robot boats as a mobile aquatic sensor networkbryanlow/pubs/essa2009.pdf · robot boats as a mobile...

Robot Boats as a Mobile Aquatic Sensor Network

Kian Hsiang Low†, Gregg Podnar‡, Stephen Stancliff‡,John M. Dolan‡, and Alberto Elfes§

Dept. Electrical and Computer Engineering†, Robotics Institute‡, Carnegie Mellon University5000 Forbes Avenue Pittsburgh PA 15213

NASA Jet Propulsion Lab§

MS:198-235 4800 Oak Grove Drive Pasadena CA 91109{bryanlow, gwp, stancliff, jmd}@cs.cmu.edu, [email protected]

ABSTRACTThis paper describes the Multilevel Autonomy RobotTelesupervision Architecture (MARTA), an architecturefor supervisory control of a heterogeneous fleet of net-worked unmanned autonomous aquatic surface vesselscarrying a payload of environmental science sensors.This architecture allows a land-based human scientist toeffectively supervise data gathering by multiple roboticassets that implement a web of widely dispersed mobilesensors for in situ study of physical, chemical or bio-logical processes in water or in the water/atmosphereinterface.

Categories and Subject DescriptorsI.2.9 [Robotics]: autonomous vehicles, sensors; I.2.11[Distributed Artificial Intelligence]: intelligentagents, multiagent systems

General TermsAlgorithms, Design, Experimentation, Human Factors

Keywordstelesupervision; autonomous surface vessels (ASV); sen-sor web; inference grids; log-Gaussian process.

1. INTRODUCTIONEarth and ocean science research use data obtained

from space, the atmosphere, the land, surface water,and the ocean to improve understanding of the Earthand its natural processes. Developing better models ofocean processes is crucial for assessing global warming

Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.ESSA Workshop ’09, April 16, 2009, San Francisco, California, USACopyright 2009 ACM 978-1-60558-533-8/09/04 ...$5.00.

and for meteorological and ecological studies, while sur-face water quality data give us insight into the suit-ability of the water for human use. Both are necessaryfor assessing water’s ability to support aquatic life. Atleast three in every four humans depend on “surface wa-ter as their primary source of drinking water”1. Waterresources in a region are critical to economic growth aswell as to quality of life.

Freshwater sensing is typically done with fixed sensorsnear tributaries and using periodic sampling. Oceansensing is typically done with satellites, buoys, andcrewed research vessels. Satellites are limited by cloudcover, temporal and geographical coverage, and resolu-tion. Manned research vessels are expensive to deploy,and fixed sensors and anchored buoys cannot be eas-ily moved to cover larger areas or to monitor regions ofincreased interest.

Deploying mobile sensor platforms to augment fixedsensors and sensor buoys will help us better understandand model how contaminants/pollutants move and af-fect our environment, from industrial chemical and oilspills, to Harmful Algal Blooms (HABs). These plat-forms will also enable in situ meteorological studies ofthe atmosphere/ocean interface for hurricane prediction.

This paper describes a multi-robot science explorationsoftware architecture and system called Multilevel Au-tonomy Robot Telesupervision Architecture (MARTA)[11]. MARTA was initially developed for supervisorycontrol of multiple robot assets exploring the lunar sur-face [4, 8, 12]. It has subsequently been expanded andinstantiated in a system called TAOSF (TelesupervisedAdaptive Ocean Sensor Fleet) for coordination and con-trol of multiple robot boat exploration vehicles for oceanic[5, 6, 13] and freshwater research [10]. For the oceanboats, we have selected as an application the character-ization of HABs [1], while for the freshwater vehicles, weare concentrating on their use in water quality studies.

1National Research Council. Earth Science and Applicationsfrom Space: National Imperatives for the Next Decade andBeyond, 2007.

!"#$%&"' ("!"&)*+",-.' */0&)/(' $,-"12/-$),' /,(' -"3-$,1.'

)*"2/-$),3.' /,(' +/$,-",/,%"' )4' -#"' 56787' 4&""-' /,('

12)9,('303-"+3:''

;#"' 56787' *&/-4)2+' $3' /**2)<$+/-"&0' =>' 4""-' &),1' /,('

?"$1#3' @93-' )!"2' ABBB' &C3:' ;#"' !"#$%&"' #/3' /' */0&)/('

%/*/%$-0')4'DBB'&C3.'/,('$3'("3$1,"('-)'C"'3"&4E2$1#-$,1'-)'

",392"' 392!$!/C$&$-0' $,' #"/!0' 3"/3:' ' 8-' 39**)2-3' /' ?$("'

2/,1"')4'%)++9,$%/-$),' &$,F3' $,%&9($,1'3*2"/('3*"%-29+'

2/($).'/'%"&&9&/2'*#),"'&$,F.'/,('/,'82$($9+'3/-"&&$-"'&$,F:'

'

!"#$%&'()'*+,-,'./0123%4'5"16'0143786&%"9'0:;'7&0'501&%'

7&:73%7'";&:1"2"&;)'

'

;?)' *&/-4)2+3' G,/+"(' 56787=' /,(' 56787HI' /2"'

%922",-&0' 9,("21)$,1' -"3-$,1' /-' J676' KLL' /,(' ?$&&'

39**)2-' )*"2/-$),3' 4)2' -#"' ;657L' *2)@"%-:' 6(($-$),/&'

*&/-4)2+3' /2"' 9,("2' *2)(9%-$),:' 56787' 3#/F"()?,'

)*"2/-$),3' #/!"' C"",' *"24)2+"(' 3$,%"' "/2&0' HBBD' $,' -#"'

?/-"23' )4' -#"' MNOP6QR6' 2"1$),.' $,%&9($,1' -#"'

S#$,%)-"/19"'T/0'/,('U)%)+)F"'7)9,(:';#"'4$23-')*",E

)%"/,'("*&)0+",-')4'-#"'56787'303-"+'?/3'*"24)2+"('$,'

J)!"+C"2' HBBV:' M92$,1' -#$3' )*"2/-$),.' -#"' 56787H'

*&/-4)2+'39%%"3349&&0',/!$1/-"(')!"2'>',/9-$%/&'+$&"3'),'/'

-2/,3"%-' &$,"' "3-/C&$3#"(' $,' -#"' 6-&/,-$%' 5%"/,' )44' -#"'

%)/3-' 42)+' KLL:' ' 56787=' /,(' 56787H' /2"' %922",-&0'

9,("21)$,1' 9*12/("3.' 3",3)2' $,-"12/-$),.' /,(' -"3-$,1' $,'

*2"*/2/-$),'4)2'",(92/,%"'-2$/&3'/,('3%$",%"')*"2/-$),3:''''

7",3)23' ),C)/2(' -#"' 56787' *&/-4)2+3' ",/C&"' -#"'

%)&&"%-$),' )4' ?/-"2' 3/&$,$-0' /,(' %),(9%-$!$-0' (/-/.' 3"/'

3924/%"' -"+*"2/-92".' /,(' %#&)2)*#0&&' +"/392"+",-3:' 6'

2#)(/+$,"'4&9)2)+"-"2'?/3'$,-"12/-"('-)'39**)2-'+/**$,1'

)*"2/-$),3'(92$,1'(0"'("*&)0+",-' -"3-3:'M0"'$3'("*&)0"('

-)' 3$+9&/-"' W6T3' -)' /&&)?' ("!"&)*+",-' /,(' -"3-$,1' )4'

,/!$1/-$),'/,('+/**$,1'/&1)2$-#+3:';#"'4)2?/2('*/0&)/('

C/0'*2)!$("3' 3*/%"' 4)2' $,3-/&&/-$),')4' /(($-$),/&' 3",3)23:'

;#$3' C/0' $,%&9("3' /' ?/-"2' 4&)?E-#2)91#' 303-"+' ?$-#'

+/,$4)&(3' /,(' /' ("EC9CC&$,1' 303-"+' -#/-' 3$+*&$4$"3'

$,3-/&&/-$),')4',"?'3",3)23:''

''''6' +/3-E+)9,-"(' +"-")2)&)1$%/&' 3-/-$),' /&&)?3'

/%X9$3$-$),' )4' /-+)3*#"2$%' +"/392"+",-3.' $,%&9($,1'

C/2)+"-2$%' *2"3392".' /$2' -"+*"2/-92".' 2"&/-$!"' #9+$($-0.'

?$,('3*""(.'/,('?$,('($2"%-$),:'56787'$3'/&3)'"X9$**"('

?$-#'/'4)2?/2(E&))F$,1'($1$-/&' $+/1$,1'303-"+'*2)!$($,1'

2"+)-"&0' &)%/-"(' 3%$",-$3-3' ?$-#' $+/1"3' )4' /-+)3*#"2$%'

/,('3"/'3-/-"'%),($-$),3:''

'''';#"' )44EC)/2(' $,42/3-29%-92"' ("!"&)*"(' C0' N7;' $3'

F,)?,' /3' -#"' 56787' P$33$),' 5*"2/-$),3' N,!$2),+",-'

GP5NI:';#"'P5N'2"3$("3' $,' -#"'K/&&)*3'S)/3-/&'5%"/,'

5C3"2!/-$),' O/C)2/-)20' GK/S55OI' %),-2)&' 2))+' /,('

*2)!$("3' /**&$%/-$),3' /,(' 3"2!$%"3' -#/-' ",/C&"' -#"'KLL'

",1$,""2$,1' /,(' 3%$",%"' )*"2/-$),3' -"/+' -)' *"24)2+'

*&/-4)2+'%)++/,($,1' /,(' -"&"+"-20'+),$-)2$,1.' /3'?"&&'

/3'%)++9,$%/-$),3'+/,/1"+",-:';#"'P5N'/&3)'*2)!$("3'

/'+$((&"?/2"'$,-"24/%"'-)'",/C&"')-#"2'%93-)+"23.'39%#'/3'

-#"' P6Q;6Y;657L' *2)@"%-.' -)' $,-"12/-"' ,"?' 303-"+3'

-#/-'492-#"2'",#/,%"'56787'3%$",%"')*"2/-$),3:''

8,'-#"'%922",-'("*&)0+",-'%),4$192/-$),.'C)-#'",1$,""2$,1'

-"&"+"-20'GZU7'*)3$-$),.'2)&&.'*$-%#.'0/?.'C/--"20'!)&-/1".'

[%:I'/,('3%$",%"'3",3)2'(/-/'/2"'%)++9,$%/-"('C"-?"",'

"/%#'2)C)-$%'*&/-4)2+'/,('J676\3'Z)((/2('7*/%"'L&$1#-'

S",-"2'!$/' -#"' 8,-"2,"-:' ' 8-' $3' 42)+'-#$3'($3-2$C9-$),'*)$,-'

-#/-' J566' ?"/-#"2' 2"3"/2%#"23' ?$&&' 2"%"$!"' )%"/,' /,('

)%"/,Y/-+)3*#"2"'$,-"24/%"'3",3)2'(/-/:''

<)<! =6&'>3?31"9',&:73%'@3017'A>,@7B'L)2'-#"'+)C$&"'("*&)0+",-')4'42"3#?/-"2'X9/&$-0'3",3)23.'

/'4&""-')4'3+/&&'/,('2"&/-$!"&0'$,"<*",3$!"'Q)C)-$%'7",3)2'

T)/-3' GQ7T3I' $3'C"$,1'("!"&)*"('/-'SP]'GL$1:'HI:' ';#"'

("3$1,' $3' C/3"(' ),' %)++"2%$/&&0' /!/$&/C&"' %)+*),",-3'

39%#' /3' -#"' #9&&.' (2$!"' 303-"+.' /,(' %)++9,$%/-$),3'

303-"+:' S93-)+$^/-$),' $3' F"*-' -)' /' +$,$+9+' ?$-#' -#"'

*2$+/20' 4)%93' C"$,1' F"*-' ),' -#"' 3",3$,1' /,(' ,/!$1/-$),'

2"X9$2"+",-3:'

;#"' #9&&' $3' /' 2"&/-$!"&0' $,"<*",3$!"' 2)-)E+)&("('

2"%2"/-$),/&'F/0/F'-#/-'$3'4$--"('?$-#'/'(9%-"('-#293-"2'),'

"/%#' 3$(":' ' ;#"' -#293-"23' /2"' C/3"('),'*"23),/&'7S]T6'

*2)*9&3$),'9,$-3'?$-#'3%2"",3'-)'"<%&9("'4)2"$1,'+/-"2$/&3'

/,(' -)' *2)!$("' 3/4"-0' -)' #9+/,3' )*"2/-$,1' $,' -#"' 3/+"'

/2"/:' ' ' ;#"' (9/&' -#293-"23' /&&)?' 3$+*&"' ($44"2",-$/&'

3-""2$,1'/,('*2)!$("'/("X9/-"'3*""('?#",'("*&)0$,1'"/%#'

C)/-' -)' $-3' 3/+*&$,1' /2"/:' ' 6' */%F/1"' )4' -"&"+"-20' /,('

?/-"2' X9/&$-0' 3",3)23' $3' 39**)2-"(:' ' 6' 3"/&"(' 3*$2/&E%"&&'

(""*E%0%&"' +/2$,"' C/--"20' *2)!$("3' *)?"2' /("X9/-"' 4)2'

29,3' )4' +/,0' #)923' /,(' /&3)' /%-3' /3' C/&&/3-:' ;"&"+"-20'

Figure 1: OASIS Platform with atmospheric andsea water sensors identified.

In the following sections, we will discuss the overallMARTA architecture, describe the robot boats beingdeveloped and used, and describe field tests conductedunder controlled conditions. We conclude with an out-line of the next steps in the development of MARTA.

2. ROBOT BOAT PLATFORMS

2.1 The OASIS Ocean PlatformsThe OASIS (Ocean-Atmosphere Sensor Integration

System) vessels are long-duration solar-powered autono-mous surface vehicles (ASVs), designed for global open-ocean operations (Figure 1). Their development isfunded by the National Oceanic and Atmospheric Ad-ministration (NOAA). The NOAA-funded OASIS Plat-form Build Team, which consists of EG&G, Zinger En-terprises, and Emergent Space Technologies, providesvehicle development, payload integration and testing,operations, and maintenance of the OASIS fleet andground systems.

The OASIS platform is approximately 18 feet long andweighs just over 3000 lbs. The vehicle has a payloadcapacity of 500 lbs, and is designed to be self-rightingto ensure survivability in heavy seas. It supports a widerange of communication links including spread spectrumradio, a cellular phone link, and an Iridium satellite link.

Three platforms (named OASIS-1, OASIS-2, andOASIS-3) are operated by NASA WFF and support op-erations for the TAOSF project. OASIS shakedown op-erations have been performed since early 2005 in thewaters of the DELMARVA region, including the Chin-coteague Bay and Pocomoke Sound. The first open-

ocean deployment of the OASIS system was performedin November 2006. During this operation, the OASIS-2platform successfully navigated over 8 nautical miles ona transect line established in the Atlantic Ocean off thecoast from WFF.

Sensors onboard the OASIS platforms enable the col-lection of water salinity and conductivity data, sea sur-face temperature, and chlorophyll measurements. Arhodamine fluorometer was integrated to support map-ping operations during TAOSF dye deployment tests.Dye is deployed to simulate HABs to allow develop-ment and testing of navigation and mapping algorithms.The forward payload bay provides space for installationof additional sensors. This bay includes a water flow-through system with manifolds and a de-bubbling sys-tem that simplifies installation of new sensors.

A mast-mounted meteorological station allows acqui-sition of atmospheric measurements, including baromet-ric pressure, air temperature, relative humidity, windspeed, and wind direction. OASIS is also equipped witha forward-looking digital imaging system providing re-motely located scientists with images of atmosphericand sea state conditions.

The off-board infrastructure developed by EST isknown as the OASIS Mission Operations Environment(MOE). The MOE resides in the Wallops Coastal OceanObservation Laboratory (WaCOOL) control room andprovides applications and services that enable the WFFengineering and science operations team to perform plat-form commanding and telemetry monitoring, as well ascommunications management. The MOE also providesa middleware interface to enable other customers, suchas the MARTA/TAOSF project, to integrate new sys-tems that further enhance OASIS science operations.

In the current deployment configuration, both engi-neering telemetry (e.g., GPS position, roll, pitch, yaw,battery voltage) and science sensor data are communi-cated between each robotic platform and NASA’s God-dard Space Flight Center via the Internet. It is fromthis distribution point that NOAA weather researcherswill receive ocean and ocean/atmosphere interface sen-sor data.

2.2 The Robotic Sensor Boats (RSBs)For the mobile deployment of freshwater quality sen-

sors, a fleet of small and relatively inexpensive RoboticSensor Boats (RSBs) is being developed at CMU (Fig-ure 2). The design is based on commercially availablecomponents such as the hull, drive system, and commu-nications system. Customization is kept to a minimumwith the primary focus being kept on the sensing andnavigation requirements.

The hull is a relatively inexpensive roto-molded recre-ational kayak that is fitted with a ducted thruster on

and science data, and well as control signals are all handled over a cellular modem. To develop and test the control and communication packages while the RSB is being built, a smaller platform was adapted (Fig. 3).

Figure 2. Robot Sensor Boat (RSB) under development.

Figure 3. Initial testing of the telemetry and communications

package on a small (1m length) platform.

3. THE MARTA ARCHITECTURE The MARTA system architecture (Fig. 4) provides an integrated approach to multi-robot coordination and multi-level robot-human autonomy. It allows multiple robotic sensing assets (both mobile and fixed) to function in a cooperative fashion, and the operating mode of different robot platforms to vary from full autonomy to teleoperated control.

High-level planning and monitoring allows a human telesupervisor to assign to a fleet of robotic assets high-level goals, such as specifying an area of ocean to investigate. The steps to achieve the goals are planned, and operational commands sent to each robot by the Robot Team Coordinator. As the robots execute these plans, their operation is monitored both by the Robot Team Coordinator module and by the human telesupervisor. Adaptive replanning of the robot assignments is based on sensor inputs (dynamic sensing) and coordination between multiple assets, thereby increasing data-gathering effectiveness while reducing the human effort required for tasking, control, and monitoring of the vehicles. The assets depicted in Fig. 4 include two OASIS platforms, a drift buoy that reports its position as it moves with the current, and a Robot Sensor Boat for surface water operations. Multi-level autonomy includes low-level autonomy on each independently operating robot; autonomous monitoring of the fleet; adaptive replanning; and when necessary, intervention by the human telesupervisor either with manual replanning, or by taking direct control of a robot via teleoperation. Algorithms for science analysis of the acquired data can perform an initial assessment of the presence of specific science signatures of immediate interest. This can be done both onboard each robot and at the telesupervisor’s workstation. Web-based communications support both control and communications of the robotic fleet over long distances, and the sharing of newly sensed and historical data with remote experts.

Figure 4. MARTA Telesupervision Architecture

4. MAPPING HARMFUL ALGAL BLOOMS (HABS) Interest in Harmful Algal Bloom (HAB) detection has grown in recent years for scientific, commercial and

Figure 2: Robot Sensor Boat (RSB) under de-velopment.









Figure 3: Initial testing of the telemetry andcommunications package on a small (1m length)platform.each side. The thrusters are based on personal SCUBApropulsion units with screens to exclude foreign mate-rials and to provide safety to humans operating in thesame area. The dual thrusters allow simple differen-tial steering and provide adequate speed when deployingeach boat to its sampling area. A package of telemetryand water quality sensors is supported. A sealed spiral-cell deep-cycle marine battery provides power adequatefor runs of many hours and also acts as ballast. Teleme-try and science data, and well as control signals are allhandled over a cellular modem. To develop and test thecontrol and communication packages while the RSB isbeing built, a smaller platform was adapted (Figure 3).

3. THE MARTA ARCHITECTUREThe MARTA system architecture (Figure 4) provides









Figure 4: MARTA Telesupervision Architecture.

an integrated approach to multi-robot coordination andmulti-level robot-human autonomy. It allows multiplerobotic sensing assets (both mobile and fixed) to func-tion in a cooperative fashion, and the operating modeof different robot platforms to vary from full autonomyto teleoperated control.

High-level planning and monitoring allows a humantelesupervisor to assign to a fleet of robotic assets high-level goals, such as specifying an area of ocean to in-vestigate. The steps to achieve the goals are planned,and operational commands are sent to each robot bythe Fleet Planning and Navigation Monitoring. As therobots execute these plans, their operation is monitoredboth by the Fleet Planning and Navigation Monitor-ing module and by the human telesupervisor. Adaptivereplanning of the robot assignments is based on sen-sor inputs (dynamic sensing) and coordination betweenmultiple assets, thereby increasing data-gathering effec-tiveness while reducing the human effort required fortasking, control, and monitoring of the vehicles. The as-sets depicted in Figure 4 include two OASIS platforms,a drift buoy that reports its position as it moves withthe current, and a Robot Sensor Boat for surface wateroperations.

Multi-level autonomy includes low-level autonomy oneach independently operating robot; autonomous mon-itoring of the fleet; adaptive replanning; and when nec-essary, intervention by the human telesupervisor eitherwith manual replanning, or by taking direct control ofa robot via teleoperation.

Algorithms for science analysis of the acquired datacan perform an initial assessment of the presence of spe-cific science signatures of immediate interest. This canbe done both onboard each robot and at the telesupervi-sor’s workstation. Web-based communications supportboth control and communications of the robotic fleetover long distances, and the sharing of newly sensedand historical data with remote experts.

4. MAPPING HARMFUL ALGAL BLOOMS(HABS)

Interest in Harmful Algal Bloom (HAB) detection hasgrown in recent years for scientific, commercial publichealth reasons. Depending on the type of algae present,HABs have been shown to be dangerous to sea life andto human health.

There is a significant interest in identifying environ-mental factors that contribute to the occurrence of HABs,so that these may be incorporated in bloom predic-tion algorithms. These factors may include time ofyear, salinity, and sea-surface temperature to predict theabundance (low, medium, or high) and type of HABs.

MARTA will provide the following advantages overexisting systems for observing and analyzing HABs:• Dynamic tasking and adaptation;• Higher in situ resolution and greater insensitivity to

cloud cover in comparison to current satellite systems;• Access to and greater agility in coastal waters than

what is available through buoys;• Real-time multipoint science data observations and

generation of associated interpretations by remotelylocated oceanographers.

Since HABs are sporadic ocean phenomena, we are us-ing rhodamine WT (water-tracing) dye as a HAB sur-rogate for initial experiments and tests.

For investigating an algal bloom, researchers at Carne-gie Mellon University take control of the OASIS plat-forms and provide high-level planning and monitoring.The OASIS collection of weather-related data mentionedearlier is not interrupted, but the additional sensing forinvestigating the algal bloom is brought online to mapthe extent of a bloom, chlorophyll concentrations, andwith additional sensors, eventually allow a shore-basedbiologist the ability to assess whether the algal speciesare harmful. If they are, local authorities can be notifiedat aquaculture businesses, fisheries, and beaches.

5. FIELD TESTS

5.1 Field Test 1 - Aug. 2007For controlled experiments such as mapping while com-

pensating for drift currents, geometric patches of rho-damine water-tracing dye are laid (Figures 5 and 6).An aerostat, tethered to a manned tender boat, carriesan instrument package aloft including GPS, altimeter,compass, and video camera to provide validation imagesof data gathered by OASIS platforms that are mappingwith fluorometers.

Methodology: Inference Grids. Observation of nat-ural processes is limited by spatial and temporal sensorfootprint, coverage, resolution, sampling rate, and mea-

public health reasons. Depending on the type of algae present, HABs have been shown to be dangerous to sea life and to human health. There is a significant interest in identifying environmental factors that contribute to the occurrence of HABs, so that these may be incorporated in bloom prediction algorithms. These factors may include time of year, salinity, and sea-surface temperature to predict the abundance (low, medium, or high) and type of HABs. MARTA will provide the following advantages over existing systems for observing and analyzing HABs:

! Dynamic tasking and adaptation ! Higher in situ resolution and greater insensitivity to

cloud cover in comparison with current satellite systems

! Access to and greater agility in coastal waters than what is available through buoys

! Real-time multipoint science data observations and generation of associated interpretations by remotely located oceanographers.

Since HABs are sporadic ocean phenomena, we are using rhodamine WT (water-tracing) dye as a HAB surrogate for initial experiments and tests. For investigating an algal bloom, researchers at Carnegie-Mellon University take control of the OASIS platforms and provide the high-level planning and monitoring. The OASIS collection of weather-related data mentioned earlier is not interrupted, but the additional sensing for investigating the algal bloom is brought online to map the extent of a bloom, chlorophyll concentrations, and with additional sensors, eventually allow a shore-based biologist the ability to determine the species of the algae to assess whether it is harmful. If it is, local authorities can be notified at aquaculture businesses, fisheries, and beaches.

5. INFERENCE GRIDS Observation of natural processes is limited by spatial

and temporal sensor footprint, coverage, resolution, sampling rate, and measurement uncertainty. Markov Random Fields (MRFs) provide a natural formulation to represent spatially and temporally distributed observations. We use a discretized version of MRFs, called a spatio-temporal Markov Random Lattice (ST-MRL), to encode the data obtained by the different sensors and agents. Each cell in the lattice corresponds to a spatial volume and a time slice, and stores a stochastic vector with the probabilistic state description of the various processes that have been measured at the given location and time interval. Efficient updating methods are used to improve the probabilistic descrption encoded in

the lattice as new observations flow in from the various sensors and platforms being used [11].

Associated with the ST-MRL we also maintain additional stochastic lattice-based layers for inference and decision, which are used to plan and control the activities of the robot platforms [10]. These layers include vehicle navigation cost and risk to reach an area of interest; hypotheses of scientific events to be explored further; information metrics such as entropy to determine how the knowledge of a natural process is evolving, and where critical information is missing; and others. The augmented informational structure that incorporates both the ST-MRL and the information-theoretic inference and decision layers is called an Inference Grid (IG) [10, 12]. In the MARTA/TAOSF system, we use the Inference Grid (IG) model to represent multiple spatially- and temporally-varying properties, concerning both ocean processes and HABs. The rhodamine dye concentration measurements taken by the OASIS platforms during field tests are used as input to the Inference Grid mapping process. The fluorometer measurements are used to compute the presence or absence of dye for each cell in the area traversed by each OASIS boat. The probabilistic sensor model was derived from information on the sensitivity and performance of the sensors. The Inference Grid map of rhodamine dye for one of our mapping experiments (Fig. 5) is shown in Fig. 6.

Figure 5. OASIS-2 and OASIS-3 mapping a

patch of rhodamine water tracing dye.

6. FIELD TESTS For controlled experiments such as mapping while compensating for drift currents, geometric patches of rhodamine water-tracing dye are laid (Figs. 5 and 6). An aerostat, tethered to a manned tender boat, carries an instrument package aloft including GPS, altimeter, compass, and video camera to provide validation images of data gathered by OASIS platforms that are mapping with fluorometers. Fig. 6 shows an aerial view from the aerostat of the test area in the Chesapeake Bay; the OASIS robot boat

Figure 5: OASIS-2 and OASIS-3 mapping apatch of rhodamine water tracing dye.

platform is in the lower part of the image, close to the rhodamine dye tracks that serve as a surrogate for algal blooms during field testing. Fig. 7 shows the search pattern executed by the OASIS boat to find the surrogate algal bloom, and Fig. 8 displays an Inference Grid showing the areas with high probability of dye presence (in red), high probability of dye absence (in green), and high entropy or lack of information (in grey). Finally, in Fig. 9 an Inference Grid with inferred hypotheses of algal blooms (dye tracks) is presented. The areas of high entropy can be used to replan the search pattern of the boat.

Figure 6. Aerial view from the aerostat observation

platform. The chase boat (with aerostat tether) is in the upper part of the image, the OASIS boat in the lower part, and the rhodamine dye tracks laid by the chase boat which

serve as surrogate HABs are visible towards the right part of the image.

Figure 7. A slanted view of the ocean area where the test was conducted. A spiral search pattern was executed by the boat. The measured rhodamine concentrations are shown as a vertical “ribbon” along the route taken by the OASIS boat.

1 10 20 30 39

1

10

20

27

1 10 20 30 39

1

10

20

27

Figure 8. Inference Grid showing the areas with high

probability of dye presence (in red), high probability of dye absence (in green), and high entropy or lack of information

(in grey).

Figure 9. Inference Grid with inferred hypotheses of algal

blooms (dye tracks).

7. REFERENCES [1] A. Elfes, J. M. Dolan, G. Podnar, S. Mau, and M. Bergerman. "Safe and Efficient Robotic Space Exploration with Tele-Supervised Autonomous Robots." Proceedings of the AAAI Spring Symposium 2006, Stanford, CA, USA, March 27-29, 2006. [2] E. Halberstam, L. Navarro-Serment, R. Conescu, S. Mau, G. Podnar, A. D. Guisewite, H. B. Brown, A. Elfes, J. M. Dolan, and M. Bergerman. "A Robot Supervision Architecture for Safe and Efficient Space Exploration and Operation." Proceedings of the 10th Biennial International Conference on Engineering, Construction, and Operations in Challenging Environments, Houston, TX, USA, March 5-8, 2006.

Figure 6: Aerial view from the aerostat obser-vation platform. The chase boat (with aerostattether) is in the upper part of the image, the OA-SIS boat in the lower part, and the rhodaminedye tracks laid by the chase boat to serve as sur-rogate HABs are towards the right part of theimage.

surement uncertainty. Markov Random Fields (MRFs)provide a natural formulation to represent spatially andtemporally distributed observations. We use a discretiz-ed version of MRFs, called a spatio-temporal MarkovRandom Lattice (ST-MRL), to encode the data obtainedby the different sensors and agents. Each cell in thelattice corresponds to a spatial volume and a time slice,and stores a stochastic vector with the probabilistic statedescription of the various processes that have been mea-sured at the given location and time interval. Efficientupdating methods are used to improve the probabilisticdescription encoded in the lattice as new observationsflow in from the various sensors and platforms beingused [2].

Associated with the ST-MRL we also maintain addi-tional stochastic lattice-based layers for inference anddecision, which are used to plan and control the activ-ities of the robot platforms [3]. These layers include






1 10 20 30 39

1

10

20

27

1 10 20 30 39

1

10

20

27



(in grey).




Figure 7: A slanted view of the ocean area wherethe test was conducted. A spiral search patternwas executed by the boat. The measured rho-damine concentrations are shown as a vertical“ribbon” along the route taken by the OASISboat.vehicle navigation cost and risk to reach an area of in-terest; hypotheses of scientific events to be exploredfurther; information metrics such as entropy to deter-mine how the knowledge of a natural process is evolving,and where critical information is missing; and others.The augmented informational structure that incorpo-rates both the ST-MRL and the information-theoreticinference and decision layers is called an Inference Grid(IG) [3, 7].

In the MARTA/TAOSF system, we use the InferenceGrid (IG) model to represent multiple spatially- andtemporally-varying properties, concerning both oceanprocesses and HABs. The rhodamine dye concentra-tion measurements taken by the OASIS platforms dur-ing field tests are used as input to the Inference Gridmapping process. The fluorometer measurements areused to compute the presence or absence of dye for eachcell in the area traversed by each OASIS boat. Theprobabilistic sensor model was derived from informa-tion on the sensitivity and performance of the sensors.The Inference Grid map of rhodamine dye for one of ourmapping experiments (Figure 5) is shown in Figure 8.

Results and Analysis. Figure 6 shows an aerial viewfrom the aerostat of the test area in the ChesapeakeBay; the OASIS robot boat platform is in the lowerpart of the image, close to the rhodamine dye tracksthat serve as a surrogate for algal blooms during fieldtesting. Figure 7 shows the search pattern executedby the OASIS boat to find the surrogate algal bloom,and Figure 8 displays an Inference Grid showing theareas with high probability of dye presence (in red), highprobability of dye absence (in green), and high entropyor lack of information (in grey). Finally, in Figure 9, an






1 10 20 30 39

1

10

20

27

1 10 20 30 39

1

10

20

27



(in grey).




Figure 8: Inference Grid showing the areas withhigh probability of dye presence (in red), highprobability of dye absence (in green), and highentropy or lack of information (in grey).






1 10 20 30 39

1

10

20

27

1 10 20 30 39

1

10

20

27



(in grey).




Figure 9: Inference Grid with inferred hypothe-ses of algal blooms (dye tracks).

Inference Grid with inferred hypotheses of algal blooms(dye tracks) is presented. The areas of high entropy canbe used to replan the search pattern of the boat.

5.2 Field Test 2 - Jul. 2008In this field test, the rhodamine WT dye is laid down

in two stripes. The OASIS-2 boat follows a raster scan,which progresses along the dye stripes. The OASIS-2boat’s trajectory in global (GPS) coordinates (i.e., withadded drift) is drift-corrected to produce a correspond-ing trajectory in (water) surface coordinates (i.e., with-out drift). From Figure 10, we can observe that thelatter drift-corrected trajectory stretches over a shorterdistance (i.e., 233m) as compared to that of the formertrajectory (i.e., 380.9m). With drift correction, the dyestripes do not “run away” from the boat.

Our aim is to reconstruct the (drift-corrected) rho-damine WT dye map (i.e., in surface coordinates) usingthe sparse dye data collected along OASIS-2 boat’s tra-

0 50 100 150 200 250 300 350 4000

20

40

60

80

100

120

140

Global coordinates

Surface coordinates

t=780s

t=780s

t=1s

START

Figure 10: Comparison of OASIS-2 boat’s tra-jectories with and without added drift (respec-tively, green and black paths). The units for theaxes are in meters.jectory over a 380.9m×87.5m sampling region (i.e., 780fluorometer readings over a period of 13 minutes). Inparticular, we want to investigate if the dye-stripe phe-nomenon can be inferred from the reconstructed map.

Methodology: Log-Gaussian Process. The dyedata (see Figure 11) collected by OASIS-2 boat is char-acterized by continuous, spatially correlated, positivelyskewed (Figure 12) fluorometer measurements. Usingthis sparse dye data, we will reconstruct the dye mapin surface coordinates, which spans 233m by 102.4m.The properties of the fluorometer measurements sug-gest the use of the non-parametric probabilistic infer-ence model called the log-Gaussian process (`GP). Theadvantage of this modeling technique is that it does notrequire any assumptions on the distribution underlyingthe sampling data. Furthermore, the prediction of flu-orometer measurements at unobserved locations can beperformed at any desired map resolution. We will nowdescribe the log-Gaussian process briefly and refer theinterested reader to [9] for more details.

Let X be the domain of the dye map corresponding toa finite, discretized set of locations. Each location x ∈ Xis associated with an observed fluorometer measurementyx and its corresponding random counterpart Yx. Let{Yx}x∈X denote a `GP defined on the domain X . Thatis, if we let Zx = log Yx (i.e., Yx = exp{Zx}), then{Zx}x∈X is a Gaussian process (GP). This means thejoint distribution over any finite subset of {Zx}x∈X isGaussian. The GP can be completely specified by its(prior) mean and covariance functions

µZx

4= E[Zx] ,

σZxZu

4= cov[Zx, Zu]

for x, u ∈ X . Let the data d denote n pairs of sam-pled locations and their corresponding observed mea-surements. Let x and zx denote vectors comprising thelocation and measurement components of the data d. Ifthe data d are available, the distribution of Zx remainsa Gaussian with the posterior mean and variance

µZx|d4= E[Zx | d] = µZx

+ ΣxxΣ−1xx{z>x − µZx} (1)

0 50 100 150 200 2500

20

40

60

80

100

120

4.5

5

5.5

6

6.5

7

7.5

8

8.5

t=780 t=1

0

50

100

150

200

250 0

50

100

150

4

5

6

7

8

9

4.5

5

5.5

6

6.5

7

7.5

8

8.5

Figure 11: Continuous, spatially correlated flu-orometer readings (measured in mg/m3) of thedrift-corrected trajectory (measured in meters).

σ2Zx|d

4= var[Zx | d] = σZxZx − ΣxxΣ−1

xxΣxx (2)

where, for the location components v, w in x, µZx isa column vector with mean components µzv

, Σxx is acovariance vector with components σzxzv

, Σxx is thetranspose of Σxx, and Σxx is a covariance matrix withcomponents σzvzw . Note that the posterior mean µZx|d(1) is the best unbiased predictor for predicting thelog-measurement zx at the unobserved location x. Animportant property of the Gaussian posterior varianceσ2

Zx|d (2) is its independence of zx.The `GP has the (prior) mean and covariance function

µYx

4= E[Yx] = exp{µZx + σZxZx/2} ,

σYxYu

4= cov[Yx, Yu] = µYx

µYu(exp{σZxZu

} − 1)

for x, u ∈ X . We know that the distribution of Zx

given the data d is Gaussian. Since the transformationfrom zx to yx is invertible, the distribution of Yx giventhe data d is log-Gaussian with the posterior mean andvariance

µYx|d4= E[Yx | d] = E[exp{Zx} | d] = exp{µZx|d+σ2

Zx|d/2}(3)

σ2Yx|d

4= var[Yx | d] = µ2

Yx|d(exp{σ2Zx|d} − 1) (4)

where µZx|d and σ2Zx|d are the Gaussian posterior mean

(1) and variance (2) respectively. Note that the log-

4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 90

50

100

150

200

250

300

Measurement

Sam

ple

freq

uenc

y

Figure 12: Positively-skewed fluorometer read-ings (mg/m3): The positive skew results from asmall number of high measurements and a hugepool of low measurements.

Gaussian posterior mean µYx|d (3) is the best unbi-ased predictor for predicting the actual measurementyx = exp{zx} at the unobserved location x. On theother hand, exp{µZx|d} is a biased predictor of the ac-tual measurement yx = exp{zx}. This can be observedfrom (3) that µYx|d ≥ exp{µZx|d}.

Results and Analysis. The log-Gaussian process isused to predict the fluorometer measurements at 233×103 = 23999 unobserved locations that are evenly spacedthroughout the sampling region in surface coordinates(Figure 13). Figure 13a shows the predicted map (3)while Figure 13b shows the variance (4) associated withthe predicted locations. From Figure 13a, we can ob-serve two roughly parallel dye stripes running from theright with higher concentration to the left with lowerconcentration. The left part of the dye stripes appearsto be more diffused, thus resulting in a smaller inter-stripe gap. From the lengthscale hyperparameters ofthe log-Gaussian process (obtained using maximum like-lihood estimation), we also learn that there is a muchhigher degree of spatial correlation between fluorome-ter measurements along the horizontal axis than alongthe vertical axis. This allows the two dye stripes to beinferred relatively well.

Using the predicted dye map (Figure 13a), we furtherperform binary classification (Figure 13c) via column-wise thresholding to determine if each map location isin a dye stripe or not. From Figure 13, we can see thatthe top dye stripe is better identified than the bottomone. Figure 14 shows the row-wise profile of the pre-dicted dye map (Figure 13a); each circle denotes themean predicted fluorometer measurement for a partic-ular row (≈ 0.9945m wide). If we set the threshold tothe mean fluorometer measurement of the predicted dyemap (i.e., ≈ 5.3188mg/m3), Figure 14 then shows thatthe top and bottom dye stripes in Figure 13 are ap-proximately 15.9m and 22.9m wide, respectively, withan inter-stripe gap of roughly 9.9m.

0 50 100 150 200

10

20

30

40

50

60

70

80

90

1004.5

5

5.5

6

6.5

7

7.5

8

8.5

(a)

0 50 100 150 200

10

20

30

40

50

60

70

80

90

100

0.5

1

1.5

2

2.5

(b)

20 40 60 80 100 120 140 160 180 200 220

10

20

30

40

50

60

70

80

90

100

(c)Figure 13: Map inference of rhodamine dye con-centration: (a) Prediction, (b) variance, and (c)binary classification via thresholding. The unitsfor the axes are in meters.

0 10 20 30 40 50 60 70 80 90 100

5

5.2

5.4

5.6

5.8

6

6.2

6.4

Figure 14: Row-wise profile showing the plotof mean predicted fluorometer measurement(mg/m3) vs. row number. The red circles areabove the threshold, which is set to the meanfluorometer measurement of the predicted dyemap. The blue circles are below the threshold.

6. CONCLUSIONSThe MARTA-based TAOSF multi-level autonomy con-

trol architecture provides many advantages over exist-ing systems for observing and analyzing HABs includ-ing: dynamic tasking and adaptation; higher in situresolution and greater insensitivity to cloud cover (ascompared with satellite systems); access to and greateragility in coastal waters than that available throughbuoys; real-time multipoint science data observationsand generation of associated interpretations by remotelylocated oceanographers.

By using the TAOSF architecture, it increases data-gathering effectiveness and science return while reducingdemands on scientists for robotic asset tasking, control,and monitoring. The data are also made available toother scientists via the world-wide-web both as they arecollected, and from an historical data archive server.

AcknowledgementsThis work was supported by NASA award NNX06AF27G,“Telesupervised Adaptive Ocean Sensor Fleet”, grantedunder the Advanced Information Systems Technologyprogram of NASA’s Earth Systems Technology Office(ESTO). The TAOSF project is a collaboration amongCarnegie Mellon University (CMU), NASA GoddardSpace Flight Center (GSFC), NASA Goddard’s Wal-lops Flight Facility (WFF), Emergent Space Technolo-gies, Inc. (EST), and the Jet Propulsion Laboratory(JPL). Work on the OASIS platforms is conducted byEmergent Space Technologies, Inc., EG&G, and ZingerEnterprises under award NA03NOS4730220 from theNational Oceanic and Atmospheric Administration(NOAA), U.S. Department of Commerce.

7. REFERENCES[1] J. M. Dolan, G. W. Podnar, A. Elfes, S. Stancliff,

E. Lin, J. Higinbotham, J. C. Hosler, J. Moisan,and T. A. Moisan. Smart ocean sensing using thetelesupervised adaptive ocean sensor fleet system.In Proc. ESTC, 2008.

[2] A. Elfes. Multi-source spatial fusion usingBayesian reasoning. In M. A. Abidi and R. C.Gonzalez, editors, Data Fusion in Robotics andMachine Intelligence. Acad. Press, 1992.

[3] A. Elfes. Robot navigation: Integratingperception, environmental constraints and taskexecution within a probabilistic framework. InL. Dorst, M. van Lambalgen, and F. Voorbraak,editors, Reasoning with Uncertainty in Robotics,pages 91–130, LNCS 1093, Springer, Berlin, 1996.

[4] A. Elfes, J. M. Dolan, G. W. Podnar, S. Mau, andM. Bergerman. Safe and efficient robotic spaceexploration with tele-supervised autonomous

robots. In Proc. AAAI Spring Symposium,Technical Report SS-06-07, pages 104–113, 2006.

[5] A. Elfes, G. W. Podnar, J. M. Dolan, S. Stancliff,E. Lin, J. C. Hosler, T. J. Ames, J. Higinbotham,J. R. Moisan, T. A. Moisan, and E. A. Kulczycki.The telesupervised adaptive ocean sensor fleet(TAOSF) architecture: Coordination of multipleoceanic robot boats. In Proc. IEEE AerospaceConference, 2008.

[6] A. Elfes, G. W. Podnar, J. M. Dolan, S. Stancliff,E. Lin, J. C. Hosler, T. J. Ames, J. R. Moisan,T. A. Moisan, J. Higinbotham, and E. A.Kulczycki. The telesupervised adaptive oceansensor fleet. In Proc. SPIE Conference onAtmospheric and Environmental Remote SensingData Processing and Utilization III: Readiness forGEOSS, volume 6684, 2007.

[7] A. Elfes, G. W. Podnar, R. F. Tavares Filho, andA. Pavani Filho. Inference grids for environmentalmapping and mission planning of autonomousmobile environmental robots. In Proc. WorkshopSensing a Changing World, 2008.

[8] E. Halberstam, L. Navarro-Serment, R. Conescu,S. Mau, G. W. Podnar, A. D. Guisewite, H. B.Brown, A. Elfes, J. M. Dolan, and M. Bergerman.A robot supervision architecture for safe andefficient space exploration and operation. In Proc.10th Biennial Int’l Conf. on Engineering,Construction, and Operations in ChallengingEnvironments (Earth & Space 2006), 2006.

[9] K. H. Low, J. M. Dolan, and P. Khosla. Adaptivemulti-robot wide-area exploration and mapping.In Proc. AAMAS, pages 23–30, 2008.

[10] G. W. Podnar, J. M. Dolan, and A. Elfes.Networked architecture for robotic environmentalocean science sensors. In Proc. Workshop Sensinga Changing World, 2008.

[11] G. W. Podnar, J. M. Dolan, A. Elfes, andM. Bergerman. Multi-level autonomy robottelesupervision. In Proc. ICRA 2008 Workshop onNew Vistas and Challenges in Telerobotics, 2008.

[12] G. W. Podnar, J. M. Dolan, A. Elfes,M. Bergerman, H. B. Brown, and A. D.Guisewite. Human telesupervision of a fleet ofautonomous robots for safe and efficient spaceexploration. In Proc. HRI, pages 325–326, 2006.

[13] G. W. Podnar, J. M. Dolan, A. Elfes, S. Stancliff,E. Lin, J. C. Hosler, T. J. Ames, J. R. Moisan,T. A. Moisan, J. Higinbotham, and E. A.Kulczycki. Operation of robotic science boatsusing the telesupervised adaptive ocean sensorfleet system. In Proc. ICRA, pages 1061–1068,2008.

robot boats as a mobile aquatic sensor networkbryanlow/pubs/essa2009.pdf · robot boats as a mobile...

Documents