The COMRADE System for Multi- 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, USA 2Mechanical Engineering Department, NIT, Karnataka, India 3Computer Science Department, INAOE, Mexico 4Computer Science Department, Allegheny College, PA, USA

Abstract 1 Introduction

We consider the problem of autonomous landmine Humanitarian demining is a crucial effort for the detection using a team of mobile . Previous safety and sustenance of human lives in post-conflict research on robotic landmine detection mostly regions. Unfortunately, recent surveys on landmine employs a single robot equipped with a landmine monitoring report that humanitarian demining ef- detection sensor to detect landmines. We envisage forts are considerably lagging behind anti-personnel that the quality of landmine detection can be signif- landmine planting activities due to several techno- icantly improved if multiple robots are coordinated logical and economic reasons [27]. This results in to detect landmines in a cooperative manner by enormous loss to human lives; e.g., in 2010 alone, incrementally fusing the landmine-related sensor explosions of landmines and similar devices resulted information they collect and to then use that in- in 4, 191 casualties, with civilians accounting for 70% formation to visit locations of potential landmines. of the casualties. One the major technological chal- Towards this objective, we describe a multi-robot lenges in humanitarian demining is to detect land- system called COMRADES to address different mines rapidly and with reasonable accuracy, while aspects of the autonomous landmine detection prob- reducing the number of false positives. We envis- lem including distributed area coverage to detect age that automating landmine detection operations and locate landmines, information aggregation to using multiple, off-the-shelf autonomous robots will fuse the sensor information obtained by different provide a reasonably accurate yet economical solu- robots, and, multi-robot task allocation (MRTA) tion to the problem of detecting landmines. Towards to enable different robots determine a suitable this objective, we describe a multi-robot system sequence to visit locations of potential landmines called COMRADE (COoperative Multi-Robot Au- while reducing the time required and battery tomated DEtection) System for humanitarian dem- expended. We have used commercially available ining. The central objective of COMRADES is to all-terrain robots called Coroware Explorer that are develop novel coordination techniques between mul- customized with a metal detector to detect metallic tiple low-cost, mobile robots, which enable them to objects including landmines, as well as indoor autonomously and collaboratively detect landmines Corobot robots, both in simulation and in physical with high accuracy in post-conflict regions. COM- experiments, to test the different techniques in RADES includes techniques that allows each robot COMRADES. to explore an initially unknown region while search- ing for landmines, recognize landmine-like objects on its sensors, share and fuse the landmine-related sen- Keywords: robotic landmine detection: coverage sor information with other robots and coordinate its and exploration; sensor information fusion; multi- actions with other robots, so that multiple robots robot task allocation can converge on the object to analyze and confirm it as a landmine. In this paper, we present the that our proposed techniques offer suitable means description and experimental results from different to rapidly perform autonomous landmine detection techniques for coverage, task allocation, and, multi- with inexpensive robots. sensor information aggregation and sensor schedul- The rest of our paper is structured as follows: In ing using multiple robots, that we have developed as the next section we provide an overview of existing part of COMRADES. Specifically, we describe the research on robotic landmine detection. In Section following aspects of multi-robot autonomous land- 3, we describe the main features of our proposed mine detection in COMRADES: system, the robots and the landmine detector used, and the user interface. The specific algorithms, tech- • A distributed area coverage technique that al- niques and results related to the three main techni- low a set of robots to dynamically partition an cal aspects of COMRADES - distributed area cover- initially unknown environment into a set of non- age, distributed task allocation and information fu- overlapping regions and search for landmines sion are addressed in Section 4 and finally we discuss within each region. The techniques are robust future directions of our work and conclude. to individual robot failures and are able to scale with the number of robots and size of the envi- ronment. 2 Related Work

• A distributed information fusion technique to Autonomous landmine detection using robotic sys- aggregate landmine-related sensor information tems has been an active research topic over the past from different robots using a prediction market decade. Excellent reviews of the state-of-the-art based technique and a decision making tech- techniques in robotic landmine detection are avail- nique that uses the fused information to allo- able in [4, 24]. The research in this topic can be di- cate additional robots (sensors) to rapidly clas- vided into three major directions - designing robots sify the object. attached with suitable sensor devices to detect and • A multi-robot task allocation (MRTA) tech- possibly extract landmines, developing data and in- nique using a spatial queueing model that en- formation fusion techniques to improve the accuracy ables a set of robots to determine a suitable or- of detecting landmines, and, computational tech- der or performing a set of landmine detection niques to coordinate multiple robots and present the related tasks while reducing the time and en- information collected by the robots in a structured ergy spent in performing the tasks. and visualizable format to a human supervisor. Much of the recent research on autonomous land- To realize the above techniques, we have cus- mine detection has been concentrated on develop- tomized commercially available all-terrain robots ing robotic systems for detecting landmines; most called Coroware Explorer with a metal detector to of these systems consist of a single robot attached enable them to detect metallic objects including with appropriate sensors for landmine detection. For landmines. We have also developed a user inter- example, some of these robots include a mecha- face that allows shared autonomy between robots nism mounted on small robot platforms to flail the and humans. Humans can visualize information ground and detonate landmines along with vegeta- about the health and status of the robots and their tion clearing tools [26]. Many deployed robotic sys- progress in the landmine detection operation on a tems for landmine detection rely on tele-operation control station, as well as selectively supersede their rather than autonomy. Examples of such systems autonomous operations by remotely controlling the include a remote operated vehicle called MR-2, the movement and some operations of the robots. We enhanced tele-operated ordnance disposal system have verified the operation of the robots in differ- (ETODS), TEMPEST robot, etc. These robots ben- ent types of outdoor terrain and different operational efit from the improved precision in detecting and conditions. We have also used indoor robots called neutralizing landmines due to a human’s presence in Coroware Corobots, which have very similar features the loop, but they require humans to be in the vicin- to the outdoor Explorer robots, both in simulation ity of landmines to tele-operate the vehicle. Later and within an indoor arena to test the different improvements to some of these systems such as the techniques used in COMRADES. Our results show MR-2 have added partial autonomy in navigation

2 and increased the tele-operation range to 5 km us- sisting of metal detector, an infra-red sensor and a ing feedback sensors. In contrast to these larger chemical sensor is described in [47]. In [48], the robots, researchers have also investigated smaller authors report that Bayesian inference techniques robots that are highly agile, have a small footprint for fusing data using multiple sensors - metal de- and low weight, to reduce the risk of accidental ex- tector, GPR, infra-red camera and magnetometer, plosion of a landmine. Examples of such robots can significantly improve the detection rate of land- are the Ares, Shrimps, Pemexs, Dervishs and Tri- mines. The Advanced Landmine Imaging System dem robots and legged robots such as the AMRU, (ALIS) uses signatures from metal detectors and Shadow Deminer and COMET [24]. Despite their GPR to more accurately locate deep mines, although agility, smaller robots are limited in the weight of the sensors are operated manually and their signa- sensors that they can carry on-board and are not tures are inspected manually as well. This idea has suited for heavier, more robust sensors like ground been extended to robotic landmine detection by col- penetrating radar (GPR) or large-coil metal detec- lecting data from a metal detector array and GPR tors. To accommodate such sensors, robots such as mounted on a single robot and using a combination the Titan [13], Gryphon [19], mine detection robot of Bayesian inferencing and clustering algorithms and mine hunter vehicle (MHV) [24] have been de- depending upon the context of the collected data, veloped. Most of the robots discussed above are for to get improved detection rate of landmines [20]. detecting and locating landmines. In contrast, the Across the world, several recent projects for human- PEACE robot [41] and a mechanical hand called the itarian demining using robots are also utilizing mul- Minehand [19] have been developed mainly to ex- tiple sensors to perform landmine detection more ac- cavate detected landmines. Researchers have also curately. The ongoing TIRAMISU project in the Eu- proposed unmanned aerial vehicles (UAVs) in land- ropean Union [1] proposes to use multi-sensor data mine detecting robotic systems to aid in terrain map- fusion techniques for combining information from a ping before deploying ground robots [4] or to unob- metal detector and a chemical sensor [46] to improve trusively detect landmines using a conceptual sys- the location and detection accuracy of landmines. In tem of sensors attached to cables suspended from [30], authors describe field tests with the ALIS and UAVs [25]. Gryphon systems while using only metal detectors, For detecting landmines, a wide variety of sensors and metal detectors along with GPR, for landmines including metal detectors (MD) [39], ground pene- buried in different types of soil in test mine-fields in trating radar (GPR) [21, 48], infra-red cameras [14] Croatia. Similarly, in [17], the author describes a and chemical sensors for detecting plumes emanating mechanical system equipped with nuclear detectors from landmines [6] have been proposed. For most re- for measuring gamma radiation and backscattered search as well as commercially deployed applications, thermal neutrons, which is planned to be deployed metal detectors, either individually or in combina- in Libya. The challenges reported in these projects tion with GPR [19, 18] are the sensors of choice, as include accurately localizing a landmine’s depth and they provide reasonably accurate source localization mitigating the false alarm rate subject to the prob- and are relatively straightforward to acquire as off- ability of detection of the sensor used to detect the the-shelf components and integrate on robots. For landmine. Soil composition and clutter in the soil this reason, we have used metal detectors on the are also important factors affecting the accuracy of robots for detecting landmines in our system. landmine detection [30]. Yet another important aspect of landmine detec- Recently, some researchers have proposed using tion is fusing sensor readings obtained by multiple multi-robot systems for landmine detection[30, 50, sensors. The Joint Multi-Sensor Mine-Signatures 36]. The clear advantage afforded by multiple robots project was one of the earliest research efforts to is the ability to include sensors of different types on collect landmine detection data using multiple sen- different robot platforms, and, making the system sors [51]. Milisavljevic et al. [8] have proposed robust against the failure of single or multiple robots. several sensor fusion techniques based on Dempster- In this direction, the distributed field robot architec- Shafer Theory including techniques for incorporating ture (DFRA) [36] proposed a software framework for human confidence values with the data collected by the coordinated operation of an aerial and a wheeled sensors [40]. A multi-sensor demining robot con- ground robot with visual and thermal sensors to de-

3 tect landmines, while in [50], the authors included of an AOI is given in Table 2 and a diagram show- wheeled, legged and aerial robots using a combina- ing the deployment of robot teams based on the AOI tion of GPR, metal detectors and vapour sensors for classification is shown in Figure 1. As shown in Fig- landmine detection. Both these works are mentioned ure 1, for high-risk areas, the entire AOI is covered to be preliminary research, and mainly focus on a using all sensor types. For moderate risk areas, as a suitable software architecture for integrating multi- trade-off between landmine detection costs (time and ple robots into a single system for landmine detec- energy expended in detection) and accuracy, higher tion. Our work in this paper is mainly along this accuracy sensors are deployed at a certain location direction of coordinating multiple robots to perform only when a lower accuracy sensor has detected a landmine detection with a focus on specific tech- suspicious object at that location. If 100% detection niques and algorithms that can be used by the robots is required, the entire AOI can be classified as high- for performing various aspects of autonomous land- risk area to ensure that it is covered at least once by mine detection. every sensor. For the sake of legibility, in the rest of the paper, we refer to the sub-area in which a team 3 Description of the COMRADE of robots is deployed as their environment and con- sider algorithms for coordinating the robots that are System only within that sub-area.

The central objective of the COMRADE (COop- We consider an initially unknown environment erative Multi-Robot Automated DEtection) system whose boundaries are known, but the locations of for humanitarian demining is to develop novel coor- obstacles within the environment are initially not dination techniques between multiple low-cost, au- known or known with inaccuracies. The environ- tonomous, mobile robots which enable them to col- ment contains landmines as well as non-landmine laboratively detect landmines in post-conflict re- objects that are buried underground and can be gions. The robots used in the COMRADE system detected by the landmine detection sensors on the are off-the-shelf, relatively inexpensive, autonomous robots. These objects are together referred to as ob- robots that are equipped with appropriate sensors jects of interest and their locations are not known for detecting landmines. We consider three main a priori. When a robot is in the vicinity of an ob- candidates for sensors - metal detectors (MD), IR- ject of interest, we assume that the robot is appro- based multi-locator device and ground penetrating priately positioned, so that the object’s signature radar (GPR). The costs, accuracy and capabilities can be registered on its detection sensor. We re- of the different sensors are given in Table 1. Be- fer to the operations performed by a robot to de- cause of the differences between sensors across these termine the signature of a buried object of interest three factors, it is important to deploy the sensors using its landmine detection sensor as a task. The in a region based on the possibility of existence objective of the robots is to search for landmines of landmines and risks to human lives in that re- using their landmine detection sensors. When an gion. For example, in low-risk, low-incidence ar- object of interest is detected on a robot’s sensor, it eas, more robots with low cost/low accuracy sen- calls other robots to the location at which the ob- sors (e.g., MD only) can be deployed using loosely ject was detected to analyse the object’s signature coordinated robot teams that offer very coarse guar- using the other robots’ sensors. Finally, the data antees on the time required to confirm a detected collected from the object of interest by the robots object as a landmine. On the other hand, in a high- has to be fused so that the object can be classified risk, high-incidence area, it would make sense to in- as a landmine or non-landmine. To realize these ac- clude more accurate and more expensive sensors, us- tivities, the robots in COMRADES have to perform ing tightly coupled robot teams so that a potential three major functions: (1) Distributedly cover the landmine could be confirmed rapidly. To achieve free space in the environment to search for objects this in the COMRADE system, the area of interest of interest. (2) For each robot, determine a sched- (AOI) is classified by human experts into sub-areas ule or order to perform tasks it is aware of, so that based on landmine incidence possibility and risks to the cost of the overall schedule in terms of energy ex- human life, and robots with appropriate sensors are pended by the robots is reduced. The set of tasks can deployed in each sub-area. An example classification change dynamically as robots discover new objects

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

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-board sensors include a color VGA Webcam capable of 2 MP resolution, Wi-Fi to communicate with the con- trol station, and a Hokuyo URG-04LX-UG01 laser rangefinder with a detection range of 2 cm to 4 m and a sweep angle of 240◦. Figure 1: General schematic of the COMRADE The Corobot robot is equipped with a Hagisonic system for landmine detection using multiple, au- Stargazer localization device that uses IR-based po- tonomous, mobile robots sitioning using overhead markers to determine the location and heading of the robot with an accuracy of ±2 cm. To avoid obstacles at close proximity, the Corobot robot is also fitted with two cross-beam IR sensors mounted on the front bumper, one IR sensor on each side, oriented sixty degrees from the front of the vehicle; the IR sensors provide proximity mea- surements within a range of 10 − 80 cm. For the Corobot robots, we used their Webcams to detect specific objects or marks on the ground correspond- Table 1: Costs of different types of landmine detection ing to virtual landmines due to the complications sensors in the COMRADE system. in using landmine detectors such as metal detectors in indoor environments (e.g., inside buildings with metal frames). 3.1 COMRADES Robots Additional on-board sensors on the Explorer robot include an inertial measurement unit (IMU) and a We have used two robots called the Explorer and Garmin GPS16x LVS differential GPS that provides Corobot, manufactured by Coroware Inc. Both localization with an accuracy of ±3m. To localize robots are four-wheeled and use a skid-steer mech- the robot, a Kalman filtering technique was used to anism for maneuvering. The Explorer robot has combine the GPS, IMU and encoder readings, re- rugged construction with a higher clearance and is sulting in a localization accuracy of ±1 m.1 The suitable for outdoor navigation over rough terrains, Explorer robots were also customized with a metal while the Corobot robot is a lighter, scaled-down detector attached with a fixed arm to the front of version of the Explorer robot and more suited for in- the robot, as described below. door usage and testing purposes. Photographs of the Explorer and Corobot robots used in COMRADES are shown in Figure 2. Each robot is equipped with 1 Techniques to fuse localization data from multiple sen- an on-board computer, an AMD Athlon Dual Core sors [20] can be used to further improve the localization accu- Processor 5050e running at 2.6 GHz with 1.87 GB racy of detected landmines.

5 decay. The difference between those two times indi- cates the presence of metal. The slope rate of the curve will indicate how strong the signal is from the metal. An advantage of the pulse induction technol- ogy over other comparable technologies is that it is affected very little by mineralization in soil and wa- ter which means it can be used in a broader range of soil types and locations. To verify the operation of the MD, we conducted a simple experiment - we used a mockup landmine (a) with a very little amount of metal in its construc- tion, shown in Figure 4 (a). Figure 4 (b) shows the metal detector coil placed on the landmine when it is placed 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 when the mockup landmine is placed at different depths below ground ranging from 0 − 100 mm, directly un- der the coil of the MD. The strength of the signals show that the MD used is able to detect the landmine and isolate it from non-metallic material despite its low metal content. Controllers for a repertoire of low-level behaviors such as obstacle avoidance, wandering or random (b) walk, boundary following (both physical and virtual boundaries) and mine avoidance were programmed Figure 2: Robots used in COMRADES along with in C++ on each robot. These behaviors are utilized their sensors. (a) Explorer robot, (b) Corobot robot. for implementing more complex coordinated opera- tions for coverage and task allocation on multiple robots, which are described later. 3.2 Metal Detector System

The metal detector (MD) system used in COM- RADES consists of an Infinium LS Metal Detector manufactured by Garrett Inc. attached to a fixed, forward facing arm on the Explorer robot. The MD Figure 3: The Explorer robot searches for landmines is designed to work in moist and heavily mineralized in different types of environments. It is possible to environments. The device has been customized and scan areas covered by grass, snow and rocks due to integrated on the Explorer Robot via a USB inter- the characteristics of the MDS. face to detect landmines over different environments such as grass, snow and rock based surfaces, as show in Figure 3. The metal detector implements Pulse 3.3 Graphical User Interface Induction (PI) technology which works by sending short (50µs), high current (20A) pulses to a search A graphical user interface (GUI) has been designed coil. It then listens for the “echo” from a metallic using Java to interact with a team of physical robots object. This echo is actually the residual magnetic during a mission of collective detection of landmines. field induced in an object near the coil and changes The primary goal of the GUI is to visualize in real- the decay rate from the natural decay of the coil. time the state of a mission including the position If there is no metal near the coil this pulse will de- and current status of the robots, their available bat- cay rapidly and predictably. When metal is near the tery power and the location of potential landmines coil it will decay at a different rate than the natural detected by robots, etc., as shown in Figure 5(a).

6 current operation and navigate to a specific location in 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 two physical robots deployed in an indoor environment of 3.5 × 5 square meters and the corresponding state of (a) the GUI. The paths followed by robots are dynam- ically recorded and highlighted, and the landmine- like objects detected by robots are represented with red spots in the GUI.

(b)

1200

996.80 1000

800

600

387.73 400

182.41 200 160.78 (a) 122.41

Metal Detector Data (RAW data) Metal Detector Data (RAW 104.45

0 0 20 40 60 80 100 Depth (mm) (c)

4

3.51 3.5 3.26 3.23 3.28 3.03 3.06 3

2.5 (b) 2

1.5 Figure 5: (a) Snapshot of the COMRADES graphi- cal user interface. The image illustrates a panoramic 1 Standard Deviation Error of the experimental environment using two Corobot 0.5 robots. After 5 minutes of exploration the robots 0 have detected 4 landmines shown by red circles. The 0 20 40 60 80 100 Depth (mm) path followed by the robots is shown by the green (d) trail. (b) Communication architecture of the COM- RADE system. Dotted lines indicate wireless com- Figure 4: (a) A real action paintball mine is used munication via Wi-Fi between the robots and the to simulate mines made with few metallic parts, (b) robot server at the control station. Landmine has been buried 100mm in soil with the purpose of mine detection analysis, (c) Metal detec- Every two seconds each robot transmits data such tor raw data for different buried depths of the paint- as its pose and battery state to the control sta- ball mine, (d) Standard deviation error for different tion. When a landmine-like object is detected the depths at which the paintball mine is buried. robot sends the estimated position of such object. Landmine-like objects are represented in indoor ex- periments by red pieces of paper detectable by the The operator can also command robots to do spe- robots’ camera. The communication between the cific actions through the GUI such as to abort its robots’ server and the GUI is established through

7 TCP ports. The robot server periodically updates tition. After partitioning, each robot is responsible the state of robots to the GUI. Each robot receives for covering the Voronoi region it is situated in. This and transmits data to the robot server located at the removes the overhead for avoiding repeated coverage control station via Wi-Fi, as shown in Figure 5(b). and collision between robots. One issue with Voronoi Videos showing the operation of COMRADES along partitioning is that the size of the Voronoi regions are with its GUI are available at [2]. dependent on the initial positions of the robots and a bad initial spatial distribution of the robots (e.g., robots being very close to each other) might result 4 Robot Planning Techniques in in disproportionate regions. To alleviate this situa- COMRADES tion, we propose to use simple dispersion strategies [7] between robots to achieve a well-spaced initial Planning techniques form a central part of COM- spatial distribution. RADES to enable the deployed robots to navigate autonomously and reach objects of interests (poten- 4.1.1 Voronoi Partition Coverage (VPC) Al- tial landmines) while avoiding obstacles as well as gorithm other robots. We consider two categories of planning in COMRADES - coverage path planning to ensure The main contribution of our research on cover- that the robots cover the entire free space within the age in COMRADES is an algorithm called Voronoi environment using their landmine detection sensors Partition-based Coverage (VPC). The detailed de- while searching for landmines, and, task planning scription of the algorithm is avaialable in [32], here or task allocation, to determine the order in which we provide an overview of the key aspects of the a robot will visit the locations at which objects of algorithm. In the VPC algorithm, each robot first interests have been discovered by other robots. partitions the environment into a set of disjoint re- gions given by the Voronoi partition of the environ- 4.1 Distributed Terrain Coverage in ment, using the robots’ initial positions [12, 16]. The COMRADES robots in the adjacent Voronoi regions of a robot are called its Voronoi neighbors. In [23], we have de- Coverage path planning [11] techniques enable scribed a fully distributed technique for computing robots to plan their paths so that they can cover the Voronoi partition using communication between the entire free space of their environment using their a set of robots. Once the partitions are determined, coverage sensors. Distributed coverage with multiple each robot proceeds to cover the region in which it robots offers several advantages over using a single is situated. For this, a robot decomposes the free robot to perform coverage, such as reduced time to space of its region using a grid-like cellular decompo- complete coverage and improved robustness against sition; each cell of the grid corresponds to four times single or multiple robot failures. However, a chal- footprint size of the robot’s coverage tool (landmine lenging problem in distributed, multi-robot coverage detection sensor). is to ensure that different robots do not impede each The robot then uses a cellular coverage technique others’ movement, or, robots with the same set of called Spanning Tree Coverage (STC) [54] to cover sensors repeatedly cover the same region while leav- the cells2. The STC algorithm allows the robot ing portions of the environment uncovered. Previ- to cover successive cells in its direction of motion. ous approaches to multi-robot distributed coverage When an obstacle is encountered, the robot selects assume that the environment is decomposed into a a previously uncovered cell from the neighbors of cellular or grid-like structure before deploying the its current cell, while checking the neighbors in a robots [43, 49, 29, 53]. They then use graph traver- clockwise direction from its current cell. If no free sal techniques to completely cover the environment. neighboring cell is found (e.g., robot is in a cave), Robots send messages to each other with their cov- the robot backtracks to the previous cell from which erage information to ensure that they cover disjoint it had arrived at the current cell. The STC algo- regions. In our research, we have used Voronoi par- rithm terminates when the robot reaches its start titioning [3] to divide the free space in the environ- 2 Although we have used STC for implementing our algo- ment into disjoint cells or regions. The robots’ initial rithm, any other single-robot coverage algorithm can be used positions are used as the sites for generating the par- in 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-shaped region 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 coverage completion message to its Voronoi neighbors. We verified the operation of the VPC algorithm using Corobot robots on the Webots simulator. Figures 6(a) and 6(b) show snapshots from the simulation using 4 robots in a 20 × 20 m2 environment with no obstacles and an X-shaped obstacle respectively; the coverage paths followed by the different robots are marked with colored trails. (a) (b) The VPC algorithm is also robust to failure of in- Figure 7: (a) The Voronoi cells of two robots are partially dividual or a few robots. To ensure robustness, each inaccessible due to obstacles. The blue solid arrows show robot periodically exchanges alive messages with its the path taken by a robot to reach the inaccessible por- Voronoi neighbors. If a non-responsive neighbor that tions of its cell using a bug-like path planning algorithm. has not completed its coverage is discovered, it is (b) Robots coordinate with each other to repartition the marked as a failed robot. The Voronoi regions are initial Voronoi cells so that each robot has a contiguous recomputed while discarding the failed robot and the region to cover. Voronoi neighbors of the failed robot then proceed to cover their new Voronoi regions after excluding already covered portions in the region. Figure 6(c) shows when the robot with the purple trail fails, one of its Voronoi neighbors (robot with red trail) takes of the location and geometry of obstacles. A po- over and finishes coverage of its unfinished Voronoi tential problem while using the VPC algorithm in region. This process ensures that the entire environ- such a scenario is that a robot might discover that ment will be covered as long as at least one robot a portion of its Voronoi region is occluded by obsta- remains operational. We have also shown analyti- cles, as shown in Figure 7(a). The robot then has to cally that the VPC algorithm ensures complete, non- use path planning techniques to find a path to reach overlapping coverage provided the single-robot cov- the inaccessible portions of its Voronoi region; such erage algorithm achieves complete non-overlapping path planning technique can involve complex com- coverage [32]. putations [11] and increase the robots’ time and en- ergy requirements. To avoid excessive path planning costs, we investigated an approach that adaptively 4.1.2 Repartitioning Coverage Algorithm repartitions regions where coverage is impeded by In COMRADES, because the environment is initially obstacles and reallocates the repartitioned portions unknown, we assume that the robots are not aware to 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). time to visit these locations and analyze the object, At the core of our repartitioning approach is an al- the robots need to determine a suitable itinerary for gorithm called Repartition Coverage [28]. The main visiting the locations. Multi-robot task allocation insight of this algorithm is that even if the Voronoi (MRTA) techniques provide a structured method to region that a robot is covering gets disconnected due solve this problem - how to find a suitable assignment to obstacles, because the free space is connected, the of tasks to robots so that the tasks performed by the inaccessible portion of the region must be adjacent robots can be completed in an efficient manner in to at least one of the neighboring regions and acces- terms of time and energy expended by the robots. sible to the robot in that region. Consequently, the We consider a category of MRTA problems called robot performing coverage in the adjacent neighbor- ST-MR-TA (single task robot, multi-robot tasks, ing region could be requested to augment its cover- time extended assignment) [5], where ST stands for age with the inaccessible portion of the disconnected single-task robots, i.e., each robot is able to execute region, as shown in Figure 7(b). While using the as most one task at a time, MR means multi-robot VPC algorithm, if a robot determines that it cannot tasks, tasks that require multiple robots to be com- reach portions of its coverage region due to obsta- pleted, and TA means time-extended assignment, cles, it uses an auction-based protocol to systemat- problems where the information to allocate tasks to ically repartition and reallocate the inaccessible re- robots arrives over time. A task in the COMRADES gions to other robots. We have shown analytically scenario corresponds to a robot visiting the location that the Repartition Coverage algorithm guarantees of a potential landmine (not necessarily at the same complete, non-overlapping coverage and that it con- time as other robots) to analyze the object using the verges to termination within a finite number of steps, robots’ sensors. The location of potential landmines determined by the number of robots in the environ- arrives dynamically as they are detected using the ment. The performance of the algorithm was also coverage techniques described in Section 4.1. MRTA verified in Webots for different environments and in such a scenario corresponds to the multiple travel- different obstacle geometries to verify its complete- ing salesman problem (mTSP) that has been shown ness and coverage times for environments of different to be NP-hard [44]. Previous work in MRTA for sizes and different obstacle geometries. Some of the ST-MR-TA problem considers local or market-based results are shown in Figure 8. heuristics [22, 15, 35]. In COMRADES, we have used a stochastic queueing-based technique to address the 4.2 Multi-Robot Task Allocation in MRTA problem [52, 10]. Using spatial queueing is COMRADES attractive for our ST-MR-TA MRTA problem as it provides a formal framework for distributed decision In COMRADES, when robots have detected objects making by the robots so that they can respond effi- of interest they request other robots, possibly with ciently to dynamic changes in the task distribution. other types of sensors, to visit the location of the In the next section, we give an overview of a spa- object and inspect it with their sensors. Conse- tial queuing based MRTA algorithm used in COM- quently, each robot might receive requests to visit RADES; details of the algorithm along with exten- objects of interest at different locations from multi- sive simulation results are available in [42, 34]. ple robots. To avoid expending excessive energy and

10 4.2.1 Spatial Queueing for Multi-Robot be common knowledge. The objective of the robots Task Allocation is to visit the location of each task and perform certain operations related to the task. We assume To motivate our MRTA problem we consider an au- that task τi requires operations to be performed by tomated landmine detection scenario where a set dτ ≤ |R| distinct robots to be completed. Because of robots are deployed within a bounded 2D envi- i the focus of this paper is on the task allocation algo- ronment with potential landmines. The location of rithm, we assume that techniques for appropriately the landmines is not known a priori. Robots are positioning the robots to perform operations at the equipped with sensors that are capable of detecting locations of the tasks are already available. The dis- landmine-like objects, albeit within a certain level tance between two tasks, τi and τj, is denoted by of uncertainty due to sensor noise. Robots initially dij = ρτi − ρτj while the distance between robot ri explore the environment and when a robot finds an ˆ and task τj is denoted by dij = ρri − ρτj . Also, we object of interest that could potentially be a land- 0 let ρi denote the initial position of robot ri and τr1 mine, it requests other robots, possibly with differ- i denote the first task selected by robot r . Within this ent sensor types to visit the location of the detection i setting the MRTA problem can be formally defined and confirm the object on their sensors. Within this as the following: setting, a task corresponds to a set of robots visit- Definition. Multi-robot Task Allocation. Given a ing the location of an object of interest and recording set of robots R and a set of tasks T find a suitable the object’s signature on their sensors. For legibility, allocation A : 2R → T such that, ∀R ⊆ R : r ∈ R: we have referred to each robot’s visit to the object’s i location and taking its reading, as the robot per-  0  forming its portion of the task. Robots can perform min ρ − ρτ 1 + ρτ − ρτ , ri r j k ∈ i a task asynchronously by performing their portion of ri R  (τj ,τk)∈A(R)  the task at different times. A task is considered to be complete when the desired number of robots have subject to: performed their portion of the task. Finally, tasks τj = τk ∀(τj,τk) ∈ A(R), ∀R ⊆ R : ri ∈ R, ∀ri can arrive dynamically as robots explore the envi- |A(R)| = d A(R)= τ , R ⊆ R, ∀τ ronment and find objects of interest. Within this τi i i context, the MRTA problem corresponds to finding a suitable allocation of tasks to robots so that the The above formulation of the MRTA problem at- total time required to complete the tasks is reduced. tempts to find an allocation for each robot such that The MRTA problem described above corresponds the distances traveled by the robots to perform the to the MR-ST-TA setting [22], where MR (multi- tasks is minimized. The two constraints of the prob- robot task) denotes that multiple robots are required lem ensure that the same task does not get allocated to complete a task, ST (single task robot) denotes to the same robot more than once, and, the total each robot can perform a single task at a time and number of robots allocated to perform a task equals TA (time-extended assignment) denotes that each the demand for the task. robot can determine and update its schedule or or- The solution to the MRTA problem has been der of tasks to perform over a finite time window, as shown to be an instance of the dynamic traveling opposed to determining the task schedule instanta- salesman problem and proven to be NP-hard [45]. neously. In this paper, we propose spatial queue-based [9] We consider a set of mobile robots R = {ri : i = MRTA solution technique that attempts to attempts 1, 2, ..., m} that are deployed within a bounded en- the MRTA problem using a heuristic that represents 2 vironment E ⊂ ℜ . We assume that each robot is the distances between robots and tasks as an ordered capable of localizing itself with respect to the en- queue based on the robots’ locations and inter-task vironment and its pose at any instant is given by distances. To achieve this is in a systematic manner, ρri . The environment also contains a set of tasks each robot utilizes the following four mathematical T = {τi : i = 1, 2, ..., n} that are distributed arbi- constructs: trarily within the environment; the location of task

τi is denoted by ρτi . Robot and task positions are 1. Inter-task Transition Matrix. Inter-task dis- initially shared between the robots and assumed to tances form the basis of our method as the

11 objective of the MRTA technique is to enable ´ robots find a suitable schedule or order of navi- Vri (t)= Vri (t) × M(t) (3) gating between tasks so that the total distance traveled by them is reduced. The inter-task dis- 4. Robot Spatial Task Queue. The spatial task tances are represented as a transition matrix. queue of a robot denotes the order in which the The transition matrix at time-step t is denoted robot plans to perform the tasks in the environ- by M(t) and given by the normalized inverse ment. It is calculated by removing all tasks from Euclidean distances between every task pair, as the task proximity vector that are either occu- shown in Equation 1: pied or completed, and sorting the remaining tasks in descending order based on their prox- imity vector values. The task queue for robot π11 π12 ... π1n   r t Q t π π ... π i at time-step , ri ( ), given by: M(t)= 21 22 2n (1)  ...  ´   Qri (t)= {q1,q2, ..., qn : qk ≥ qk+1∀k,qk ∈ Vri (t)}  πn πn ... πnn  1 2 (4) 1 di,j where πi,j = 1 P = j6 i di,j DG RA SQ HA 1000 Each entry πij of M(t) represents the proba- bility of a robot to select task τj following τi, 800 based on the distance between the tasks’ loca-

tions. Note that πii = 0 and therefore the di- 600 agonal elements of the matrix are zeros. The transition matrix values are calculated indepen- 400 dently by all robots. Initially, the transition Time to200 complete 24 tasks (sec) matrix is computed for all task pairs, but as time proceeds, each robot recalculates the ma- 5 10 15 20 trix when it completes a task, or when it re- No. of robots ceives information that a task has been com- 800 pleted by other robots. The transition probabil- ities of completed tasks are set to zero and the probability values in M(t) are re-normalized. 600

2. Robot State Vector. The state vector of a robot comprises of the inverse Euclidean distances be- 400 tween the robot and each task in the environ-

ment. The state vector for robot i, Vri at time- step t is given by: 200

V (t)=(ˆπ (t), πˆ (t), ..., πˆ (t)) (2) ri i1 i2 in 0 6 tasks 12 tasks 18 tasks 24 tasks 1 whereπ ˆij(t) = and dˆi,j(t) is the distance dˆi,j (t) between robot ri and task τj at time-step t. Figure 9: Completion times with fixed task load of 24 tasks for 5, 10, 15, and 20 robots (top) and with 3. Task Proximity Vector. The task proximity vec- fixed number of 20 robots fir 6, 12, 18 and 24 tasks tor of a robot ri represents its preference for (bottom) for the compared MRTA approaches. each task τj in the environment based on per- forming task τj first followed by the remaining Robots use the spatial queueing framework to se- tasks. It is calculated as the product of the lect tasks using the Spatial Queueing MRTA (SQ- robot ri’s state vector and the inter-task tran- MRTA) algorithm. In the SQ-MRTA algorithm, sition matrix. The task proximity vector for each robot sorts the available tasks based using ´ robot ri at time-step t, Vri (t), is given by: Equations 1-3. The robot then selects the task at

12 peated auctions (RA) algorithm. For more combina- DG Algorithm tions of robots and tasks, Figure 10 shows that our 2 SQ-MRTA algorithm performs very closely in com- parison to the repeated auctions (RA) algorithm, with their simulation times lying between ±10% of each other. On the other hand, the HA algorithm

1 performs poorly as the numbers of robots and tasks

SQ-MRTA Algorithm increases because it assigns robots to tasks based RA Algorithm on the initial placement of robots and tasks; delays in robots reaching tasks due to inter-robot collision

0 avoidance are not considered by it while determining w. r. t. simulation time of Hungarian algorithm Hungarian of time simulation t. r. w. Ratio of simulation times of different algorithms different of times simulation of Ratio the robot to task assignments. The DG algorithm is 5R, 6T 5R, 12T5R, 18T 15R, 6T 20R, 6T 5R, 24T10R, 10R,6T 12T10R, 18T10R, 24T 15R, 12T15R, 18T15R, 24T 20R, 12T20R, 18T20R, 24T inefficient in terms of task completion times when Robot-task combinations the number of tasks greatly outweigh the number of Figure 10: Competitive ratio of simulation times us- robots as it allocates the closest task to a robot and ing the Hungarian method as the baseline. Differ- is unable to calculate a suitable schedule when each ent robot and task numbers’ combinations used are robot needs to perform multiple tasks. shown on the x-axis. 4.3 Information Aggregation Techniques in COMRADES the head of the spatial queue and announces a bid for that task based on its cost (distance) to perform 2A central aspect of multi-robot autonomous land- that task. It then waits for a certain time period to mine detection is to combine the information about receive bids for the same task from other robots. If it the characteristics of a potential landmine from dif- is the highest bidder and the task is still available, it ferent types of sensors and make a decision whether proceeds to perform the task; otherwise it selects the the object is indeed a landmine, and, identify its next available task from its spatial queue and repeats characteristics, if it is indeed one. Previous re- the bidding process. When it finishes performing the searchers [40] have considered this problem from a task, the robot broadcasts a task performed message. static viewpoint where all information about a land- When robots receive a task performed message from mine’s characteristics is assumed to be available and another robot that results in the task being com- the main concept is to use statistical inference tech- pleted (sufficient number of robots have visited the niques to classify the landmine’s characteristics with task), they rebuild their local copy of the transition accuracy. However, in COMRADES, the process of matrix. landmine detection is not an instantaneous one; it We verified the performance of the SQ-MRTA continues over a period of time during which robots (SQ) algorithm and compared it with three state- with appropriate sensors, corresponding to a po- of-the-art MRTA algorithms - the Hungarian as- tential landmine’s initially perceived characteristics, signment (HA) based algorithm [33], a decentral- need to be deployed to the potential landmine’s lo- ized greedy (DG) allocation algorithm [42] and the cation so that the cumulative information gathered repeated auctions (RA) algorithm [37]. As before, by the robots’ sensors can improve the accuracy of the algorithm was implemented on Corobot robots the landmine’s detection. Because of this dynamic within the Webots simulator. We used 5, 10, 15, or nature of landmine detection, we consider the follow- 20 robots with 6, 12, 18, or 24 tasks within a 20 × 20 ing multi-sensor information aggregation and sensor m2 environment. Each task was required to be per- scheduling problem in COMRADES - given an initial formed by 3−5 robots. All results were averaged over signature perceived by a certain type of sensor from 10 simulation runs. We evaluated different metrics a potential landmine, what is an appropriate set of including the total time required to complete tasks sensors (robots) to deploy additionally to the loca- and the average distance traveled per robot. Two tion of the potential landmine so that the landmine illustrative graphs of our simulations are shown in is detected with higher accuracy. Details of the oper- Figures 9 and 10. In Figure 9 we see that the SQ- ation of the information aggregation technique along MRTA algorithm performs comparably with the re- with analytical and experimental results of its per-

13 formance are given in [31]; we provide an overview technique are given in [31]. The aggregation mech- of its main features and a few significant results in anism and outputs a single aggregated belief value, the next section. Bt. The aggregation mechanism also selectively in- cludes weights of the sensor reports from a human Weightage of reports human from sensor expert based on Weight environment & operaonal expert about the accuracy of the sensors’ reports condions based on the ambient conditions of the sensors. The

Belief in object Exp. Reward Report Bayesian aggregated belief value is then passed on to a deci- being landmine a1,t r Aggregaon Agg. inference Object (calculated Strategy mechanism Belief and exp. (landmine) using Bayesian Report sion maker agent that makes decisions about which Belief using market Bt ulity inference, calculated based scoring Maximizn Sensed condioned on ba,t using exp. other robots (sensor types) should be deployed to the rule data past beliefs) ulity maximizaon Agg. belief Decision potential landmine’s location using a Bayesian infer- (from last Maker Agent me step) Sensor Agent 1 Market Maker Agent encing based technique, so that the landmine can be Reports from Robot/ other agents Sensor Scheduling scheduling decision (sensors) Algorithm confirmed rapidly and accurately [31]. Predicon Market

110 MD Figure 11: Schematic of the prediction market-based 100 IR IR GPR 90 0.5

information aggregation technique used to combine 80 MD 0.4 70

reports from multiple sensors in COMRADES. 60 0.3 50 Cost

40 0.2 To solve the information aggregation and sensor 30 Average RMSE 20 0.1 scheduling problem in COMRADES, we have pro- 10 GPR 0 0 Mine Metallic Object Non−metallic Object 0 1 2 3 4 5 6 7 8 9 10 posed a novel technique that uses a market-based Object Type Number of Time Steps information aggregation mechanism called a predic- (a) (b) tion market. Each robot participating in the land- mine detection task is provided with a software agent that uses the sensory input of the robot from a po- Figure 12: (a) Cost to classify different types of ob- tential landmine and performs the calculations of the jects - mines, metal but not mine and non-metal us- prediction market technique. A schematic describ- ing a different types of sensors. Relative costs of ing the technique is shown in Figure 11. When an sensing using MD, IR and GPR were assumed to agent records readings from a potential landmine on be in the ratio of 1 : 2 : 4. (b) Root mean square its sensors, it associates a probability value, called a error from different types of sensors from different belief, ba,t, to denote the agent’s confidence in iden- types of sensors when used individually over time at tifying the sensed object as a landmine. The be- the same object of interest. Note that although us- liefs are conditioned over past belief values to pre- ing MDs along has a low cost, their RMSE is higher vent wide variations from the object’s previous read- (accuracy is lower). ings due to sensor noise or ambient conditions using a Bayesian network. Each agent then strategically We simulated our algorithm using three different calculates a sensor report, ra,t, from its belief and types of sensors - MD, GPR and and IR-based multi- expected rewards from making the report, using a sensor device. We performed different experiments utility maximization technique. It then submits this with data from different types of sources (metallic, report to a central location called the aggregator low-metallic, non-metallic) collected by different sen- or market maker agent. The information aggrega- sors at different times. A few experimental results tion or fusion is implemented using the aggregation of our algorithm while using identical data distri- mechanism that uses a technique called a logarithmic butions and settings are highlighted in Figures 12 - market scoring rule (LMSR). This technique uses a Figure 14. Figure 12 shows the relative costs of de- utility-based formulation of the costs and value to ploying the sensors and the corresponding root mean each sensor (robot) for identifying an object of inter- square errors (RMSE) from the readings when using est as a landmine. It then assigns a score or virtual one type of sensor. Figure 13(a)-(d) shows the effect reward to each sensor (robot) if it correctly reports of deploying multiple sensors of different types over the probability distribution over the different types time. For these experiments, the data was assumed of objects of interests such as landmine, metal but to arrive from the same source object of interest. Dif- not landmine, and not landmine. The details of this ferent sets of sensors were deployed over 7 time steps.

14 0.55 0.5

0.5 0.6 0.5 0.45 MD MD MD 0.4 0.45 0.5 0.4 MD 0.35 0.4 0.3 0.4 0.35 0.3 0.25 0.3 0.3 0.2 0.2 IR IR 0.25 Average RMSE Average RMSE Average RMSE 0.15 Average RMSE 0.2

0.2 0.1 0.1 GPR 0.1 GPR 0.15 0.05 GPR 0.1 0 0 0 0 1 2 3 4 0 1 2 3 4 5 0 1 2 3 4 5 6 7 8 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 Number of Time Steps Number of Time Steps Number of Time Steps Number of Time Steps 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).

0.5 5 0

0.45 DDF 0 −0.02 0.4 D−S −5 −0.04 PM 0.35 −10 −0.06 0.3 0.25 −15 DDF −0.08 DDF RMSE 0.2 NMSE −20 −0.1 0.15 D−S −25 PM Information gain −0.12 D−S 0.1 PM −30 −0.14 0.05

0 −35 −0.16 0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10 Number of Time Steps Number of Time Steps Number of Time Steps (a) (b) (c)

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

We notice that while using 5 MDs (in Figure 13(a)) 5 Conclusions and Future Direc- the RMSE reduces to 20% in 4 time-steps, but with tion a combination of MD, IR and GPR (in Figure 13(d)) the combined RMSE from the fused data from these In this paper, we have described our experience with sensors is reduced much more, to around 6%. Fi- the COMRADE system for autonomous landmine nally, Figures 14 (a)-(c) illustrate the comparison of detection using multiple robots. We have customized our proposed prediction market based techniques for a Coroware Explorer robot with a metal detector information aggregation with two techniques. The to detect landmines and performed tests with it on results show that the information fusion performed different terrains. We have also addressed various using our technique reduces the RMSE by 5−13% as aspects of landmine detection such as multi-robot compared to a previously studied technique for land- search and coverage, multi-robot task allocation and mine data fusion using the Dempster-Shafer theory multi-sensor data fusion using different algorithms [40] and by 3−8% using distributed data fusion tech- and validated our results using simulated indoor nique [38]. We also conducted several experiments Corobot robots. to test the effect of various parameters in our model, and we found that using the combination of different As future work, we are looking at several direc- sensors in the environment gives the best accuracy tions to improve our proposed techniques. Integra- for the object’s type identification. 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 combination Reduction of Mine Suspected Areas. International of sensors for different ambient conditions such as Journal of Advanced Robotic Systems., 4(2):173–186, ambient temperature, sunlight, depth of landmines, 2007. etc. is an ongoing work in our research. We are [9] L. Breuer. Spatial queues with infinitely many also investigating ways to improve our coverage al- servers. In From Markov Jump Processes to Spatial gorithm by sharing maps between robots, planning Queues, pages 119–138. Springer Netherlands, 2003. coverage paths based on expected information gain [10] Francesco Bullo, Emilio Frazzoli, Marco Pavone, Ke- about landmines and improved information fusion tan Savla, and Stephen L. Smith. Dynamic vehicle techniques for combining data from multiple sensors. routing for robotic systems. Proceedings of the IEEE, Finally, we are also looking at using aerial robots to 99(9):1482–1504, 2011. aid ground moving robots navigate more intelligently in the environment. Overall, we envisage that our [11] H. Choset, W. Burgard, S. Hutchinson, G. Kantor, Lydia E. Kavraki, K. Lynch, and S. Thrun. Princi- research will lay the foundation for further research ples of Robot Motion: Theory, Algorithms, and Im- and proliferation of multi-robot systems for land- plementation. MIT Press, June 2005. mine detection, nuclear source detection, unmanned search and rescue, emergency response services and [12] Jorge Cort´es, Sonia Mart´ınez, Timur Karatas, and other high-risk applications. Francesco Bullo. Coverage control for mobile sens- ing networks. IEEE T. and , Acknowledgements. The COMRADES project 20(2):243–255, 2004. was supported by the US Office of Naval Research, grant no. N000140911174. [13] Paulo Debenest, Edwardo F. Fukushima, Yuki Tojo, and Shigeo Hirose. A new approach to humanitarian demining. Auton. Robots, 18(3):323–336, 2005. References [14] Eric den Breejen, Klamer Schutte, and Frank Cre- [1] http://www.fp7-tiramisu.eu/. [Online; accessed mer. Sensor fusion for antipersonnel landmine detec- 01-January-2015]. tion: a case study. volume 3710, pages 1235–1245, 1999. [2] http://cmantic.unomaha.edu/projects/comrades/. [Online; accessed 14-August-2014]. [15] M.B. Dias, Robert Zlot, N. Kalra, and A Stentz. Market-based multirobot coordination: A survey [3] Okabe A, Boots B, Sugihara K, and Chiu S. Spatial and analysis. Proc. IEEE, 94(7):1257–1270, 2006. Tessellations: Concepts and Applications of Voronoi Diagrams. Wiley& Sons;, 2000. [16] Joseph W. Durham, Ruggero Carli, Paolo Frasca, and Francesco Bullo. Discrete partitioning and cov- [4] M. Acheroy. Mine action: status of sensor tech- erage control for gossiping robots. IEEE Transac- nolo.gy for close-in and remote detection of an- tions on Robotics, 28(2):364–378, 2012. tipersonnel mines. In Advanced Ground Penetrating Radar, 2005. IWAGPR 2005. Proceedings of the 3rd [17] FA El BAKKOUSH. Current research activities for International Workshop on, pages 3–13, May 2005. landmine detection by nuclear technique in libya. [5] Gerkey B and Mataric M. A formal analysis and 2009. taxonomy of task-allocation in multi-robot systems. [18] Xuan Feng and M. Sato. Landmine imaging by a The International Journal of Robotics Research, hand-held gpr and metal detector sensor (alis). In 23(9):939–954, 2004. Geoscience and Remote Sensing Symposium, Proc. [6] S. Badia, U. Bernardet, A. Guanella, P. Pyk, and IEEE International, volume 1, July 2005. P. Verschure. A biologically based chemo-sensing [19] Marc Freese, Paulo Debenest, Edwardo F. UAV for humanitarian demining. International Fukushima, and Shigeo Hirose. Development Journal of Advanced Robotic Systems., 4(2):187–198, of deminer-assisting robotic tools at tokyo in- 2007. stitute of technology. In Maki K Habib, editor, [7] Maxim A. Batalin and Gaurav S. Sukhatme. Spread- Humanitarian Demining. Intech Open, 2008. ing out: A local approach to multi-robot coverage. In [20] Hichem Frigui, Lijun Zhang, Paul Gader, Joseph N Proc. of 6th International Symposium on Distributed Wilson, KC Ho, and Andres Mendez-Vazquez. An Autonomous Robotic Systems, pages 373–382, 2002. evaluation of several fusion algorithms for anti-tank [8] I. Bloch, N. Milisavljeniv, and M. Acheroy. Mul- landmine detection and discrimination. Information tisensor Data Fusion for Spaceborne and Airborne Fusion, 13(2):161–174, 2012.

16 [21] T. Fukuda, Y. Hasegawa, Y. Kawai, S. Sato, [33] H. W. Kuhn. The Hungarian method for the assign- Z. Zyada, and T. Matsuno. GPR signal processing ment problem. Naval Research Logistics Quarterly, with geography adaptive scanning using vector radar 2(1-2):83–97, 1955. for antipersonal landmine detection. International Journal of Advanced Robotic Systems., 4(2):199–206, [34] W. Lenagh. Multi-robot task allocation: A spatial 2007. queueing approach. Master’s thesis, University of Nebraska, Omaha, 2013. [22] Brian P. Gerkey and Maja J. Mataric. A formal analysis and taxonomy of task allocation in multi- [35] L. Liu and D. Shell. A distributable and robot systems. I. J. Robotic Res., 23(9):939–954, computation-flexible assignment algorithm: From 2004. local task swapping to global optimality. Proceed- [23] K. R. Guruprasad and Prithviraj Dasgupta. Dis- ings of Robotics Science and Systems, pages 33–41, tributed voronoi partitioning for multi-robot systems 2012. with limited range sensors. In IROS, pages 3546– [36] M. Long, A Gage, R. Murphy, and K. Valavanis. Ap- 3552, 2012. plication of the distributed field robot architecture to [24] M. Habib. Humanitarian demining: Reality and the a simulated demining task. In Robotics and Automa- challenge of technology - The state of the arts. In- tion, Proc. IEEE International Conference on, pages ternational Journal of Advanced Robotic Systems., 3193–3200, April 2005. 4(2):151–172, 2007. [37] Lingzhi Luo, Nilanjan Chakraborty, and Katia P. [25] Stefan Havlik. Land robotic vehicles for demining. Sycara. Distributed algorithm design for multi-robot In Maki K Habib, editor, Humanitarian Demining. generalized task assignment problem. In IROS, pages Intech Open, 2008. 4765–4771, 2013. [26] Stefan Havlik. Some robotic approaches and tech- nologies for humanitarian demining. In Maki K [38] J. Manyika and H. Durrant-Whyte. Data Fusion and Habib, editor, Humanitarian Demining. Intech Sensor Management: A Decentralized Information- Open, 2008. Theoretic Approach. Prentice Hall, Upper Saddle River, NJ, USA, 1995. [27] M. Hiznay. Landmine monitor report. Technical re- port, 2011. [39] S. Masunaga and K. Nonami. Controlled metal [28] Kurt Hungerford, Prithviraj Dasgupta, and K. R. detector mounted on mine detection robot. In- Guruprasad. Distributed, complete, multi-robot cov- ternational Journal of Advanced Robotic Systems, erage of initially unknown environments using repar- 4(2):207 – 218, 2007. titioning. In AAMAS, pages 1453–1454, 2014. [40] Nada Milisavljevic and Isabelle Bloch. Sensor fusion [29] Rekleitis I, New A, Rankin E, and Choset H. Ef- in anti-personnel mine detection using a two-level be- ficient boustrophedon multi-robot coverage: an al- lief function model. IEEE Transactions on Systems, gorithmic approach. Annals of Math and Artificial Man, and Cybernetics, Part C, 33(2):269–283, 2003. Intelligence, 52:109–142, 2008. [41] Yoshikazu Mori. Peace: An excavation-type dem- [30] Jun Ishikawa, Katsuhisa Furuta, and Nikola ining robot for anti-personnel mines. In Maki K Pavkovic. Test and evaluation of japanese gpr-emi Habib, editor, Humanitarian Demining. Intech dual sensor systems at the benkovac test site in croa- Open, 2008. tia. In Katsuhisa Furuta and Jun Ishikawa, editors, Anti-personnel Landmine Detection for Humanitar- [42] Ang´elica Mu˜noz-Mel´endez, Pritviraj Dasgupta, and ian Demining, pages 63–81. Springer London, 2009. William Lenagh. A stochastic queueing model for [31] Jumadinova J and Dasgupta P. Multirobot au- multi-robot task allocation. In ICINCO (1), pages tonomous landmine detection using distributed mul- 256–261, 2012. tisensor information aggregation. In Proceedings of SPIE Conference on Multisensor, Multisource Infor- [43] Hazon N and Kaminka G. On redundancy, ef- mation Fusion: Architectures, Algorithms, and Ap- ficiency, and robustness in coverage for mulitple plications 8407, pages 1–12, 2012. robots. Robotics and Autonomous Systems, 56:1102– 1114, 2008. [32] Guruprasad K, Wilson Z, and Dasgupta P. Complete coverage of an initially unknown environment by [44] P. Oberlin, S. Rathinam, and S. Darbha. A trans- multiple robots using voronoi partition. In Proceed- formation for a heterogeneous, multi-depot, multi- ings of 2nd International Conference on Advances ple traveling salesman problem. In Proceedings of in Control and Optimization of Dynamical Systems the American Control Conference, pages 1292–1297, (ACODS), Bangalore, India, 2012. 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 in Defense Support Systems - 2, pages 5–20. Springer- Verlag. [46] J. Prado, G. Cabrita, and L. Marques. Bayesian sensor fusion for land-mine detection using a dual- sensor hand-held device. In Industrial Electronics Society, IECON 2013 - 39th Annual Conference of the 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 context model. In Geoscience and Remote Sensing Sym- posium (IGARSS), 2010 IEEE International, pages 4192–4195, July 2010. [49] Hert S and Lumelsky V. Polygon area decomposition for multiplerobot workspace division. International Journal of Computational Geometry & Applications, 8:437–466, 1998. [50] Pedro Santana, Jos´eBarata, H. Cruz, A. Mestre, J. Lisboa, and Lu´ısFlores. A multi-robot system for landmine detection. In ETFA, 2005. [51] P. Verlinde, M. Acheroy, G. Nesti, and A Sieber. Preparing the joint multi-sensor mine-signatures project database for data fusion. In Geoscience and Remote 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 unweighted terrain. 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 . Annals of Math and Artificial Intelligence, 31:77–98, 2001.

18