Realization of the Sensor Web Concept for Earth Scienceusing Mobile Robotic Platforms

Ayanna M. Howard, Brian Smith, Magnus EgerstedtSchool of Electrical and Computer EngineeringGeorgia Institute of Technology, Atlanta, GA

404-385-4824ayanna(ece.gatech.edu, brian(ece.gatech.edu, magnus(ece.gatech.edu

Abstract In this paper, we discuss the realization of arobotic mobile sensor network that allows for controlledreconfiguration of sensor assets in a decentralized manner.The motivation is to allow the construction of a new systemof in-situ science observations that requires higher spatialand temporal resolution models that are needed forexpanding our understanding of Earth system change. Theseobservations could enable recording of spatial and temporalvariations in environmental parameters required for suchactivities as monitoring of seismic activity, monitoring ofcivil and engineering infrastructures, and detection of toxicagents throughout a region of interest. The difficulty inestablishing these science observations are that globalformation properties must be achieved based on the localinteractions between individual sensors. As such, wepresent a novel approach that allows for the sensor networkto function in a decentralized manner and is thus able toachieve global formations despite individual sensor failure,limitations in communication range, and changing scientificobjectives. Details on the sensing and control algorithms forcontrolled reconfiguration will be discussed and results offield deployment will be presented. 12

TABLE OF CONTENTS

1. INTRODUCTION ............. .........................12.BACKGROUND......................................2

3. SPIDERMOTE - ROBOTIC SENSOR NETWORK .....24. RECONFIGURATION OF THE NETWORK ................3

5.EXPERIMENTAL RESULTS ......................................46.CONCLUSIONS......................................5ACKNOWLEDGEMENTS ......................................5

REFERENCES .......... ............................5BIOGRAPHY ......... ............................6

1. INTRODUCTIONSensor networks (or webs) have been shown to be apowerful tool for in-situ Earth science applications rangingfrom earthquake forecasting to understanding climatechange [1]. These networks capitalize on their ability todeploy cheap nodes throughout a region of interest in orderto gather information relevant for scientific analysis.Recently, there has been growing interest in mobilenetworks to deal with the limitations of static networks,11 1-4244-0525-4/07/$20.00 C 2007 IEEE.2 IEEEAC paper #1099, Version 1, Updated December 1, 2006

including issues of network deployment, coverage, and faulttolerance. However, a number of issues still exist indeploying mobile sensor networks for Earth scienceapplications, including the effectiveness of adapting to theenvironment and to changing science requirements,balancing power usage, and selecting betweencommunication and control strategies.

To address current limitations, we discuss a naturalextension to the sensor web concept that enables controlledreconfiguration of sensor assets for fault-tolerant in-situsampling. The main motivation behind our approach is toapply decentralized (i.e. local) control algorithms fornetwork deployment while establishing the global sensingcapability required for Earth science investigations. Theintegrated sensing platform, which we refer to as aSpiderMote, combines hardware, in the form ofcommunication/sensor devices (motes), and simple mobilityplatforms (spiders) for re-positioning sensor devices inresponse to changes in science demand, sensor failure,and/or communication dropout. This system of mobilesensors is conceptually described as a decentralized networkof in-situ sensors. A desired science formation is achievedby defining distances between sensors that correlates withspecific topologies that the sensor network should assume.

In most sensor web applications, individual sensor agentscollect information about their environment and neighboringagents using peer-to-peer communication. Unfortunately, asthe size of the network increases, bandwidth limitations andthe absence of feasible communication channels severelylimits the possibility of conveying and using globalinformation. Thus, it cannot be assumed that each sensoragent has complete information about the states of everyother agent in the network. And yet, a network formation isinherently a global property and, as such, novel solutionsmust be implemented for the reconfiguration process tosuccessfully occur. Subsequently, in addition to thehardware elements that comprise the SpiderMotes, softwarecontrol is instituted for adaptive reconfiguration of thenetwork that changes the spatial resolution of the network,in effect establishing a self-adapting sensor network. Thisoccurs in order to maintain the desired science-drivenconfiguration in spite of changes in science demand.

In this paper, we will discuss realization of the sensor webconcept for Earth Science using mobile robotic platforms.In Section 2, we provide an overview of related work.Section 3 discusses the realization of the hardware platformused to construct the sensor web, while Section 4 documentsthe control algorithms for controlled reconfiguration of thesensor assets. Finally, results of deployment using themobile robotic platform is discussed in Section 5 withconclusions drawn in Section 6.

2. BACKGROUNDRecently, there has been growing interest in mobilenetworks to deal with the limitations of static networks. In[2], a system of mobile sensor networks that combinesdistributed robotics and low power embedded systems ispresented. The focus of this work is not on control, butrather on the development of an integratedhardware/software platform. [3] also presents thedevelopment of a new type of mobile (robotic) sensor nodeto meet the constraints of mobile sensor networks. In thiswork, a hardware paradigm was presented that balancesmodularity and efficiency in order to perform a navigationtask in a complex environment. In terms of control, theproblem of controlling multiple, mobile sensor agents in acoordinated fashion has received considerable attentionduring the last few years (e.g. 4, 5). In these efforts, theunderlying assumption has typically focused on havingcomplete knowledge for each individual agent such that thelocations and positions of the other agents in the networkare known. However, this is not always the case, especiallyas the number of sensors within the network grows, causingbandwidth and range limitations to invalidate this globalassumption. In [6], the design of localized algorithms andthe use of a directed diffusion model was presented to dealwith the limitations found in assuming complete knowledge.The focus in this effort was not on reconfiguration though,but rather on how to deal with communication limitationsarising from the sheer numbers of sensors within thenetwork. There has also been some work that addresses theproblem of decentralized control where individual agentsmove according to limited range potential fields or byaveraging orientation rules (e.g. 7, 8). However, theseefforts do not allow changes to desired formation and, asyet, little work has been done on how to choose formationsin a decentralized and autonomous fashion as a reaction tochanges in the environment.

To realize the sensor web concept for Earth Science, ourfocus is on controlled reconfiguration of sensor assets usingmobile robotic platforms. This differs from the concept ofmobility in Mobile Adhoc Networks (MANETs) in thatcontrolled mobility provides the network the ability to movesensors intentionally. The problem though is that themanner in which network sensors gain information aboutother sensors is inherently limited. Communication channelsare typically constrained and have bandwidth as well asrange limitations. As such, our method is based on

representing our network though the use of graph models,where an edge between two nodes (sensors) exists only ifthese sensors can share information through localinteractions. This construct, given a desired scienceformation, allows us to develop decentralized algorithms forrealizing a desired Earth science formation, whileguaranteeing that the agents stay connected, that they reachthe target science formation in a safe manner, and that theydetermine their individual locations in the network withoutaccess to global information. This decentralizedmethodology represents the novel nature of our approach.

3. SPIDERMOTE - ROBOTIC SENSOR NETWORK

Figure 1. SpiderMote construct for inclusion in the roboticsensor network

We define a sensor node within the network as an integratedsensing platform that combines communication and sensingcapability with mobility. The integrated sensing platform,which we refer to as a sensor agent, allows for controlled re-positioning of the sensor devices such that scienceobservations at desired spatial and temporal resolution canbe achieved. Specific attributes that must be included in theagent design require the following:

* Linear and rotational velocities of the sensor agentmust be controllable.

* Each sensor agent must possess a communicationchannel such that it has the ability to receive andtransmit information to other agents.

* Each sensor agent must possess a sensing channelsuch that it has the ability to verify the presenceand location of other sensor nodes within thenetwork.

* Each sensor agent must have on-board computationsuch that it can function independently and doesnot require a centralized host for decision-making.

2

This generic definition of a node allows us to define thesensor web concept without being limited to the utilizationof specific robotic hardware, communication, or sensingtechnology. Figure 1 depicts our rendition of a sensor agentwithin the robotic sensor network that qualifies under thisdefinition.

For sensors deployed for Earth Science applications, a keytechnical difficulty in instituting the sensor web concept iscompensating for the non-global characteristic of thenetwork (i.e. a decentralized configuration of sensors thatmust interface individually to achieve collective behavior).As such, we model our sensor network using graphformalisms [9-10], which allows us to represent therelationship of the sensor agents within the network. Thegraphs are used to model local interactions between sensorswithin a planar network, when individual sensor agents areconstrained by limited knowledge of other sensor agents.We define our network as consisting ofN agents, each withidentical dynamics and carrying a pre-assignedidentification tag n Etl,2,...,N] such that:

(5)

This implies that an agent can assume a maximum distancevalue if it can communicate with another agent, but can onlydetermine absolute distance and orientation if it can sensethe other agent. Using this construct, two agents can berepresented as having a fully connected relationship withinthe graph when6' yC and. Q c o

where xnJn defines each agent's position and On defines

the absolute heading, 6n is the distance between Agentn and

Agenti, V4 is the angle of robot Agentn to Agenti, and fi isn ~~~~~~~~~~~~~~~n

the relative angle ofAgentn to the heading /i of Agenti.

As this is a decentralized network of sensors, each agent can

communicate to other agents through a communicationchannel, having a limited range of fl. This communicationchannel has a bi-directional property and is designated as

C1i. In addition, each agent can sense another agentthrough a sensing channel, having a limited range of y and a

perception cone of centered around 0. The sensingchannel Sn,j between agents is a directed channel since,depending on the orientation of the agents' sensing cone,

Sn pi # Si n' In other words, although agent n may be ableto sense agent i, agent i may not be able to sense agent n.

Using this construct, we can represent the relationshipbetween agents by a graph in which the vertices of thegraphs represent agents, and connected edges are

established when agents are within communication or

sensing range of each other (Figure 2) such that:

Cn pi = Sni = Sin

(2)

(3)

(4)

_ _. -

Cn-i

Figure 2. Depiction oftwo sensor agents and theirassociated connection graphs

4. RECONFIGURATION OF THE NETWORKThe utilization of a robotic sensor network provides thesystem the ability to physically change sensor resolutionassociated with a desired network topology. This drivingneed can be as a result of changes in science demand, sensor

failure, and/or communication dropout. In this section, we

discuss our methodology of using decentralized controlstrategies based on local agent communication schemes toachieve global reconfiguration of the sensor networknecessary to change network topology.

We assume that specific formations that define desiredspatial resolutions are provided by the scientist afterdeployment of the network in the field. This networkconfiguration is classified as the desired science formation,

3

(1)Agentn :{Xnnynq}

X -X= ta -( n iXn -xi

On=4'n- On So that ipn = Q' + O

- kh.,WrIq P-0

n-i > 6n 17s ' < Yn-i > 6n

INW"q

which the current network configuration must bereconfigured to in order to align with science demands. Wesolve the reconfiguration problem by using a limitedconsensus algorithm [9] (Figure 3) that capitalizes on theknowledge available from both the communication andsensing channnckls. ol

rocity,~~~~ ItIPereieVe4inotlier

l Y Itoto ttoa_tr__ns_l____atio_nal

_ r S~~~~~~~~~elocitv

Figure 3. Logic flow for limited consensus controlalgorithm

In this coordination scheme, each sensor agent n tries tomaintain a desired distance between other sensor agentspresent within the network. A network formation error withrespect to the desired science formation under considerationis therefore defined as:

VnE{1,N}:ViEn+1,N}:

if SnivSin Total Error+=|i-aelseif Cni Total_Error+ = -a

where a designates the desired spatial resolution betweensensors, N is the number of sensor agents within thenetwork, and Total Error is the accumulation of distanceerrors derived from assumed knowledge based on thecommunication channel and sensed knowledge based on thesensing channel. The goal of the reconfiguration controlproblem is therefore to change network topology byminimizing the network formation error. To allow eachsensor agent to adjust the network topology, decentralizedcontrol laws are used to control the distance between robotsin the network. Each sensor agent has a controllabletranslational and rotational velocity, denoted as vi andawirespectively. The goal of the control law it to allow thesensor agents to achieve a desired spatial distances, suchthat:

4

Jv 'sign(61 -a) if QJi<d and 8.al° otherwise

O sign(/n) if ni< 0{C ( otherwise

where vo,Co are constants. In this decentralized method, ifno agent is perceived, the agent stops and rotates until itsees another agent(s) in the field. At that point, translationalvelocity achieves a nonzero value, and the agent movestoward the perceived agent(s) until the desired spatialdistance is achieved.

This proposed control strategy enables a desired scienceformation to be achieved with the use of mobile roboticplatforms. Using this approach, the network isautonomously reconfigured from any starting configurationto a desired network formation by executing the trajectoriescalculated.

5. EXPERIMENTAL RESULTSFor assessment of the methodology, we implement ourstrategy on the SpiderMote platform, which consists of amechanical body frame, high torque servos, vision sensor,communication module, and a controller. The skeletal bodyof the platform is equipped with six legs and has dimensionsof [7.5 x 26.5 x 24] cm. Movement of each unit isaccomplished using three HI-Tec HS-645MG high torqueservos to control movement of three sets of legs. One servois used to control a set of left legs (front and back) while theother controls a set of right legs (front and back). The thirdservo is used to control a center set of legs (left and right).When oriented correctly, this center set allows a vertical tiltof the unit from one side to the other. Each servo is orientedin-between its respective set of legs, such that each set iscoupled using a push-pull scheme. Thus, as a servo pushesone leg, it simultaneously pulls the other in the samedirection. The CMUCam2 vision sensor, with multiple portsfor servo control, provide the primary sensor data. Finally,a RidgeSoft IntellibrainTm robotics controller is used toimplement the algorithm and provide the control signals forcommanding robot movement.

Communication ChannelThe wireless control module is capable of transmitting andreceiving a single 10 byte message at any given time at arange of up to 150 meters. There are two types of messages:requests and responses. Because messages could potentiallybe lost, requests are reattempted after a timeout period untilthe appropriate response is received.

Sensing ChannelThe sensing channel consists of visual data retrieved fromthe CMUCam2 camera mounted on a 15cm boom. Also,each robot is equipped with a color-coded tetrahedral cover,which is mounted below the camera in order to provide eachsensor agent with a unique identification tag (Figure 4). Thisallows the robots to identify each other, and to determine therelative position of other robots that are in camera view. Theassociated perception cone e9 is approximately 7t/7 radianscentered around ).

Figure 4. Uniquely tagged SpiderMote sensor agent

To validate our approach, a range of experimental runs wereimplemented in various terrain environments as depicted inFigure 5. Runs consisted of starting the agents at randominitial positions and orientations, and commanding them toachieve a formation having discrete separation distances.The maximum initial separation value between robots wasapproximately 5 meters. Corresponding final separationdistances ranged to approximately 1.5 meters. Averageexecution time was 2-3 minutes to achieve the desiredscience formation. In these experimental cases, the robotswere able to converge to a final configuration with anaccuracy of + 0.3 meters with respect to the desiredseparation distance.

Figure 5. Snapshot of experimental run segment

6. CONCLUSIONSIn this paper, we discuss realization of the sensor webconcept for Earth Science using mobile robotic platforms.By representing our network though the use of graphmodels, where an edge between two sensors exists only ifthese sensors can share information through localinteractions, we can develop decentralized controlalgorithms for realizing a desired Earth science formation.The motivation is to provide a means for sensor agents toachieve distances associated with the target scienceformation in a robust manner without access to globalinformation. We have integrated our methodology with theSpiderMote system, which acts as our sensor agent withinthe robotic sensor network. Future work will involvefurther deployment of the agents in hazardous terrainenvironments.

ACKNOWLEDGEMENTSThis work was performed at the Georgia Institute ofTechnology and supported under a grant from the NationalAeronautics and Space Administration, Earth ScienceTechnology Office, AIST Program.

REFERENCES1. NSF Report: "National Workshop on Future Sensing

Systems - Living, Nonliving, and Energy Systems,"Lake Tahoe, NV, August 2002.

2. K. Dantu, et. al, "Robomote: Enabling Mobility inSensor Networks", 4th Int. Symp. on InformationProcessing in Sensor Networks, pp. 404-409, 2005.

3. J. Friedman, et. al., "RAGOBOT: A New HardwarePlatform for Research in Wireless Mobile Networks,"4th Int. Conf. on Information Processing in SensorNetworks, 2005.

4. R.W. Beard, J. Lawton, F.Y. Hadaegh, "A CoordinationArchitecture for Spacecraft Formation Control," IEEETran. on Control Systems Technology, Vol. 9, pp. 777-790, 2001.

5. P. Ogren, E. Fiorelli, and N.E. Leonard, "FormationsWith a Mission: Stable Coordination of Vehicle GroupManeuvers," Proc. 15th Int. Symp. on MathematicalTheory ofNetworks and Systems, 2002.

6. D. Estrin, R. Govindan, J.S. Heidemann, and S. Kumar,"Next Century Challenges: Scalable Coordination inSensor Networks", Mobile Computing and Networking,pp.263 270,1999.

7. J.R. Lawton, et. al., "A Decentralized Approach toElementary Formation Maneuvers," Proc. IEEE Int.Conf. on Robotics and Automation, Vol. 3, pp. 2728-2733,2000.

8. R. Olfati-Saber and R. Murray, "DistributedCooperative Control of Multiple Vehicle Formationsusing Structural Potential Functions," 15th IFAC WorldCongress, Barcelona, 2002.

9. B. Smith, J. McNew, M. Egerstedt, E. Klavins, A.Howard, "Embedded Graph Grammers (EGGs) for

5

Multi-Robot Coordination," submitted to IEEE Int.Conference on Robotics and Automation, April 2007.

10. A. Muhammad and M. Egerstedt, "Connectivity Graphsas Models of Local Interactions," Journal of AppliedMathematics and Computation, Vol. 168, No. 1, pp.243-269, Sept. 2005.

BIOGRAPHYAyanna Howard is an Associate Professor at the Georgia

Institute of Technology. Her area ofresearch is centered around theconcept ofhumanized intelligence, theprocess of embedding humancognitive capability into the controlpath of autonomous systems. Thiswork, which addresses issues ofautonomous control as well as

aspects of interaction with humansand the surrounding environment, has

resulted in over 60 written works in a number ofprojects -

from autonomous rover navigation for planetary surfaceexploration to intelligent terrain assessment algorithms forlanding on Mars. To date, her unique accomplishments havebeen documented in over 12 featured articles - includingbeing named as one of the world's top young innovators of2003 by the prestigious MIT Technology Review journaland in TIME magazine's "Rise of the Machines" article in2004. Dr. Howard received the IEEE Early Career Awardin Robotics and Automation in 2005 and is a SeniorMember ofIEEE.

Magnus Egerstedt was born in Stockholm, Sweden. Hereceived the B.A. degree inphilosophy from StockholmUniversity in 1996, and the MS.degree in engineering physicsand the Ph.D. degree in appliedmathematics, both from the RoyalInstitute Of Technology,Stockholm, in 1996 and 2000,respectively. He spent 2000-2001

as a Postdoctoral Fellow with the Division ofEngineering and Applied Science, Harvard University,Cambridge, MA. In 1998, he was a Visiting Scholar at theRobotics Laboratory at the University of California,Berkeley. He is currently an Associate Professor with theSchool of Electrical and Computer Engineering, theGeorgia Institute of Technology, Atlanta. His researchinterests include optimal control as well as modeling andanalysis ofhybrid and discrete-event systems, with emphasison motion planning and control of (teams ofi mobile robots.He has authored over 100 articles in the areas of roboticsand controls. Dr. Egerstedt received the CAREER Awardfrom the National Science Foundation in 2003 and is aSenior Member ofIEEE.

Brian S. Smith grew up in southern Georgia, graduatedvaledictorian from Flint RiverAcademy, and went on to earn hisBS in Computer Engineering fromthe Georgia Institute of Technologyin 2004. During his undergraduatecareer, Brian assisted teachingobject-oriented programming and adigital design laboratory course.

He also assisted Dr. Alan Doolittlefor three years at the Advanced

Semiconductor Research Facility at Georgia Tech, andassisted Dr. Sung-Kyu Lim for a year in CAD research atthe GTCAD laboratory, earning a President'sUndergraduate Research Awardfor research in CAD VLSIapplications in 2004. Brian began his graduate work in2005, focusing on multi-agent control with roboticsapplications. He is currently building a multi-robotnetwork, while also developing decentralized controlalgorithms for such networks. His advisors are Dr. AyannaHoward at the Human-Automation Systems (HUMANS)Lab, and Dr. Magnus Egerstedt at the GRITS Lab.

6