Multi-Robot Communication-Sensitive. reconnaisance

Transcription

1 Mult-Robot Communcaton-Senstve Reconnassance Alan Wagner College of Computng Georga Insttute of Technology Atlanta, USA Ronald Arkn College of Computng Georga Insttute of Technology Atlanta, USA Ths paper presents a method for mult-robot communcatonsenstve reconnassance. Ths approach utlzes collectons of precompled vector felds n parallel to coordnate a team of robots n a manner that s responsve to communcaton falures. Collectons of vector felds are organzed at the task level for reusablty and generalty. Dfferent team szes, scenaros, and task management strateges are nvestgated. Results ndcate an acceptable reducton n communcaton attenuaton when compared to other related methods of navgaton. Onlne management of tasks and potental scalablty are dscussed. Keywords:behavor-based robotcs, nternalzed plan, multrobot, reconnasance I. INTRODUCTION Ths paper contrbutes a novel mult-robot method for performng communcaton-senstve reconnassance, whch nvolves explorng an urban area n a manner that prevents a team member from becomng a lone dsconnected network. Ths type of reconnassance s one goal of DARPA s MARS Vson 22 program and has mplcatons for much of the mult-robot communty. Communcaton senstvty s an mportant consderaton for teams of robots operatng n dynamc and potentally hazardous envronments. In partcular, communcatng robots may be more capable of self-rescue, better equpped to relay nformaton back to a human operator, and have advantages n terms of localzaton. [3]. These types of envronments requre agents capable of coordnated sensng, processng, and communcaton [6]. Multple archtectures have been created for the purpose of autonomous navgaton. Implemented systems range from purely reactve [4] to sense-plan-act [9]. Hybrd delberatve/reactve archtectures attempt to address the shortcomngs of these two extremes [1, 5]. Other approaches nclude contnuous calculaton of a local gradent feld [8] or plannng only when reactve behavors fal [12]. Payton also delneates a method for combnng plannng wth reactve navgaton [11]. In ths method, a pror map knowledge becomes an enablng resource for decson-makng. From hs perspectve, tradtonal plans are artfcally abstracted from knowledge that often results n over- or underspecfcaton of a msson s objectves. By mnmzng symbolc abstracton, a plan for acton s developed that can be used drectly by a reactve agent. Payton brands ths type of plan an nternalzed plan. These nternalzed plans dffer from tradtonal plans by ther lack of abstract symbol use and ther tght representatonal couplng to the needs of a reactve robot. Moreover, the plans are used only as advce, where njectng world map or other types of knowledge s performed only at the dscreton of the robot. Combnng plans wth reactve navgaton s not new. Rather than smply mplementng a plannng algorthm on top of a reactve archtecture, the method outlned n ths paper both extends and generalzes an earler hybrd approach. In prevous research [13], nternalzed plans were ntegrated wth Arkn s motor schema archtecture [2] usng the Mssonlab [1] behavor specfcaton software. An nternalzed plan s created by runnng a unform cost search algorthm to produce a gradent feld on a grd mesh. The resultng vector feld drects a robot from any locaton on a map to a goal locaton. Ths hybrd approach allevates some of the problems assocated wth purely reactve systems (e.g., local mnma, box canyons, and mazes) whle stll provdng tmely response to unplanned obstacle encounters. Ths earler work also developed an effcent method for usng sets of multple plans n parallel, enablng a robot to focus attenton on one plan over another n a gven stuaton or va a weghted combnaton of plans. By stackng multple vector felds on top of one another, advce can be ether arbtrated or weghted based on a rank ordered attenton mechansm (fg. 1). Buldng from ths ntal research, new procedures and technques for plan selecton, organzaton, and coordnaton are developed that can potentally be extended to novel envronments and generalzed across a wde varety of domans. Ths approach shows promse as a method for coordnatng teams of robots whle performng communcatons-senstve reconnassance. II. METHOD OVERVIEW Our method operates on the premse that reconnassance of a large urban area can be reduced nto a collecton of smaller reconnassance tasks. Moreover, f each of these tasks s senstve to communcaton attenuaton then, overall, the entre operaton wll be senstve to communcaton attenuaton. Naturally there may be many dfferent tasks; crcle a target area, perform reconnassance wthn a buldng, or ext a locaton, to name a few. Each ndvdual task may also need to be repeated several tmes, dfferng only by some parameter, such as locaton. When a task s repeated most, f not all, of the Ths research s funded under DARPA/DOI contract #NBCH1212 as part of the MARS Vson 22 program.

2 Drecton Advce Communcaton Plan Advce Vector: α P 1 +βp 2 Movement vector Robot Controller Plan Controller Coverage Plan Sensory data Feature Vector Fgure 1. Plans used n parallel. The top plan represents a communcatons plan. The bottom plan represents a coverage plan. Output advce s determned va arbtraton or weghted summaton where α s the weght of plan P 1 and β s the weght of plan P 2. underlyng structure that composes the task may be smlarly repeated. Hence by ntellgently developng and combnng smple communcaton-senstve reconnassance tasks ths method can be made to perform reconnassance on a much larger scale. For example, reconnassance of one partcular buldng may dffer from reconnassance of another buldng n the detals of the rooms or the ntermedate locatons to be navgated, but not regardng the overall procedures performed. In ths research, a task s represented by a plan element, and repetton of the same task s represented by dfferent nstances of that same element. The precedng argument has mportant consequences from a plannng perspectve. Although the moment-to-moment nuances of the envronment may change radcally, the procedures, steps, and mlestones for accomplshng a msson wll lkely not change. Granted some plans may be made untenable by changes n a dynamc envronment. In ths case t s often acceptable for the team to recognze that ts plan s nvald and gve up (and replan) rather than attempt to solve a problem for whch t has lttle or no resources. Ths lne of reasonng draws from work on cognzant falure [7]. In our case snce plannng s used only as advce to a reactve controller, the system should reman robust to dynamc and unaccounted for obstacles and mpedments. Moreover, even the method by whch the advce s realzed may not be mportant, as long as the underlyng reactve controller has the fnal determnaton of whch headng and velocty to select. Team coordnaton s another mportant aspect for multrobot reconnassance. Ths approach provdes mechansms for both between-task and wthn-task coordnaton. Between-task coordnaton s accomplshed by restrctng the transton from one task to another untl all robots have completed the precedng task. Wthn-task coordnaton s accomplshed usng specfcally desgned progress ponts or stages for the task. Movement from one progress pont to another s smlarly restrcted untl the precondtons of a transton functon are satsfed. The larger components of ths method are descrbed below and depcted graphcally n fgure 2. One component s the plan controller. Ths component s devoted to four tasks: 1) managng the plan elements whch compose the overall msson, 2) coordnatng the progresson from one plan element to the next 3) determnng the role each robot wll play when executng an element and 4) communcatng the plan controller s advce to the robot Fgure 2. Interacton of the robot controller, plan controller, and feature vectors. The robot controller mantans control of the robot but receves drectonal advce from the plan controller. The plan controller, n turn, receves nformaton regardng the status of the current envronment encoded n a feature vector and presents drectonal advce to the robot controller. controller. Pseudocode for ths procedure s provded n fgure 13. As mentoned, ths software represents tasks as ndvdual plan elements. These elements can be managed n several ways. The overall msson could, for example, demand a strct orderng of tasks. Or, on the other hand, a least commtment orderng of tasks mght be acceptable (see fgure 13 for pseudocode). The plan controller also coordnates the transton from one plan element to the next. Ths guarantees the completon of one task (ether successfully or unsuccessfully) pror to begnnng another task. The plan controller may need to assgn robots to partcular roles f the underlyng task demands t. Currently, the controller assumes the robot team to be homogeneous, but we foresee no dffculty extendng ths approach to heterogeneous teams and hope to address ths task n future work. Fnally the plannng system s advce s communcated to the robot controller va a unt vector. The robot controller executes the underlyng reactve system, n ths case Mssonlab s motor schemas. The robot controller nterfaces wth the plan controller recevng ts advce n the form of an egocentrc unt vector. The robot controller mantans the opton to gnore plan advce at any tme and provdes the plan controller wth postonal nformaton and sensor-derved feature vectors. Feature vectors provde a snapshot of the robot s envronment to the plannng apparatus. They contan the nformaton necessary to produce an advce vector. Feature vectors are created by the robot controller based on ncomng sensory nformaton. III. PLAN ELEMENTS Plan elements have already been descrbed as the ndvdual tasks nto whch the overall msson s decomposed. Each plan element represents a recommended soluton to a problem. A sngle element gudes each robot of a team n a coordnated manner to accomplsh a specfc task. Many unque nstantatons of a sngle element may exst. Fgure 3 depcts the components of a plan element and fgure 13 provdes pseudocode. All plan elements share a common nterface format. Ths serves several purposes. Frst, t allows unform handlng and management of all elements. Second, t ensures that the

3 Feature Vector Output Advce Team Sze Plan Element Decson Tree Internalzed plans Plan Element Interface Role 3 2 Progress Comms. Status Fgure 3. The nternal structure of a plan element. The major components that characterze a plan element are depcted. mplementaton underlyng an element s restrcted to the element tself. Fnally, t defnes and restrcts the type of nformaton that an element can produce and receve. It s expected that by usng a common nterface the development of addtonal plan elements wll be made easer and standardzed. Currently the nterface lmts each plan element s nput to a feature vector, the role of the robot, and the number of robots n the team. Ths allows the element to be generalzed wth respect to team sze, the robot s role n the task, and the status of the envronment. Ths desgn decson has benefts and drawbacks. On the one hand, an nterface that lmts nput to all elements vastly smplfes the management and operaton of the elements. On the other hand, because dfferent elements may requre dfferent nput, ths strategy wll lkely fal when the number and type of elements becomes more complex. One soluton mght be to mantan two separate nput channels: one common among all elements and one specfc to each element. In any case, as the quantty of planned tasks a robot can perform ncreases so does the dffculty of managng and usng those tasks. The common element nterface smlarly lmts each element s output to a unt vector wth drectonal advce. The output s then passed from the plan controller to the robot controller. Because ths method utlzes several plans n parallel addtonal channels for element output are not necessary. Ths s an mportant and defnng characterstc of our approach. The plannng system dstlls several rapdly changng and complex nput varables nto a sngle pece of output advce that wll not requre any addtonal processng for use by the robot controller. Moreover ths s performed n real-tme n a manner that s reusable from element nstance to element nstance. Between each plan element s nput and output, two components a decson tree and a parallel plan accomplsh the reducton n complexty. Each element mantans a decson tree mappng the element s nput to an array of gan values. Gan values are multplers of the basc output vectors of each behavor [2], or, n ths case, the relatve strength of the nternalzed plans. When nput s presented to an element by the plan controller t flows down the element s decson tree eventually resultng n an array of gans. Ths array wll determne the nfluence (gan) of each ndvdual nternalzed plan n a parallel-nternalzed plan. Whle the decson tree for each nstance of an element s the same, the decson tree for each type of element may be dfferent. In other words each Gan A Gan B Fgure 4. The decson tree used for the ReconBuldng element. The frst node selects based on team sze, the next node on role, the thrd nodes on progress and the fnal node based on communcatons status. Example values are dsplayed below and to rght of each node. Gan A and B are meant to represent arbtrary gan arrays. decson tree s task-specfc but not nstance-specfc. The decson tree mantans the gudelnes by whch a task s conducted. An element s decson tree may be extremely complex or as smple as a drect ponter to a sngle statc gan. For example, f the robot s ablty to communcate s acceptable, then the decson tree may select an array of gans that favor a coverage plan over a communcatons plan, perhaps drvng two robots n dfferent drectons. If, on the other hand, the robot s current ablty to communcate s unacceptable then the resultng gan values wll prefer a communcaton plan over a coverage plan, perhaps forcng the robots toward one another. Fgure 4 outlnes the decson tree for one plan element used n ths study ReconBuldng. Due to space consderatons only a sngle path through the tree s shown. The frst node of the decson tree branches based on team sze. Later the tree branches accordng to role, progress, and communcaton status. Arbtrary values have been ncluded for completeness: team sze-3, role-2, and progress-. These values map to the gan arrays A and B. The fnal stage n the operaton of a plan element apples the array of gans determned by the decson tree to a parallel plan producng the element s output. An nternalzed plan [11] has already been descrbed as a gradent feld generated on a grd mesh usng a map. Multple nternalzed plans can be utlzed n parallel by stackng ndvdual plans on top of one another. Output advce s then determned by multplyng the advce for each ndvdual plan at a locaton by a gan as determned by the decson tree. The detaled use of parallelnternalzed plans appears n [13]. IV. IMPLEMENTATION The system was developed wth the ntent of enablng teams of autonomous robots to perform coordnated communcaton-senstve reconnassance as part of DARPA s MARS Vson 22 program. Pror to mplementaton t was

4 Fgure 5. The tranng vllage located at Fort Bennng GA. necessary to determne the types of tasks that are necessary for successful completon of a communcaton-senstve reconnassance msson. The felded system wll be tested n an outdoor mockup of a European vllage (fg. 5) at Ft. Bennng, Georga. A team of robots must explore ths small town of less than a.25 x.25km area. Our overall role n Vson 22 has guded the choce of tasks for the robot team. It s not mantaned that these are the best, most optmal, or most characterstc tasks and assocated elements for performng reconnassance. They smply represent the tasks chosen to address ths problem, and other desgn choces are possble. Ths secton s organzed nto three parts. Frst, the two major tasks are descrbed n detal. The computatonal process necessary for completng each task s then explaned. Fnally the procedure for usng a plan element s descrbed. A. Descrpton of Elements Two types of plan elements were defned, one for each task deemed necessary. ReconBuldng s an element devoted to surroundng and movng around buldngs, n a communcaton-senstve manner. Ths allows the team to explore the entre vllage, ncludng alleys, streets, and passageways, wth mnmal lose of communcaton; eventually coverng the entre area and succeedng n ts reconnassance msson. MoveTo s an element that gudes the team of robots from one locaton to another, resortng to a contngency plan f communcaton fals. Ths element gudes robots between the reconnassance of ndvdual buldngs. Other methods, such as reactve navgaton, could have been used nstead. A second plan element was mplemented n order to examne the generc nature of the plan element nterface. The MoveTo plan element drects the robots to a goal poston from any locaton on the map. Network sgnal strength, along wth other unused nformaton, s nput to ths element by the plan controller. Its decson tree assgns a gan of 1. to the nternalzed plan leadng to the goal and. to the contngency plan whle the network sgnal strength to all teammates s above a 1% sgnal strength threshold level. If the network sgnal strength decreases below the threshold level, a gan of. s assgned to the nternalzed plan leadng to the goal and a gan of 1. assgned to the contngency plan, Fgure 6. Progress stages for the ReconBuldng plan element. (A) depcts the start poston and the dfferent drectons for each robot s role at ths stage. (B) shows the robots n poston for the transton to stage 2. (C) dsplays the robots n poston for the transton to stage three. (D) depcts the robots at the goal locaton. effectvely swtchng between the plans. Once the contngency plan has been selected t contnues to functon wthn ths element regardless of network sgnal strength. Ths prevents the robots from thrashng between plans at the border of a communcatons falure. The ReconBuldng plan element s more complex. Ths element employs all data nput from the plan controller: the robot s current poston, the locatons of teammates, the sgnal strength to each teammate, and the robot s role. The feature vector, transmtted to the element by the plan controller, contans the pertnent sensory nformaton. The robot s role s assgned to the element when the element s nstantated. Input from the plan controller traverses the decson tree n fgure 4. The frst branch of the tree segregates based on team sze. The team sze attemptng to surround the buldng s an mportant factor n determnng where and when a team member should move. If the team conssts of two robots the robots ntally attempt to pass the object from opposte sdes. Intutvely sendng each robot around a dfferent sde of a buldng may seem mproper when communcaton mantenance s one of the stated goals of the system. However, these tasks were desgned to prefer opportunstc plan advce. Thus an attempt s made to frst surround the buldng even f communcaton attenuaton s lkely. Other choces are possble. Three robot teams attempt to surround the buldng by leavng one robot behnd to act as a communcaton relay for the other two members. Thus, n a team of three, one robot s assgned the role of communcaton relay, one s tasked wth explorng one sde of the buldng, and the thrd robot s task wth explorng the opposte sde. Teams of four robots, depcted n fgure 6, attempt to surround the buldng by creatng a rectangle of communcatng robots around the buldng. Agan two robots are assgned (by the plan controller) the role of explorng opposte sdes of the buldng. The remanng two robots act as communcaton relays by

5 movng to nearsde postons. In ths research, the number of roles a team has s equal to ts team sze. Surroundng a buldng wth a robot team requres coordnaton. For ths reason the thrd node of the decson tree branches based on the robot s progress. A two robot team has no progress stages. Teams of three and four have three progress stages. The frst stage begns wth the robot s ntal locatons and ends when the robots assgned to explore opposte sdes of the buldng have reached ther assgned locatons and reestablshed contact wth one another. The begnnng of stage one for a team of four robots s depcted n fgure 6a. The arrows ndcate each robot s ntal trajectory. Fgure 6b shows the team at the end of stage one. Progress stage two gudes the two-team members actng as communcaton relays to ther fnal goal locaton. Ths stage ends when the communcaton relay robots have reached the same locaton as the explorng robots, depcted n box three of fgure 6c. The fnal stage gudes all of the robots the goal locaton. The last node of the decson tree selects based on communcaton status. Acceptable versus unacceptable communcaton s nfluenced by team sze, role, and progress stage. For example, n a team of four robots at the frst progress stage, the explorng robot only needs to mantan a lnk wth a partcular communcaton relay robot. A communcaton relay robot, on the other hand, must mantan a lnk wth both an explorng robot and the other communcaton relay robot. The ReconBuldng element s parallel plan conssts of four ndvdual nternalzed plans: a contngency plan, two communcaton relay plans, and a coverage plan. The contngency plan gudes each robot to the element s goal locaton and was the same for all robots n the team. The two communcaton relay plans gude the robots to relay locatons. Fnally, a coverage plan gudes each robot to the element s goal locaton but n a manner that sends the robots down dfferent sdes of the buldng hence comng n two types. Thus, n total ths element requres producton of fve nternalzed plans. Currently the arrays produced by the decson tree set the gan for one plan to 1. and all others to zero. In future work we ntend to blend advce from plans at each tme step. B. Creatng a Plan For Reconnassance The process of determnng whch tasks decompose nto the msson has already been dscussed. Smlarly, fleshng out a task nto the decson tree and the parallel plans s currently an nexact and emprcal process. The components of the MoveTo and ReconBuldng have been descrbed n the precedng secton. Next, the map locatons that underle each element s nternalzed plans are requred. All postons are determned pror to runnng the msson and were, n ths case, determned emprcally. The MoveTo element utlzes a parallel plan consstng of two nternalzed plans. The goal locaton for ths element s contngency plan drects t to an arbtrary poston outsde the vllage. Ths s meant to regroup the team f communcaton fals when movng from one buldng to another. Ths element s other plan smply drects t to the element s goal poston. The ReconBuldng element employs a parallel plan consstng of four nternalzed plans. Because the contngency plan and both types of the coverage plan gude the robot to the Fgure 7. An example of the vector feld produced for the ReconBuldng plan element. Temporary boundares gude the producton of the vector gradent. same goal, only one goal locaton needs to be determned for these three plans. Addtonal postons are necessary for each of the two communcaton relay ponts, for a total of three postons. The dfferent types of coverage have been mentoned brefly above. If one type leads the robot to the element s goal along a westerly (or northerly) path then the other type leads the robot along an easterly (or southerly) path. Constructng each type of coverage plan requres basng the unform cost algorthm to favor paths n one drecton over paths n another drecton. Ths s accomplshed by addng cost to less favored path areas or by hallucnatng a temporary boundary preventng access along one sde of a buldng, whch s the method we chose for ease of mplementaton. Fgure 7 depcts a ReconBuldng element durng constructon. The gradent feld drects the robot around the buldng from below. The temporary boundares that are used to nfluence path drecton and to grow the obstacle are marked. A temporary boundary that grows the buldng s employed to prevent communcaton attenuaton at the far left corner of the buldng snce the vector feld slopes away from the communcaton relay robot. Both of these alteratons to the nternalzed plan generaton were used only for the two types of coverage plans n the ReconBuldng plan element. In order to reduce plan computaton tme and the resultng resource fle sze, the nternalzed plans for the ReconBuldng element were not computed over the entre map. Rather a restrcted rectangle was constructed around each buldng usng temporary obstacles vsble only to the nternalzed plan generaton algorthm. These rectangles requred four addtonal postons. Fgure 7 shows the temporary boundary of the rectangle surroundng the buldng. In all, the generaton and alteraton of a sngle ReconBuldng element requres knowng ffteen postons: four ponts for the rectangle constructed around the buldng, four ponts markng the obstacle boundares used to nfluence the coverage plans path (two for each type/sde), four ponts used to grow the buldng (two for each type/sde), and three postons for the goals of the underlyng plans.

6 After all the locatons necessary for the reconnassance msson have been collected, the ndvdual nternalzed plans are compled [13]. Upon completon, a resource fle s generated. Software tools have been created to enable a user to alter specfc nternalzed plans wthn a resource fle. Indvdual plans can be added, deleted, or replaced wthout complete recomplaton of all of the plans. Fnally a msson s constructed employng the followplan behavor n the MssonLab behavor specfcaton system. C. Utlzng a Reconnasance Plan At startup, but pror to msson executon, the resource fle contanng data for each nternalzed plan s parsed and loaded nto plan objects. At runtme, feature vectors are generated that nclude data produced by ether a network model when runnng n smulaton or a hardware network component developed by BBN Technologes and sent to the plan controller. The plan controller adds nformaton pertanng to the sze of the team and the robot s partcular role n the plan. Ths nformaton s provded to the current plan element. The plan controller selects whch element to employ. Ths research explored usng both a strct orderng of plan elements and a least commtment selecton mechansm. Fnally the element produces plan-based advce to be utlzed by the robot controller or gnored. A task ends when all of the robots n the team have reached the element s goal locaton. V. EXPERIMENTS Several experments were run n smulaton to test the new system. All experments were run usng the map and obstacle representaton of a tranng vllage (fgure 8). Experments were conducted on the entre map and requred thrteen MoveTo and thrteen ReconBuldng elements. The actual map contans 15 buldngs. In two cases separate buldng were treated as a sngle complex due to ther close proxmty. These experments conssted of thrty trals startng from a random locaton outsde the urban terran of the map. The robot teams were expected to navgate to and through the cty. The map of the Fort Bennng vllage s accurate to approxmately onemeter resoluton and covers a terran of 22 square meters. Prevously conducted tests have verfed the ablty of real robots to operate n and around the vllage wth nternalzed plans usng ths map. The network model used for these experments reduced network sgnal strength when the robots were occluded by terran. Evaluaton was based on total msson tme, total dstance traveled, urban terran coverage, and percent tme wth at least one lone network. These metrcs are felt to characterze performance n real world reconnassance operatons. Estmates of baselne performance were obtaned by comparng the plan element system to control experments that used a three-robot team wanderng the ste and another usng navgaton based on Mssonlab s exstng waypont planner [14]. Although the control experments do not provde a perfect comparson, they are meant to convey a general sense of the performance of uncoordnated reactve team behavor. All experments used the same randomzed start locatons and dentcal gans for obstacle avodance. The effect of team sze was also nvestgated. As mentoned prevously, ths system operates for teams of up to four robots. Performance and Fgure 8. A map of the tranng vllage located at Fort Bennng GA. All smulaton trals were run on ths test ste. Early experments wth real robots have examned the valdty of utlzng nternalzed plans at ths ste. scalablty were examned for two, three, and four robot team szes. We also expermented wth dfferent strateges for selectng plan elements. Even though plan elements are predefned n terms of ther underlyng locatons, the plan controller selects elements dynamcally at runtme for use by the robot. Usng teams of three robots, a least commtment element selecton mechansm was compared to selectng elements usng a strct orderng determned by the expermenter. The least commtment mechansm consstently selected the plan element from the set of elements whch mnmzed the dstance from the current locaton to the element s goal locaton, thus contnually selectng the nearest buldng for whch reconnassance had not yet been performed. Fnally experments were performed usng a scenaro n whch one robot of a three-robot team was constantly fxed (tethered) at a sngle locaton. Ths experment explores an mportant real-world scenaro where a human-operated vehcle deploys the reconnassance robots that must explore an urban space whle mantanng contact wth the statonary base vehcle. VI. RESULTS It was conjectured that by utlzng coordnated plannng elements the percentage of lone or solated networks would decrease sgnfcantly n comparson to the control systems. It was further hypotheszed that as team sze ncreases, network connectvty would mprove. It was beleved that as the total number of robots ncreases, although the task of coordnatng becomes ncreasngly dffcult, more opportunty for network connectvty exsts. Fgures 9-12 dsplay the results for all experments. Fgure 9 examnes communcaton attenuaton. Due to lack of team coordnaton, the waypont planner performed sgnfcantly worse then all other all other experments on ths metrc (p = on two-taled t-test). Teams of three performed sgnfcantly worse then teams of other szes (p = vs. team sze of two and team sze of three). Teams of four mantaned nearly perfect network connectvty wth lttle varance over randomzed start locatons. The wander behavor

7 Percent Tme wth A Lone Network Tme Performance Metrc Percent Tme (tme steps) WayPt Wander 2-Team 3-Team 3-Team Tet 4-Team LeastCom Experment Tme (tme steps) WayPt Wander 2-Team 3-Team 3-TeamT 4-Team LeastCom Experment Fgure 9. Percent tme wth a lone network. The waypont control performs the worst. Teams usng plan elements mproves communcaton preformance. Urban Coverage (m^2) Coverage Performance Metrc WayPt Wander 2-Team 3-Team 3-Team T 4-Team LeastCom Experment Fgure 1. Coverage. The wander control performs the worst. Teams usng plan elements cover large areas of urban terran. also performed well on ths metrc. Ths control s excellent communcaton results, however, reflect the behavor s lack of urban exploraton rather then proper performance. More nterestngly, the tethered team mantaned a great deal of network connectvty resultng n a lone network only 4.6% of the tme (p =). Ths seems to ndcate that a tethered team of robots could perform a probe-lke reconnassance msson n whch communcaton mantenance was vtal. The tethered robot team lkely outperforms the untethered team because t has less opportunty for falure t s tethered. Overall the communcaton results ndcate the value of plan elements coordnated plannng reduces the number of solated robots. We also see that addtonal teammates help mantan network connectvty, although ths relatonshp s not strctly lnear. Teams of three lkely fare the worst because they tend to be unable to completely surround buldngs and have two network connectons to sever rather then one. Fgure 1 examnes urban terran coverage for each expermental condton. Greater coverage denotes better performance. Due to ts lack of locaton coordnate the wander control performs the worst. As would be expected, the tethered team also performs poorly. Although the coverage performance of the remanng experments occasonally dffers sgnfcantly (p = 3-Team vs. 4-Team), ths s lkely the result of our overly strct standard for coverage. In all experments, t was assumed that each robot has only the ablty to sense and hence cover a three by three meter area n all drectons from ts current locaton excludng obstacles. We suspect that relaxng ths standard would result n approxmately equvalent coverage for all experments except the tethered experment and the wander dfferences resulted. The tethered team s addtonal tme s an Fgure 11. Tme. Only the tethered team requres sgnfcantly greater tme. Dstance (m) Dstance Performance Metrc WayPt Wander 2-Team 3-Team 3-Team T 4-Team LeastCom Experment Fgure 12. Dstance A measure of energy expendture. When the plan controller utllzes a least commtment strategy the total dstance traveled by a team of four robots deceases. experment. Overall the coverage results ndcate that large porton of the urban terrtory s beng explored by the robots. Fgure 11 dsplays the tme requred for each experment. Wth the excepton of the tethered team no sgnfcant ndrect result of beng tethered. These teams typcally traverse most of the map, break communcaton, and then retreat back to the tethered robot. The back and forth nature of ths scenaro ncreases the tme requred to perform ths msson. Fgure 12 depcts the total dstance traveled by all robots n the team. Ths graph gves an dea of the energy requrements for the dfferent types of experments. The use of a least commtment strategy appears to have reduced the average energy consumpton of a team of four robots to the equvalent of a team of three. As one would expect the varance of ths strategy s large, however. Sometmes a least commtment approach to element selecton works very well. Sometmes t works very poorly. However, on average ths strategy performs well. VII. CONCLUSIONS Ths paper has presented a method for mult-robot communcaton-senstve reconnassance. More generally, t has outlned an approach by whch larger tasks can be decomposed nto smaller tasks and planned for by usng nternalzed plans. Experments usng ths method demonstrate mproved performance over other control system experments, and llustrate the effect of team sze, plan selecton, and scenaro. Three nterestng characterstcs of ths approach are worth notng: 1) Plannng s offlne n the sense that each plan element s computed a pror. Ths lmts the planner to the elements that

8 have been created n advance but affords reactve utlzaton of each element. 2) Plannng s onlne n the sense that the plan controller may select elements dynamcally at runtme. Ths allows for dynamc reconfguraton of the plan and greater adaptablty. 3) Several plans are utlzed n parallel allowng the type of advce to be altered as envronmental condtons change. It could be clamed that smple teraton through the same seres of wayponts by a team of robots would lkely maxmze our performance metrcs. Ths s ndeed the case. However, the prmary am of ths work s the development of a technque sutable for battlefeld scenaros. These stuatons demand the ablty to deal wth uncertan, unpredctable condtons and also utlze a pror nformaton as much as possble. Equally mportant s the ablty to react opportunstcally to uncertanty. Ths approach s opportunstc n that plan elements can be managed dynamcally; precompled vector felds lmt the tendency for unnecessary wayponts; and changes n the envronment are reflected n output advce. Smple teraton through a seres of wayponts s ll equpped to handle dynamc, unpredctable envronments and lacks the ablty to act opportunstcally. Currently, plan elements and ther assocated machnery are desgned by hand. Ths lmts the scalablty of ths approach wth respect to the number of tasks. Scalablty n terms of team sze, terran sze, and nstantatons of a partcular task are not smlarly lmted. Hence ths approach may be of value when the deployment stuaton could nclude arbtrary team szes, terran sze, or repettons of smlar tasks. In the future we hope to extend ths approach to larger teams of robots, nvestgate more generalzed and vared types of tasks, and examne the system s performance on real robots. ACKNOWLEDGMENT Ths research s funded under DARPA/DOI contract #NBCH1212 as part of the MARS Vson 22 program. Ths program s a jont effort of Georga Insttute of Technology, The Unversty of Pennsylvana, Unversty of Southern Calforna, and BBN Technologes. We would lke to thank the staff of the McKenna MOUT ste at Fort Bennng GA for allowng us to test our research at ther facltes. REFERENCES [1] R.C. Arkn and T. Balch, AuRA: Prncples and Practce n Revew, Journal of Expermental and Theorretcal Artfcal Intellgence, 9(2): , [2] R.C. Arkn, Behavor-Based Robotcs, MIT Press, Cambrdge, MA [3] T. Balch and R.C. Arkn, Communcaton n Reactve Multagent Robotc Systems, Autonomous Robots, 1(1):27-52, [4] R.A. Brooks, Intellgence wthout Reason, Artfcal Intellgence, 47: , [5] J. Connell, SSS: A Hybrd Archtecture appled to Robot Navgaton, Proc. IEEE Intern. Conf. on Robotcs and Automaton, pp , [6] C.P. Dehl, M. Saptharsh, J.B. Hampshre II, and P. Khosla, Colaboratve Survellance Usng Both Fxed and Moble Unattended Ground Sensor Platforms, SPIE 13 th Internatonal Conf. on Aerospace/Defense Sensng, Smulaton, and Controls, Vol. 3713, Aprl, pp [7] E. Gat, Three-Layer Archtectures, n Artfcal Intellgence and Moble Robots: Case Studes of Successful Robot Systems, pp , MIT Press, Menlo Park CA, [8] K. Konolge, A Gradent Method for Realtme Robot Control, Proc. IEEE Inter. Conf. on Intelgent Robotcs and Systems, pp , 2. [9] J.C. Latombe, Robot Moton Plannng, Kluwer Academc Publshers, Boston, [1] D.C. MacKenze, Desgn Methodology for the Confguraton of Behavor-Based Robots, Ph.D. Dss., College of Computng, Georga Inst. Of Tech., [11] D.Payton, J. Rosenblatt, D. Kersey, Plan Guded Reacton, IEEE Transactons on Systems, Man, and Cybernetcs, 2(6): , 199. [12] A. Ranganathan and S. Koeng, A Reactve Robot Archtecture wth Plannng on Demand, Proc. IEEE Inter. Conf. on Intelgent Robotcs and Systems, , 23. [13] A.R. Wagner and R.C. Arkn, Internalzed Plans for Communcaton- Senstve Robot Team Behavors, Proc. IEEE Inter. Conf. on Intelgent Robotcs and Systems, pp , 23. [14] Arkn, R.C., Navgatonal Path Plannng for a Vson-based Moble Robot, Robotca, Vol.7, pp.49-63, Element Pseudocode Plan Controller Pseudocode Gven feature vector V Intalze weght vector W = W = DecsonTree(V) Advce vector A = A x + A y where N A x = N A y = w Q w Q x y and Q = vector from nternal plan() from parallel plan return A whle Plan not complete retreve feature vector from robot controller f element complete select new element usng method m return advce vector A = Element(Feature vector) Element selecton Pseudocode f method m = least commtment Set current cost = for all remanng elements n plan f element cost < current cost current cost = element cost return Element(current cost) else f method = strct orderng return Element() Fgure 13. Pseudocode for plan controller operaton, element selecton, and element advce producton.