the task allocation algo-rithm, we assume that techniques for appropriatelypositioning the robots to perform operations at thelocations of the tasks are already available. The dis-tance between two tasks, τi and τj , is denoted bydij = ‖ρτi −ρτj‖ while the distance between robot ri

and task τj is denoted by dij = ‖ρri − ρτj‖. Also, welet ρ0i denote the initial position of robot ri and τr1idenote the first task selected by robot ri. Within thissetting the MRTA problem can be formally definedas the following:Definition. Multi-robot Task Allocation. Given a

set of robots R and a set of tasks T find a suitableallocation A : 2R → T such that, ∀R ⊆ R : ri ∈ R:

min∑

ri∈R

‖ρ0ri − ρτr1i

‖+∑

(τj ,τk)∈A(R)

‖ρτj − ρτk‖

,

subject to:

τj 6= τk ∀(τj, τk) ∈ A(R), ∀R ⊆ R : ri ∈ R,∀ri

|A(R)| = dτi A(R) = τi, R ⊆ R,∀τi

The above formulation of the MRTA problem at-tempts to find an allocation for each robot such thatthe distances traveled by the robots to perform thetasks is minimized. The two constraints of the prob-lem ensure that the same task does not get allocatedto the same robot more than once, and, the totalnumber of robots allocated to perform a task equalsthe demand for the task.The solution to the MRTA problem has been

shown to be an instance of the dynamic travelingsalesman problem and proven to be NP-hard [45].In this paper, we propose spatial queue-based [9]MRTA solution technique that attempts to attemptsthe MRTA problem using a heuristic that representsthe distances between robots and tasks as an orderedqueue based on the robots’ locations and inter-taskdistances. To achieve this is in a systematic manner,each robot utilizes the following four mathematicalconstructs:

1. Inter-task Transition Matrix. Inter-task dis-tances form the basis of our method as the

11

objective of the MRTA technique is to enablerobots find a suitable schedule or order of navi-gating between tasks so that the total distancetraveled by them is reduced. The inter-task dis-tances are represented as a transition matrix.The transition matrix at time-step t is denotedby M(t) and given by the normalized inverseEuclidean distances between every task pair, asshown in Equation 1:

M(t) =

π11 π12 ... π1nπ21 π22 ... π2n

...

πn1 πn2 ... πnn

(1)

where πi,j =1

di,j∑j 6=i

1

di,j

Each entry πij of M(t) represents the proba-bility of a robot to select task τj following τi,based on the distance between the tasks’ loca-tions. Note that πii = 0 and therefore the di-agonal elements of the matrix are zeros. Thetransition matrix values are calculated indepen-dently by all robots. Initially, the transitionmatrix is computed for all task pairs, but astime proceeds, each robot recalculates the ma-trix when it completes a task, or when it re-ceives information that a task has been com-pleted by other robots. The transition probabil-ities of completed tasks are set to zero and theprobability values in M(t) are re-normalized.

2. Robot State Vector. The state vector of a robotcomprises of the inverse Euclidean distances be-tween the robot and each task in the environ-ment. The state vector for robot i, Vri at time-step t is given by:

Vri(t) = (πi1(t), πi2(t), ..., πin(t)) (2)

where πij(t) =1

di,j(t)and di,j(t) is the distance

between robot ri and task τj at time-step t.

3. Task Proximity Vector. The task proximity vec-tor of a robot ri represents its preference foreach task τj in the environment based on per-forming task τj first followed by the remainingtasks. It is calculated as the product of therobot ri’s state vector and the inter-task tran-sition matrix. The task proximity vector forrobot ri at time-step t, Vri(t), is given by:

Vri(t) = Vri(t)×M(t) (3)

4. Robot Spatial Task Queue. The spatial taskqueue of a robot denotes the order in which therobot plans to perform the tasks in the environ-ment. It is calculated by removing all tasks fromthe task proximity vector that are either occu-pied or completed, and sorting the remainingtasks in descending order based on their prox-imity vector values. The task queue for robotri at time-step t, Qri(t), given by:

Qri(t) = {q1, q2, ..., qn : qk ≥ qk+1∀k, qk ∈ Vri(t)}(4)

5 10 15 20

200

400

600

800

1000

Tim

eto

co

mp

lete

24

tas

ks

(se

c)

No. of robots

DG RA SQ HA

6 tasks 12 tasks 18 tasks 24 tasks

400

200

600

800

0

Figure 9: Completion times with fixed task load of24 tasks for 5, 10, 15, and 20 robots (top) and withfixed number of 20 robots fir 6, 12, 18 and 24 tasks(bottom) for the compared MRTA approaches.

Robots use the spatial queueing framework to se-lect tasks using the Spatial Queueing MRTA (SQ-MRTA) algorithm. In the SQ-MRTA algorithm,each robot sorts the available tasks based usingEquations 1-3. The robot then selects the task at

12

0

1

2

Rati

o o

f sim

ula

tio

n t

imes o

f d

iffe

ren

t alg

ori

thm

sw

. r.

t. sim

ula

tio

n t

ime o

f H

un

gari

an

alg

ori

thm

DG Algorithm

RA AlgorithmSQ-MRTA Algorithm

5R, 6

T

5R, 1

8T5R

, 24T

10R, 6

T

10R, 1

2T10

R, 1

8T10

R, 2

4T15

R, 6

T15

R, 1

2T15

R, 1

8T15

R, 2

4T20

R, 6

T20

R, 1

2T20

R, 1

8T20

R, 2

4T

5R, 1

2T

Robot-task combinations

Figure 10: Competitive ratio of simulation times us-ing the Hungarian method as the baseline. Differ-ent robot and task numbers’ combinations used areshown on the x-axis.

the head of the spatial queue and announces a bidfor that task based on its cost (distance) to performthat task. It then waits for a certain time period toreceive bids for the same task from other robots. If itis the highest bidder and the task is still available, itproceeds to perform the task; otherwise it selects thenext available task from its spatial queue and repeatsthe bidding process. When it finishes performing thetask, the robot broadcasts a task performed message.When robots receive a task performed message fromanother robot that results in the task being com-pleted (sufficient number of robots have visited thetask), they rebuild their local copy of the transitionmatrix.

We verified the performance of the SQ-MRTA(SQ) algorithm and compared it with three state-of-the-art MRTA algorithms - the Hungarian as-signment (HA) based algorithm [33], a decentral-ized greedy (DG) allocation algorithm [42] and therepeated auctions (RA) algorithm [37]. As before,the algorithm was implemented on Corobot robotswithin the Webots simulator. We used 5, 10, 15, or20 robots with 6, 12, 18, or 24 tasks within a 20× 20m2 environment. Each task was required to be per-formed by 3−5 robots. All results were averaged over10 simulation runs. We evaluated different metricsincluding the total time required to complete tasksand the average distance traveled per robot. Twoillustrative graphs of our simulations are shown inFigures 9 and 10. In Figure 9 we see that the SQ-MRTA algorithm performs comparably with the re-

peated auctions (RA) algorithm. For more combina-tions of robots and tasks, Figure 10 shows that ourSQ-MRTA algorithm performs very closely in com-parison to the repeated auctions (RA) algorithm,with their simulation times lying between ±10% ofeach other. On the other hand, the HA algorithmperforms poorly as the numbers of robots and tasksincreases because it assigns robots to tasks basedon the initial placement of robots and tasks; delaysin robots reaching tasks due to inter-robot collisionavoidance are not considered by it while determiningthe robot to task assignments. The DG algorithm isinefficient in terms of task completion times whenthe number of tasks greatly outweigh the number ofrobots as it allocates the closest task to a robot andis unable to calculate a suitable schedule when eachrobot needs to perform multiple tasks.

4.3 Information Aggregation Techniques

in COMRADES

2A central aspect of multi-robot autonomous land-mine detection is to combine the information aboutthe characteristics of a potential landmine from dif-ferent types of sensors and make a decision whetherthe object is indeed a landmine, and, identify itscharacteristics, if it is indeed one. Previous re-searchers [40] have considered this problem from astatic viewpoint where all information about a land-mine’s characteristics is assumed to be available andthe main concept is to use statistical inference tech-niques to classify the landmine’s characteristics withaccuracy. However, in COMRADES, the process oflandmine detection is not an instantaneous one; itcontinues over a period of time during which robotswith appropriate sensors, corresponding to a po-tential landmine’s initially perceived characteristics,need to be deployed to the potential landmine’s lo-cation so that the cumulative information gatheredby the robots’ sensors can improve the accuracy ofthe landmine’s detection. Because of this dynamicnature of landmine detection, we consider the follow-ing multi-sensor information aggregation and sensorscheduling problem in COMRADES - given an initialsignature perceived by a certain type of sensor froma potential landmine, what is an appropriate set ofsensors (robots) to deploy additionally to the loca-tion of the potential landmine so that the landmineis detected with higher accuracy. Details of the oper-ation of the information aggregation technique alongwith analytical and experimental results of its per-

13

formance are given in [31]; we provide an overviewof its main features and a few significant results inthe next section.

sch

ed

ulin

gd

ecis

ion

Belief in object

being landmine

(calculated

using Bayesian

inference,

condi!oned on

past beliefs)

Report

calculated

using exp.

u!lity

maximiza!on

Aggrega!on

mechanism

using market‐

based scoring

rule

Robot/

Sensor Scheduling

Algorithm

Decision

Maker Agent

ba,t

ra1,t

Bt

Agg. belief

(from last

!me step)

Predic!on Market

Weightage of reports human

from sensor expert based on

environment & opera!onal

condi!ons

Object

(landmine)

Exp. Reward

Strategy

Sensed

data

Belief

Report

Weight

Reports from

other agents

(sensors)

Sensor Agent 1 Market Maker Agent

Bayesian

inference

and exp.

u!lity

Maximiz‐n

Agg.

Belief

Figure 11: Schematic of the prediction market-basedinformation aggregation technique used to combinereports from multiple sensors in COMRADES.

To solve the information aggregation and sensorscheduling problem in COMRADES, we have pro-posed a novel technique that uses a market-basedinformation aggregation mechanism called a predic-tion market. Each robot participating in the land-mine detection task is provided with a software agentthat uses the sensory input of the robot from a po-tential landmine and performs the calculations of theprediction market technique. A schematic describ-ing the technique is shown in Figure 11. When anagent records readings from a potential landmine onits sensors, it associates a probability value, called abelief, ba,t, to denote the agent’s confidence in iden-tifying the sensed object as a landmine. The be-liefs are conditioned over past belief values to pre-vent wide variations from the object’s previous read-ings due to sensor noise or ambient conditions usinga Bayesian network. Each agent then strategicallycalculates a sensor report, ra,t, from its belief andexpected rewards from making the report, using autility maximization technique. It then submits thisreport to a central location called the aggregatoror market maker agent. The information aggrega-tion or fusion is implemented using the aggregationmechanism that uses a technique called a logarithmicmarket scoring rule (LMSR). This technique uses autility-based formulation of the costs and value toeach sensor (robot) for identifying an object of inter-est as a landmine. It then assigns a score or virtualreward to each sensor (robot) if it correctly reportsthe probability distribution over the different typesof objects of interests such as landmine, metal butnot landmine, and not landmine. The details of this

technique are given in [31]. The aggregation mech-anism and outputs a single aggregated belief value,Bt. The aggregation mechanism also selectively in-cludes weights of the sensor reports from a humanexpert about the accuracy of the sensors’ reportsbased on the ambient conditions of the sensors. Theaggregated belief value is then passed on to a deci-sion maker agent that makes decisions about whichother robots (sensor types) should be deployed to thepotential landmine’s location using a Bayesian infer-encing based technique, so that the landmine can beconfirmed rapidly and accurately [31].

Mine Metallic Object Non−metallic Object0

10

20

30

40

50

60

70

80

90

100

110

Object Type

Cos

t

MDIRGPR

0 1 2 3 4 5 6 7 8 9 100

0.1

0.2

0.3

0.4

0.5

Number of Time Steps

Ave

rage

RM

SE

MD

GPR

IR

(a) (b)

Figure 12: (a) Cost to classify different types of ob-jects - mines, metal but not mine and non-metal us-ing a different types of sensors. Relative costs ofsensing using MD, IR and GPR were assumed tobe in the ratio of 1 : 2 : 4. (b) Root mean squareerror from different types of sensors from differenttypes of sensors when used individually over time atthe same object of interest. Note that although us-ing MDs along has a low cost, their RMSE is higher(accuracy is lower).

We simulated our algorithm using three differenttypes of sensors - MD, GPR and and IR-based multi-sensor device. We performed different experimentswith data from different types of sources (metallic,low-metallic, non-metallic) collected by different sen-sors at different times. A few experimental resultsof our algorithm while using identical data distri-butions and settings are highlighted in Figures 12 -Figure 14. Figure 12 shows the relative costs of de-ploying the sensors and the corresponding root meansquare errors (RMSE) from the readings when usingone type of sensor. Figure 13(a)-(d) shows the effectof deploying multiple sensors of different types overtime. For these experiments, the data was assumedto arrive from the same source object of interest. Dif-ferent sets of sensors were deployed over 7 time steps.

14

0 1 2 3 40.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

0.55

Number of Time Steps

Ave

rage

RM

SE MD

0 1 2 3 4 50

0.1

0.2

0.3

0.4

0.5

Number of Time Steps

Ave

rage

RM

SE

MD

GPR

0 1 2 3 4 5 6 7 80

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

Number of Time Steps

Ave

rage

RM

SE

MD

IR

GPR

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50

0.1

0.2

0.3

0.4

0.5

0.6

Number of Time Steps

Ave

rage

RM

SE

MD

IR

GPR

a b c d

Figure 13: Average RMSE in the environment with 5 MD sensors(a), 5 MD and 1 GPR sensor(b), 5 MD,1 IR, and 1 GPR sensors(c), 2 MD, 2 IR, and 2 GPR sensors(d).

0 2 4 6 8 100

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

Number of Time Steps

RM

SE

PM

D−S

DDF

0 2 4 6 8 10−35

−30

−25

−20

−15

−10

−5

0

5

Number of Time Steps

NM

SE

PM

D−S

DDF

0 2 4 6 8 10−0.16

−0.14

−0.12

−0.1

−0.08

−0.06

−0.04

−0.02

0

Number of Time Steps

Info

rmat

ion

gain

PM

D−S

DDF

(a) (b) (c)

Figure 14: Comparsion between our prediction market-based information aggregation and Dempster Shafertheory based fusion(DS) and Distributed Data Fusion (DDF) (a) Root mean squre error (b) Normalizedroot mean square error and (c) Information gain. The results show that over time, the PM-based techniqueis able to perform better than the compared techniques.

We notice that while using 5 MDs (in Figure 13(a))the RMSE reduces to 20% in 4 time-steps, but witha combination of MD, IR and GPR (in Figure 13(d))the combined RMSE from the fused data from thesesensors is reduced much more, to around 6%. Fi-nally, Figures 14 (a)-(c) illustrate the comparison ofour proposed prediction market based techniques forinformation aggregation with two techniques. Theresults show that the information fusion performedusing our technique reduces the RMSE by 5−13% ascompared to a previously studied technique for land-mine data fusion using the Dempster-Shafer theory[40] and by 3−8% using distributed data fusion tech-nique [38]. We also conducted several experimentsto test the effect of various parameters in our model,and we found that using the combination of differentsensors in the environment gives the best accuracyfor the object’s type identification.

5 Conclusions and Future Direc-

tion

In this paper, we have described our experience withthe COMRADE system for autonomous landminedetection using multiple robots. We have customizeda Coroware Explorer robot with a metal detectorto detect landmines and performed tests with it ondifferent terrains. We have also addressed variousaspects of landmine detection such as multi-robotsearch and coverage, multi-robot task allocation andmulti-sensor data fusion using different algorithmsand validated our results using simulated indoorCorobot robots.

As future work, we are looking at several direc-tions to improve our proposed techniques. Integra-tion of a wider suite of sensor devices such as ther-mal sensors, GPRs and chemical sensors on outdoor

15

robot platforms and test the correct combinationof sensors for different ambient conditions such asambient temperature, sunlight, depth of landmines,etc. is an ongoing work in our research. We arealso investigating ways to improve our coverage al-gorithm by sharing maps between robots, planningcoverage paths based on expected information gainabout landmines and improved information fusiontechniques for combining data from multiple sensors.Finally, we are also looking at using aerial robots toaid ground moving robots navigate more intelligentlyin the environment. Overall, we envisage that ourresearch will lay the foundation for further researchand proliferation of multi-robot systems for land-mine detection, nuclear source detection, unmannedsearch and rescue, emergency response services andother high-risk applications.

Acknowledgements. The COMRADES projectwas supported by the US Office of Naval Research,grant no. N000140911174.

References

[1] http://www.fp7-tiramisu.eu/. [Online; accessed01-January-2015].

[2] http://cmantic.unomaha.edu/projects/comrades/.[Online; accessed 14-August-2014].

[3] Okabe A, Boots B, Sugihara K, and Chiu S. SpatialTessellations: Concepts and Applications of VoronoiDiagrams. Wiley& Sons;, 2000.

[4] M. Acheroy. Mine action: status of sensor tech-nolo.gy for close-in and remote detection of an-tipersonnel mines. In Advanced Ground PenetratingRadar, 2005. IWAGPR 2005. Proceedings of the 3rdInternational Workshop on, pages 3–13, May 2005.

[5] Gerkey B and Mataric M. A formal analysis andtaxonomy of task-allocation in multi-robot systems.The International Journal of Robotics Research,23(9):939–954, 2004.

[6] S. Badia, U. Bernardet, A. Guanella, P. Pyk, andP. Verschure. A biologically based chemo-sensingUAV for humanitarian demining. InternationalJournal of Advanced Robotic Systems., 4(2):187–198,2007.

[7] Maxim A. Batalin and Gaurav S. Sukhatme. Spread-ing out: A local approach to multi-robot coverage. InProc. of 6th International Symposium on DistributedAutonomous Robotic Systems, pages 373–382, 2002.

[8] I. Bloch, N. Milisavljeniv, and M. Acheroy. Mul-tisensor Data Fusion for Spaceborne and Airborne

Reduction of Mine Suspected Areas. InternationalJournal of Advanced Robotic Systems., 4(2):173–186,2007.

[9] L. Breuer. Spatial queues with infinitely manyservers. In From Markov Jump Processes to SpatialQueues, pages 119–138. Springer Netherlands, 2003.

[10] Francesco Bullo, Emilio Frazzoli, Marco Pavone, Ke-tan Savla, and Stephen L. Smith. Dynamic vehiclerouting for robotic systems. Proceedings of the IEEE,99(9):1482–1504, 2011.

[11] H. Choset, W. Burgard, S. Hutchinson, G. Kantor,Lydia E. Kavraki, K. Lynch, and S. Thrun. Princi-ples of Robot Motion: Theory, Algorithms, and Im-plementation. MIT Press, June 2005.

[12] Jorge Cortes, Sonia Martınez, Timur Karatas, andFrancesco Bullo. Coverage control for mobile sens-ing networks. IEEE T. Robotics and Automation,20(2):243–255, 2004.

[13] Paulo Debenest, Edwardo F. Fukushima, Yuki Tojo,and Shigeo Hirose. A new approach to humanitariandemining. Auton. Robots, 18(3):323–336, 2005.

[14] Eric den Breejen, Klamer Schutte, and Frank Cre-mer. Sensor fusion for antipersonnel landmine detec-tion: a case study. volume 3710, pages 1235–1245,1999.

[15] M.B. Dias, Robert Zlot, N. Kalra, and A Stentz.Market-based multirobot coordination: A surveyand analysis. Proc. IEEE, 94(7):1257–1270, 2006.

[16] Joseph W. Durham, Ruggero Carli, Paolo Frasca,and Francesco Bullo. Discrete partitioning and cov-erage control for gossiping robots. IEEE Transac-tions on Robotics, 28(2):364–378, 2012.

[17] FA El BAKKOUSH. Current research activities forlandmine detection by nuclear technique in libya.2009.

[18] Xuan Feng and M. Sato. Landmine imaging by ahand-held gpr and metal detector sensor (alis). InGeoscience and Remote Sensing Symposium, Proc.IEEE International, volume 1, July 2005.

[19] Marc Freese, Paulo Debenest, Edwardo F.Fukushima, and Shigeo Hirose. Developmentof deminer-assisting robotic tools at tokyo in-stitute of technology. In Maki K Habib, editor,Humanitarian Demining. Intech Open, 2008.

[20] Hichem Frigui, Lijun Zhang, Paul Gader, Joseph NWilson, KC Ho, and Andres Mendez-Vazquez. Anevaluation of several fusion algorithms for anti-tanklandmine detection and discrimination. InformationFusion, 13(2):161–174, 2012.

16

[21] T. Fukuda, Y. Hasegawa, Y. Kawai, S. Sato,Z. Zyada, and T. Matsuno. GPR signal processingwith geography adaptive scanning using vector radarfor antipersonal landmine detection. InternationalJournal of Advanced Robotic Systems., 4(2):199–206,2007.

[22] Brian P. Gerkey and Maja J. Mataric. A formalanalysis and taxonomy of task allocation in multi-robot systems. I. J. Robotic Res., 23(9):939–954,2004.

[23] K. R. Guruprasad and Prithviraj Dasgupta. Dis-tributed voronoi partitioning for multi-robot systemswith limited range sensors. In IROS, pages 3546–3552, 2012.

[24] M. Habib. Humanitarian demining: Reality and thechallenge of technology - The state of the arts. In-ternational Journal of Advanced Robotic Systems.,4(2):151–172, 2007.

[25] Stefan Havlik. Land robotic vehicles for demining.In Maki K Habib, editor, Humanitarian Demining.Intech Open, 2008.

[26] Stefan Havlik. Some robotic approaches and tech-nologies for humanitarian demining. In Maki KHabib, editor, Humanitarian Demining. IntechOpen, 2008.

[27] M. Hiznay. Landmine monitor report. Technical re-port, 2011.

[28] Kurt Hungerford, Prithviraj Dasgupta, and K. R.Guruprasad. Distributed, complete, multi-robot cov-erage of initially unknown environments using repar-titioning. In AAMAS, pages 1453–1454, 2014.

[29] Rekleitis I, New A, Rankin E, and Choset H. Ef-ficient boustrophedon multi-robot coverage: an al-gorithmic approach. Annals of Math and ArtificialIntelligence, 52:109–142, 2008.

[30] Jun Ishikawa, Katsuhisa Furuta, and NikolaPavkovic. Test and evaluation of japanese gpr-emidual sensor systems at the benkovac test site in croa-tia. In Katsuhisa Furuta and Jun Ishikawa, editors,Anti-personnel Landmine Detection for Humanitar-ian Demining, pages 63–81. Springer London, 2009.

[31] Jumadinova J and Dasgupta P. Multirobot au-tonomous landmine detection using distributed mul-tisensor information aggregation. In Proceedings ofSPIE Conference on Multisensor, Multisource Infor-mation Fusion: Architectures, Algorithms, and Ap-plications 8407, pages 1–12, 2012.

[32] Guruprasad K, Wilson Z, and Dasgupta P. Completecoverage of an initially unknown environment bymultiple robots using voronoi partition. In Proceed-ings of 2nd International Conference on Advancesin Control and Optimization of Dynamical Systems(ACODS), Bangalore, India, 2012.

[33] H. W. Kuhn. The Hungarian method for the assign-ment problem. Naval Research Logistics Quarterly,2(1-2):83–97, 1955.

[34] W. Lenagh. Multi-robot task allocation: A spatialqueueing approach. Master’s thesis, University ofNebraska, Omaha, 2013.

[35] L. Liu and D. Shell. A distributable andcomputation-flexible assignment algorithm: Fromlocal task swapping to global optimality. Proceed-ings of Robotics Science and Systems, pages 33–41,2012.

[36] M. Long, A Gage, R. Murphy, and K. Valavanis. Ap-plication of the distributed field robot architecture toa simulated demining task. In Robotics and Automa-tion, Proc. IEEE International Conference on, pages3193–3200, April 2005.

[37] Lingzhi Luo, Nilanjan Chakraborty, and Katia P.Sycara. Distributed algorithm design for multi-robotgeneralized task assignment problem. In IROS, pages4765–4771, 2013.

[38] J. Manyika and H. Durrant-Whyte. Data Fusion andSensor Management: A Decentralized Information-Theoretic Approach. Prentice Hall, Upper SaddleRiver, NJ, USA, 1995.

[39] S. Masunaga and K. Nonami. Controlled metaldetector mounted on mine detection robot. In-ternational Journal of Advanced Robotic Systems,4(2):207 – 218, 2007.

[40] Nada Milisavljevic and Isabelle Bloch. Sensor fusionin anti-personnel mine detection using a two-level be-lief function model. IEEE Transactions on Systems,Man, and Cybernetics, Part C, 33(2):269–283, 2003.

[41] Yoshikazu Mori. Peace: An excavation-type dem-ining robot for anti-personnel mines. In Maki KHabib, editor, Humanitarian Demining. IntechOpen, 2008.

[42] Angelica Munoz-Melendez, Pritviraj Dasgupta, andWilliam Lenagh. A stochastic queueing model formulti-robot task allocation. In ICINCO (1), pages256–261, 2012.

[43] Hazon N and Kaminka G. On redundancy, ef-ficiency, and robustness in coverage for mulitplerobots. Robotics and Autonomous Systems, 56:1102–1114, 2008.

[44] P. Oberlin, S. Rathinam, and S. Darbha. A trans-formation for a heterogeneous, multi-depot, multi-ple traveling salesman problem. In Proceedings ofthe American Control Conference, pages 1292–1297,2009.

17

[45] Dasgupta P. Multi-robot task allocation for per-forming cooperative foraging tasks in an initially un-known environment. In L Jain, editor, Innovations inDefense Support Systems - 2, pages 5–20. Springer-Verlag.

[46] J. Prado, G. Cabrita, and L. Marques. Bayesiansensor fusion for land-mine detection using a dual-sensor hand-held device. In Industrial ElectronicsSociety, IECON 2013 - 39th Annual Conference ofthe IEEE, pages 3887–3892, 2013.

[47] Michael Rachkov, Lino Marques, and Anibal T.de Almeida. Multisensor demining robot. Au-tonomous Robots, 18(3):275–291, 2005.

[48] C. Ratto, P. Torrione, K. Morton, and L. Collins.Context-dependent landmine detection with ground-penetrating radar using a hidden markov contextmodel. In Geoscience and Remote Sensing Sym-posium (IGARSS), 2010 IEEE International, pages4192–4195, July 2010.

[49] Hert S and Lumelsky V. Polygon area decompositionfor multiplerobot workspace division. InternationalJournal of Computational Geometry & Applications,8:437–466, 1998.

[50] Pedro Santana, Jose Barata, H. Cruz, A. Mestre,J. Lisboa, and Luıs Flores. A multi-robot system forlandmine detection. In ETFA, 2005.

[51] P. Verlinde, M. Acheroy, G. Nesti, and A Sieber.Preparing the joint multi-sensor mine-signaturesproject database for data fusion. In Geoscience andRemote Sensing Symposium, IEEE International,volume 7, pages 3240–3242, 2001.

[52] Huang X and Serfozo R. Spatial queueing processes.Mathematics of Operations Research, 24(4):865–886,1999.

[53] Zheng X, Koenig S, Kempe D, and Jain S. Multi-robot forest coverage for weighted and unweightedterrain. IEEE Trans on Robotics, 26(6):1018–1031,2010.

[54] Gabriely Y and Rimon E. Spanning-tree based cov-erage of continuous areas by a mobile robot. Annalsof Math and Artificial Intelligence, 31:77–98, 2001.

18