mapping multiple gas/odor sources in an uncontrolled indoor environment using a bayesian occupancy...

Robotics and Autonomous Systems 59 (2011) 988–1000

Contents lists available at SciVerse ScienceDirect

Robotics and Autonomous Systems

journal homepage: www.elsevier.com/locate/robot

Mapping multiple gas/odor sources in an uncontrolled indoor environment usinga Bayesian occupancy grid mapping based methodGabriele Ferri a,∗, Michael V. Jakuba b, Alessio Mondini a, Virgilio Mattoli c, Barbara Mazzolai c,Dana R. Yoerger d, Paolo Dario a

a The BioRobotics Institute, Scuola Superiore Sant’Anna, Viale Rinaldo Piaggio 34, 56025 Pontedera (Pisa), Italyb Australian Center for Field Robotics, University of Sydney, Rose Street Bldg, J04 Sydney, NSW, 2006, Australiac Center of Micro-BioRobotics@SSSA, Italian Institute of Technology, Viale Rinaldo Piaggio 34, 56025 Pontedera (Pisa), Italyd Deep Submergence Laboratory, Woods Hole Oceanographic Institution, Woods Hole, MA 02543, USA

a r t i c l e i n f o

Article history:Received 6 November 2009Received in revised form30 May 2011Accepted 6 June 2011Available online 7 July 2011

Keywords:Indoor monitoringGas source localizationGas source mappingOccupancy grid mapping

a b s t r a c t

In this paper we address the problem of autonomously localizing multiple gas/odor sources in an indoorenvironment without a strong airflow. To do this, a robot iteratively creates an occupancy grid map. Theproduced map shows the probability each discrete cell contains a source. Our approach is based on arecent adaptation (Jakuba, 2007) [16] to traditional Bayesian occupancy gridmapping for chemical sourcelocalization problems. The approach is less sensitive, in the considered scenario, to the choice of thealgorithm parameters. We present experimental results with a robot in an indoor uncontrolled corridorin the presence of different ejecting sources proving the method is able to build reliable maps quickly(5.5 minutes in a 6 m × 2.1 m area) and in real time.

© 2011 Elsevier B.V. All rights reserved.

1. Introduction

This paper addresses the issue of localizing/mapping multiplegas/odor sources in an indoor environment with no strongairflow. Gas source localization aims at solving the problem ofrapidly localizing an emitter source ejecting a chemical agent.Improvements in this localization system could open up newhorizons in robotics: we could use autonomous agents able tolocalize a gas/odor source in order to minimize the exposure ofhuman agents to dangerous pollutants. These robotic automaticsystems could find many applications in different scenarios: forenvironmental monitoring, for localizing hidden explosives ornarcotics, for automatic humanitarian demining, for rescue tasksafter accidents or natural disasters (survivors are seen as carbondioxide sources), for the detection of dangerous chemical leakagesin a domestic or industrial environment and for the preventionof carbon monoxide poisoning caused by fire or inadequateventilation of obstructed stoves.

The gas source localization task is not trivial. At medium andhigh Reynolds numbers gas dispersion is dominated by turbulentmixing [1]. Instantaneous gradients are poorly defined, time

∗ Corresponding author. Tel.: +39 050883395; fax: +39 050883399.E-mail address: [email protected] (G. Ferri).

0921-8890/$ – see front matter© 2011 Elsevier B.V. All rights reserved.doi:10.1016/j.robot.2011.06.007

varying and frequently do not point toward the source [2]. Odor isconcentrated in ‘‘packets’’, oftenwith extremely low concentrationmeasured between immersions in an odor packet.

Indoor environments present as a defining characteristic alack of strong or persistent mean flow. The patchiness andintermittency of odor distribution typical of turbulent flows inthe presence of strong mean flows is exacerbated by the lack ofa persistent mean flow. In this case, the low energy turbulentmixing produces a patchier gas distribution. Local maxima inconcentration have been observed at some distance from thesource if the source has been active for some time [3].

The difference between an environment with the presenceand the absence of a strong wind can be seen in Fig. 1 wherethe results of our source characterization experiments with onesource ejecting ethanol are showed. In the figure, the results of fivesensors located at different distances from a source are reported.

These results indicate:• without wind there are high fluctuations of gas concentration

around the average value;• with a strong and constant wind in an area near the source,

the magnitude of fluctuations around the average gas value arereduced and gradient in concentration is relativelywell defined.The energy carried by the strong wind leads to a more rapidmixing. The airflow can also be used by a robotic agent to inferthe direction to the source.

http://dx.doi.org/10.1016/j.robot.2011.06.007

http://www.elsevier.com/locate/robot

http://www.elsevier.com/locate/robot

mailto:[email protected]

http://dx.doi.org/10.1016/j.robot.2011.06.007

G. Ferri et al. / Robotics and Autonomous Systems 59 (2011) 988–1000 989

Fig. 1. Source (a circular 8.5 cm diameter dish ejecting ethanol) characterization.Sensor 0 is placed at a distance of 30 cm from the source. All sensors arespaced 25 cm, so that the farthest sensor (sensor 4) is at a distance of 130 cm.Characterization without a predominant airflow (top). The distribution appearsto be very patchy and an instantaneous concentration gradient is not present.Characterization with the presence of airflow (about 50 cm/s) generated by a fanplaced behind the source (bottom) are presented. Here a gradient in concentrationis relatively well defined.

To address the source localization issue, most researchershave proposed and implemented on robots gradient ascent basedmethods [4–6] and mixed behaviors of plume acquisition andplume following [7,8] assuming or creating by fans a persistentmean flow. A strong and constant unidirectional airflow producesa plume of a well defined shape to be followed up to its source.

A few works have addressed indoor environments with noartificially created strong ventilation, despite the fact that thiscondition is more realistic and common in indoor environmentsand represents real environments in which robots could work inthe future [9]. In this setting, some researchers have proposedmethods that either not rely or only partially rely on windmeasurements [10–13].

Multiple-source scenarios present additional problems torobotic searchers. Most reactive or behavior based plume tracingalgorithms implicitly assume that only a single source is presentin the environment. While many of these algorithms wouldlikely succeed in finding one source even in a multi-sourceenvironment, they offer no guidance on how to continue searchingfor other sources once a source has been found, or how to avoidre-finding the same source. To solve this problem we looked

at map building techniques. Even if it may not be the mostefficient way to locate a source, this approach can give importantinformation on the features of gas dispersion or can be used inparticularly challenging environments (not well defined plume,presence of multiple sources) to localize chemical sources wherethe robot might encounter difficulties in tracking a plume up toits source. Particularly in multi-source scenarios, map building cansupplant or complement reactive and behavior based strategies byproviding high-level guidance on the regions within a search areathat have not been adequately searched.

In this work we demonstrate a map building technique thatdoes not explicitly localize sources like reactive/behavior basedmethods do, but creates maps of the environment from whichareas containing sources are easily recognizable. In the approachpursued here a robot builds iteratively an occupancy grid map[14,15] that indicates grid cells likely to contain sources.

We present a method based on an extension to Bayesian occu-pancy grid mapping standard algorithm proposed by Jakuba [16]that in this application creates maps with a reduced sensitivity tochanges in algorithm parameters. In particular, it produces mapsmore consistent with the true number of sources in the case theprobabilities with which the map is initialized (priors) underesti-mate the true number of present sources. In our setting the wellknown artifacts produced by traditional occupancy grid mappingtechniques caused by independently estimating the state of eachcell [17,18] are emphasized because of the low prior with whichthe cells are initialized [16]: unlike the classical mapping problemsusing sonar sensors, only a few cells of the map are expected to beoccupied containing a source.

A particularmodel for the sensor taking into account the sensedgas intensities to use all the information carried by the chemicalmeasurements is presented.

The paper is organized as follows. Section 2 presents anintroduction to techniques for mapping a gas concentration. InSection 3 we introduce the classical occupancy grid mappingalgorithm and we present our proposed method along with theprocedure employed to select the algorithm’s parameters. InSections 4 and 5 we describe the experimental set-up and theresults from experiments with a real robot using our occupancygrid algorithm and the standard one. We conclude with adiscussion of our results and recommendations for future research.

2. Map building for chemical source localization

Some researchers have considered map based approaches tochemical plume source localization. A map of the average gasconcentrations can be produced from the measurements acquiredby a grid of fixed sensors [19,20]. Building a map with a grid ofstatic sensors offers the advantage of reduced time consumptionbutmuch effort in sensors calibration is required. Furthermore, thesystem lacks in scalability because, increasing the area, the numberof sensors needs to be incremented implying problems of cost andmaintenance. Furthermore, a large number of metal oxide sensorscan disturb the environment because the heaters present in thesensors can produce convective flows.

The use of robotic agents to produce gas distribution mapsgives more flexibility. Hayes et al. [7] proposed to build a two-dimensional histogram over the number of ‘‘odor hits’’ (sensedvalues above a certain threshold) received in the correspondingarea. The histogram bins contain the accumulated number of‘‘odor hits’’ received by all robots (six robots were used) in thecorresponding area during a random walk movement. However,the robots have to cover completely all the area of interest toproduce a reliable map and the ‘‘odor hits’’, given their binarynature, strongly depend on the chosen threshold. During themapping the robots have to scan the area and use a number of

990 G. Ferri et al. / Robotics and Autonomous Systems 59 (2011) 988–1000

sampling locations relatively low not to make the task too muchtime-expensive. In order to produce a reliable gas distributionmapof the whole area an extrapolation process is therefore needed.The proposed methods in literature have either built maps ofconcentration using different kinds of extrapolation [3,21–23]rather than explicitly of source location or require the powerfulassumption of a single source in the environment [24,25].

In [3] the sensed gas intensities are convolved with a Gaussiankernel and experimentswere carried out in an uncontrolled indoorenvironment. The method relies on the assumption that stabilizedconcentration maps represent the true average distribution,possibly with a small distortion due to the memory effect ofsensors. This assumption is difficult to prove because there isno direct way to verify the observed concentration field (the‘‘ground truth’’) for the mapping experiments. However someindications of the consistency of the produced maps comes fromthe applicability of the produced concentration mapping methodfor gas source localization. In [21] gas measurements taken by 5robots in a ventilated room are fused together with wind velocityin an advection–diffusionmodel (a non-turbulent transport modelwas assumed) using a Kalman Filter. The experimental area wasdivided into a grid and the gas concentration in each cell isrepresented by one state variable of the filter. In [22] experimentswith a mobile robot driven on a zig-zag trajectory in a ventilatedenvironment are reported. Since the measurement locations werenot equally distributed, triangle based cubic filtering was usedfor interpolation. The more jagged distribution with respectto that produced in the same environment from the averagedvalues sensed by a stationary sensor [22] shows a flaw of theinterpolation. There is no means of averaging out the fluctuationsof the gas concentration sensed at the sampling locations. Adifferent approach aiming at addressing these problems was thatproposed by Stachniss and co-workers [23] in which a sparseGaussian process mixture model is used to represent the smoothbackground signal and to model sharp boundaries around patchesof high concentrations.

Techniques to estimate the probability a source is in a certainposition have been proposed: hidden Markov methods [24] and aBayesian method [25].

Both methods, based on flow measurements and on binarydetections and non-detections of chemical effluent, update theprobability that some cells on a grid contain a source. Farrellet al. [24] propose a hidden Markov method for plume mappingand source localization to estimate the likelihood of odor detectionversus position, the likelihood of source position versus position,the most likely path taken by the odor to a given location, and thepath between two points most likely to result in odor detection.This approach makes sense in strongly advective environmentswhen the width of the plume is small relative to the search area.

In [25] a Bayesian methodology is proposed updating the stateof a random field of possible locations for a source given binarydetections/non-detections of effluent and a record of advectivecurrents. This method is explicitly developed for short time-scaleplumes suitable to be modeled as a collection of independentlydiffusing ‘‘filaments’’ [26]: each filament moves according to arandom walk due to turbulent motion with a superimposedadvection. The key assumption of this method is that a singlesource is present in the search area. This assumption makes thismethod different from occupancy grid mapping even if the gridbased representation of possible source location is similar to theidea of occupancy grid mapping. The assumption of one singlesource results in a very powerful one: this leads to an update rulethat enables cells in portions of the grid not actually observed ata given time to have their posterior updated. The detection of thechemical at one location means all potential source locations notupwind to that position are unlikely to contain a source, while

the upwind locations are likely to contain the source. Informationabout all cells of the map is extracted by a single measurement.A similar argument applies to the case where the effluent is notdetected. Even if this method works in single-source scenario itsextension to multiple sources is not straightforward.

For these reasons we considered an occupancy grid mappingbased method that enables to build maps of the probability eachdiscrete cell contains a source without considering the constraintof considering a single active source present in the search area.

3. Proposed algorithm

Occupancy grids were developed in 1980s as a finite-dimensional means of representing uncertainty in a robot’senvironment [15,14]. The method requires discretization of theenvironment into a collection of small cells, each of which is ap-proximated as being either completely occupied or else completelyempty.

Occupancy grid mapping has been applied to many roboticapplications: mapping of indoor environment [17], navigation,path planning [27,28] and collision avoidance [29]. The mapshave usually been created using long-range sensors as laser rangefinders [28], stereo vision [30] and sonar sensors [18].

Occupancy grid mapping is the process of fusing successivenoisy and uncertainmeasurements into estimates of the likelihoodof each cell being either occupied or empty. Recently, one ofus has adapted occupancy grids to the mapping of likely plumesource locations [16]. In that work, the binary state of grid cellsis redefined to denote the presence or absence of an activeplume source. A defining characteristic of chemical plume sourcelocalization considered within the context of occupancy gridmapping is the small number of cells expected to be occupiedby active plume sources. From a Bayesian perspective, the priorprobability of occupancy associated with each cell is much lowerthan is typical in occupancy gridmaps of, for instance, the layout ofan office or a building constructed from laser or sonar range-finderdata [17,18].

In this situation, thewell knownconflicts generated by standardgrid mapping algorithm [31,17] appear to be amplified [16].These conflicts are caused by the sensors cone that sweeps onmultiple cells inducing dependences among the cells that cannotbe taken into account by the standard occupancy grid mappingalgorithm [17] that estimates independently the state of eachcell. While in traditional applications these conflicts can generatemaps where, for example, open doors appear closed, in sourcelocalization they generatemapswith high posteriors respect to thetrue number of ejecting sources. These maps are, therefore, hardlyusable to localize sources.

The map update rule employed in this work [16] is shownexperimentally to produce better maps in our application than thestandard Bayesian update rule. In particular, the resultingmaps aremore consistent with the true number of occupied cells, and lesssensitive to the correct choice of prior.

We have also to underline the use of a gas sensor is quitedifferent with respect to long-range sensors usually used ingrid mapping because it provides punctual measurements. Thesepunctual measurements (above all taking into account the gasintensities) carry information about a larger area: a high gasintensity means likely a source is near. A mechanism to retrievethe information carried by gas measurements is therefore neededto be able to build maps (see Section 3.4).


3.1. Bayesian standard occupancy grid mapping algorithm

Let C denote the number of cells in an occupancy grid map,m1:C denote the collection of all binary cell states, and z1:t denotethe collection of all measurements. Then posed as a Bayesian C-dimensional binary estimation problem, all information about thestate of the map would be contained in the posterior:

p(m1:C | z1:t) (1)

Unfortunately, the number of possible maps is 2C , and so itis usually prohibitive to store or compute this full posterior.Consequently, occupancy gridmappingmethods typically seek themarginal posteriors for the occupancy of each cell,

p(ms | z1:t). (2)

However, it is not possible in general to compute these marginalsexactly without first computing the full posterior and thenmarginalizing. To avoid the computational burden implied by thistask, occupancy grid mapping methods employ map update rulesthat attempt to generate the marginal posteriors directly frommeasurements and the current state of the map.

The standard Bayesian map update rule produces inexactmarginal posteriors, but it is simple to implement and popular [18].For numerical reasons, the standard algorithm computes the so-called log odds of occupancy for each cell at each time step,

lt,i = logp(mi | z1:t)

1 − p(mi | z1:t)(3)

fromwhich it is trivial to recover the associatedmarginal posterior.At time t = 0, this quantity is initialized to the log odds prior foreach cell,

l0 = logp(mi)

1 − p(mi). (4)

Besides the prior probability of occupancy p(mi) for each cell, theonly other quantity required is an inverse sensor model to providethe probability that the cell being updated is occupied given thecurrent measurement,

p(mi | zt). (5)

Pseudo-code for standard algorithm is provided in Table 1. Acritical assumption is required in the derivation of the algorithmof Table 1 [18]:

p(zt | z1, . . . .zt−1,mi) = p(zt | mi). (6)

This equation states that the current measurement is independentof all other measurements predicated on knowledge of the cellcurrently being updated. Because sensors sweep over multiplecells, this assumption is usually incorrect and consequently themarginal posteriors computed by the standard algorithm areinexact. However, assumption (6) greatly simplifies the problemof computing the marginal posteriors by enabling the mappingproblem to be treated as a collection of C independent binaryestimation problems [17]. It remains to specify a sensor modelsuited to indoor plume source localization.

3.2. A forward sensor model for plume source localization

We conjecture a characteristic common to all plume sourcelocalization scenarios is that measurements consist fundamentallyof plume presence (detection) and absence (non-detection). Ofcourse, auxiliary measurements of sensed concentration or winddirection are often also available, but these latter quantities onlyinfluence how plume detections and non-detections should beinterpreted in terms of likely source locations. If one accepts thebinary nature of plume detection (zt = 1) and non-detection

Table 1Pseudo-code of standard Bayesian occupancy grid mapping algorithm.

Standard occupancy grid mapping({lt−1, i}, zt )for all cellsmi do

ifmi is in the perceptual field of zt thenlt,i = lt−1,i + log_inverse_sensor_model(mi, zt ) − l0

elselt,i = lt−1,i

endifendforreturn {lt,i}where log_inverse_sensor_model(mi, zt ) is the log oddsform of the inverse sensor model, log p(mi |zt )

1−p(mi |zt )

(zt = 0), then a simple model for the probability of registering adetection given complete knowledge of the locations of all sourcesin the map (i.e. the true state of each grid cell) is [16]:

p(zt = 1 | m1:C ) = 1 − (1 − P tf )

∏s∈O

(1 − P ts ) (7)

where P tc (s and c are used as a subscript) denotes the probability

that sufficient signal from cell c arrives at time t to trigger adetection, P t

f denotes the probability that a false alarm occurs attime t , and the product is over all occupied cells (O) in themap. Theprobability of non-detection is simply one minus this quantity.

To understand how (7) is derived, we start to show theprobability of non-detection zt = 0 given none or one occupiedcell,

p(zt = 0 | 0 = ∅) = 1 − P tf

p(zt = 0 | O = s) = (1 − P tf )(1 − P t

s ).(8)

The first relation of (8) states that the probability of a non-detection given no cell is occupied is equal to the probability thata false alarm did not occur. The second states that if only one cellin the entire map is occupied, the probability of a non-detectionis equal to the probability that a false alarm does not occur andthat the single occupied cell does not present a sufficient signal totrigger a detection at the location of measurement. Assuming thatall the occupied cells of the map have independent probability totrigger a detection it follows that,

p(zt = 0 | m1:C ) = (1 − P tf )

∏s∈O

(1 − P ts ). (9)

From (9) it is trivial to derive (7). This model admits thepossibility of multiple cells presenting simultaneously sufficientsignal to the detector to cause a detection, but it does not admitmultiple occupied cells reinforcing one another’s signals to triggera detectionwhen the signal present at the detector from any singleoccupied cell is insufficient on its own. The sensormodel describedby (7) is a forward model, that is, it describes the probability of ameasurement given the map. The parameters of forward models(P t

c and P tf in the above) are usually easy to specify from calibration

data. Section 3.4 contains an example of this procedure. Howeverthe standard Bayesian occupancy grid mapping algorithm requiresthe inverse marginals of (7), i.e. p(mi | zt). It can be shown that (7)is analytically both invertible and marginalizable [16]. The resultsgiven a detection and non-detection are, respectively,

p(mc | zt = 1)

=

1 − (1 − P tf )(1 − P t

c )∏s=c

(1 − P ts p(ms))

1 − (1 − P tf )

∏s(1 − P t

s p(ms))p(mc) (10a)

p(mc | zt = 0) =1 − P t

c

1 − P tcp(mc)

p(mc) (10b)

where p(mc) is the prior for cell c.


Table 2Pseudo-code of IP algorithm.

IP algorithm ({lt−1,i}, {P ti }, {zt })

Time starts at instant t = 1for all cellsmi do

ifmi is in the perceptual field of zt thenif zt = 1 (zt results in a detection)P ti =

elt−1,i

1+elt−1,i

lt,i = log1−(1−P tf )(1−P ti )

∏s=i(1−P ts P

ts )

1−(1−P tf )∏

s=i(1−P ts Pts )

+ lt−1,i

else (zt = 0, zt results in a non-detection)lt,i = log(1 − P t

i ) + lt−1,iendif

elselt,i = lt−1,i

endifendforreturn {lt,i}where l0,i = log p(mi)

1−p(mi). It is trivial from {lt,i} to

recover p(mi|z1:t ). In fact p(mi|z1:t ) =elt,i

1+elt,i.

3.3. An alternative map update algorithm

In plume source localization problems where the expectednumber of sources (occupied cells) is small, an alternative mapupdate rule often produces better results than the standardalgorithm given in Table 1. The marginal posteriors given by theinverse sensor model (10a) are exact. Therefore, if the set of allmeasurements consists of a single measurement, i.e. if Z t

= {z1},then (10a) can be used to compute the exact marginal posteriors.Now suppose these one-step marginal posteriors are regarded asindependent:

p(m1,m2, . . . ,mc | z1) =

∏i

p(mi | z1). (11)

This same requirementwasplacedon thepriorp (m1,m2, . . . ,mC |

z1), thus posteriors satisfying this property can be regarded asspecifying a new, ‘‘revised,’’ prior. Furthermore, if at t = 2 thepriors p(mc) in (10a) are replaced by the single-step marginal pos-teriors p (mc |z1), and if these are independent, then the resultantmarginal posteriors p(mc | z2) producedwill remain exact. Replac-ing the priors in (10a) with the revised priors P t

c = p(mc | zt−1) ateach timestep defines an alternative map update rule [16]. The al-gorithm pseudo-code is showed in Table 2.

Because OG algorithms intrinsically involve conditional inde-pendence assumptions, this algorithm was named after its par-ticular assumption: Independence of Posteriors (IP). In principle,independence of the posteriors could be assumed in conjunctionwith any sensor model; however, the forward model (7) is specialbecause the priors enter explicitly into the inverse sensor modelgenerated from it, thereby providing a mechanism for interpret-ing the current measurement in terms of the present belief in thestates of the cells in the map. Like other ‘‘context-sensitive’’ OGmethods specific to sonar range finders [32], the IP algorithm inter-prets new measurements in light of the present state of the map.In fact, independence needs to hold only following detections forthemarginal posteriors computed according to Table 2 to be exact;themarginal posteriors remain independent following anynumberof non-detections [16]. This asymmetry is apparent in Table 2 inthe case of non-detection: the posterior odds do not depend on thestate of any cells other than the cell in question i. Its existence canalso be understood on intuitive grounds: correct determination ofcell occupancy following a detection is strongly influenced by thestate of nearby cells because the detected plume may have origi-nated from one of the other cells rather than the cell in question.In contrast, the state of nearby cells is irrelevant following a non-detection. Since no detection has occurred, by definition no cell has

presented signal to the detector, and consequently no ambiguityarises as to which cell is responsible for generating the measure-ment (none are responsible). Uncertainty in the state of the cellsin the perceptual field still remains because the combination ofa plume and chemical sensor represents an imperfect compositesensor with probability of detection less than one.

Like the standard algorithm (showed in Table 1), the IP algo-rithm is inexact: however, the central assumption of independentposteriors required to derive it produces an algorithm that exhibitsmarkedly different behavior. By revising the prior at each iterationa weak dependence is maintained between measurements withinthe map itself. Because measurements are interpreted differentlydepending on the state of themap, the results produced by IP algo-rithm depend on the order in which measurements are collected.By contrast, the classical algorithm produces identical maps re-gardless of the order in which measurements are added. This as-pect of the IP algorithm represents a potential weakness, since badmeasurements (false alarms ormeasurements that do not conformthe assumedmodel) collected early during the survey can alter theinterpretation of further measurements.

However, for environments with low priors, it does tend toproduce maps that are more consistent with the true number ofoccupied cells and to be less sensitive to the choice of the algorithmparameters. Furthermore, its computational and storage costs areequivalent to the standard algorithm.

3.4. Algorithm implementation

It remains to specify the parameters of the sensor model,a prior probability of occupancy, p(mi), and to choose a gridcell size. In this section, we demonstrate how to specify theparameters, P t

c and P tf , of the forward sensormodel (7) from readily

attained calibration data. The invertibility of (7) avoids the oftenmore difficult task of specifying the parameters of an inversemodel directly. Empirical expressions for generating the P t

c werederived from the sensor and source characterization data shown inFig. 1.

3.4.1. Sensor calibration and source characterizationAs odor sensors we opted for TGS 2602 commercial sensors

from Figaro Engineering Inc. They are metal oxide semiconductorgas sensors and in the presence of alcohol vapors, their internalresistance changes logarithmically. The choice of this kind of sen-sors was due to their high sensitivity, their usable long life spanand their comparatively high robustness to changing environmen-tal conditions [33]. They are small, cheap and can easily be inte-grated in the measurement circuit [33]. Their main drawbacks area needed high operating temperature that results in a high powerconsumption, a slow recovery time after the gas is removed and asignificant variability in response characteristics between individ-ual sensors. For this reason, at first, some calibration sessions werecarried out. The sensors (6 sensors: 5 for source characterization,1 mounted on the robot) were tested in a hermetically sealed box,where a known and increasing amount of alcohol was periodicallyinjected. A fan was also included to accelerate alcohol vaporiza-tion and rapid mixing. The signal coming from the sensors was ac-quired using aDAQCard-6024ETM, fromNational Instruments. Datawere sampled every 1 ms; we only stored the mean value com-puted on 500 measurements. In order to calibrate our sensors, thecollected data were fitted with the characteristic bi-logarithmicfunction given by the sensors datasheet. Best fit parameters wereextrapolated. A third degree polynomial was used to fit the data.The found coefficients of the polynomial were used to convert sen-sor voltage output in concentrations [10].

Once the sensors were calibrated, some preliminary experi-ments were carried out with environmental conditions similar to


00 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0

0 0.5 1 1.5 2 2.5 3

1

2

3

4

5

6

7

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Sin

gle

-so

urc

e p

rob

abili

ty o

f m

easu

rin

g c

on

c >=

ref

eren

ce

Reference concentration (µl/l)

Distance from the source

0.3 m

0.55 m

0.8 m

1.05 m

1.3 m

Single source probability of measuring concentration greater than or equal to reference concentration vs. reference concentration

(in dashed line the fitting Gaussians)

Sp

read

(µl

/l)

Range (m)

Experimental dataInterpolating exponential

Spread vs. range assuming P[conc >= ref] = G(ref;sigma(range))

Fig. 2. (Left) Single-source probability of measuring a concentration greater than or equal to reference concentration vs. reference concentration. The probabilities for thedifferent sensors are shown. (Right) Standard deviation of the fitted Gaussian for the different sensors and the interpolating exponential.

00 0.5 1 1.5 2 2.5 3

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5

con

c (µ

l/l)

Probability that measured concentration is above reference concentrationas a function of reference concentration and range from source

Range (m)

Fig. 3. Probability that measured concentration is above reference concentration as a function of the reference concentration and range from source.

those present in the survey area in order to characterize the gasdispersion. We used an 8.5 cm diameter circular dish containingethanol free to evaporate as a source for the calibration. The di-mension of this source was similar to those used during our ex-periments with robot (see Section 4). We used 5 sensors placed ina row, spaced 25 cm. The first one is at a distance of 30 cm fromthe source. Data acquisitions started after a source had been freeto eject for 10 min. Two different kinds of trials were carried out:one without the introduction of an external artificial airflow andthe other with a fan, placed 40 cm behind the source, generatingan airflow of about 50 cm/s. In the trials where wind was present,the sensors were placed inside the generated plume. The resultsare shown in Fig. 1.

3.4.2. Determining the parameters of the sensor modelThe percentage of measurements above a given reference

concentration is shown in Fig. 2 for each gas sensor in thearray. Based on these data, we fit analytical functions to theexperimentally observed probability that measured concentrationq exceeds a given reference concentration p(q ≥ qref ; r, qref ) as a

function of range r from the source and the reference concentrationqref . First we fit a Gaussian curve to the data from each sensorto yield the probability of the measured concentration exceedinga reference concentration at the range ri from the source (ri =

0.3, 0.55, 0.8, 1.05, 1.3 m) associated with each sensor and as afunction of reference concentration:

p(q ≥ qref ; ri, qref ) = e−12 (

qrefσ(ri)

)2 (12)

where σ (ri) denotes a range-dependent scale parameter deter-mined by the fit. Nextwe fit an exponential to σ (ri) to permit com-putation of p(q ≥ qref ; r, qref ) for all ranges:

σ(r) = e−5.849·(r−0.325)+ 0.1323. (13)

Fig. 2 shows the σ(ri) and the interpolating exponential function.Fig. 3 shows the final analytical approximation to p(q≥ref ; r, qref ).

Let rc denote the range between the robot and a cell c . Then theprobability p(q ≥ qref ; rc, qref ) determined thus far correspondsto the P t

c we seek if the reference concentration qref is set tothe measured concentration q as long as q exceeds some staticmaster threshold q0 discriminating a detection (zt = 1) and non-


Fig. 4. One robot of MOMO platform.

detection (zt = 0). For detections, the equivalence between p(q ≥

qref ; rc, qref ) and P tc holds for a dynamically altered reference

concentration (qref = q); however, non-detections require settingqref = q0 in order to compute the correct P t

c . In fact P tc denotes

the probability that sufficient signal from cell c arrives at time tto trigger a detection; since a detection would have been declaredonly for q ≥ q0, in a non-detectionwe fix qref = q0. Thus after eachmeasurement the P t

c are computed according to:

ptc =

p(q ≥ qref ; rc, qref ) q ≥ q0 (detection)p(q ≥ q0; rc, q0) q < q (non-detection). (14)

In this way, the measured concentration can be used to help inferrange to source without violating the imposed form of the forwardsensor model (7). In multiple-source environments, effluent fromtwo or more sources can mix to produce, on average, highermeasured concentrations than would be present at the samerange from a single source. The forward sensor model prohibitsinferring the number of sources contributing to the currentmeasurement, and therefore high measured concentrations willcause themapping algorithms employing thismodel to buildmapsthat favor the explanation of a single nearby source rather thanmultiple sources farther away.

On the basis of Fig. 3, the probability of measuring aconcentration above about 0.7 µl/l(1 µl/l of ethanol is equalto 383.6249 ppm or 0.789 g/m3) is essentially independent ofrange for ranges greater than about 0.5 m from the source. Thissuggests a natural static master threshold for detection q0 of about0.7 µl/l. With a threshold near this value, the P t

c associated withcells at least 1.5 m from the robot are less than about 0.02. Onthe basis of this fact, we chose to use a static perceptual radiusof 1.5 m, outside of which all P t

c are effectively regarded as zero.Computing the P t

c at each iteration for all cells in the small mapused in our experiments (∼1250 cells) presents only a modestcomputational load; however, by employing a static perceptualradius, the algorithms presented here scale without modificationto map sizes limited only by available memory. Our experimentsall employed a probability of false alarm, P t

f = 0. In our experience,the probability of measuring a concentration above about 0.7 µl/lin the absence of a nearby source is negligible. Finally, we chose abasic cell size for our experiments of 10 cm × 10 cm, on the orderof the size of the sources used. We also present results from mapswith 30 cm × 30 cm grid cells.

4. Experimental set-up

The experiments were performed using one robot of MOMOplatform [10]. The PC tracks robot position (via webcams),communicates with the robot, checks its state (through an FMcommunication channel – an AUR EL XTR-434H from AUREL spa

Fig. 5. Experimental area in which experiments were carried out.

0600

500

400

300

200 200

100 100

0

1

2

3

4

5co

nce

ntr

atio

n (

µl/l)

Concentrations sensed by the robot during a survey

X (cm)

Y (cm)

Fig. 6. Trajectory of the robot with gas samplings and source locations (blackcrosses).

Inc. transceiver is used eitherwith the PC or the robots), collects thedata produced by the robot actions and allows a human operatorto tele-operate the robotic agent. One single robot (see Fig. 4)(17 cm long and 17 cm wide) has two encoders on its wheels forrough odometry, amainmicro-controller thatmanagesmovement(PIC 18F452 from Microchip Technology Inc.), two sets of highcapacity rechargeable Ni–MHbatteries (assuring at least two hoursof durability), two bumper whiskers as touch sensors, and a gassensor.

In the experiments described in this paper, the robot was tele-operated using the FM wireless connection by a human operator.The positions were acquired using two Logitech Quickcam ExpressII webcams (640 × 480 resolution) fixed to the ceiling. The robotwas equipped with one TGS 2602 sensor from Figaro EngineeringInc. for alcohol detection.

The experiments were performed in an area of one corridorof CRIM Lab building. The dimensions of the area were 6 m ×

2.1 m (see Figs. 5 and 6). The data processing was akin to thatapplied to our calibration data (see Section 3.4.1). Once a samplingwas commanded, the robot stopped and computed the averageon a temporal window of 3 s. The robot followed a trajectory asthat showed in Fig. 6. The samplings are spaced 40 cm: one trialconsisted in 64 samplings.


Fig. 7. Posteriors in log odds form (top) and posteriors (bottom) for one trial using IP and standard algorithm. The priors are equal to 4 × 10−3 (consistent with 5 expectedsources). The maps are created from ‘‘experiment 1’’ dataset. The standard algorithm presents slight overestimated posteriors with respect to IP and the peaks related to thethree different sources are less uniform as value than in maps produced with IP (Note the different gray scales used).

Several trials were carried out in the presence of three sourcesejecting ethanol. The sources used in the trials (labeled S1, S2 andS3) are different in shape and so eject alcohol at different rates.S1 was a circular 8.5 cm diameter dish, S2 was formed by twocircular 8.5 cm diameter dishes and S3 was composed by a singlecircular 15 cm diameter dish. In all the trials the odor sources werekept at the same positions. The reported experiments were carriedout in 5 experimental sessions carried out in different days duringthe month of February (second half of the month). In each sessionthe first trial started after the sources had been ejecting ethanolfor 10 min. The next trial started about 10 min after the end ofthe previous one. The temperature was between 18 and 20 C° andthere were no drastic changes of environmental conditions duringthe experiments. No artificial airflow was introduced. The heatingsystem of the building was not turned off in order to create arealistic indoor scenario.

5. Experimental results

In this section we report the results of 15 experiments with therobot. Experimental results show that the proposed method canprovide suitable maps for localizing multiple gas ejecting sources.A goodmap provides an indication both where sources are and arenot likely to be, thus indicating which areas have been exploredand which still need to be further investigated.

Figs. 7 and 8 show the posteriors (also in log odds form) ofmaps generated using the samplings acquired in two differentexperiments (they are named in the following ‘‘experiment 1’’and ‘‘experiment 2’’). The parameters used to generate both themaps are those presented in Section 3.4. The priors for the cellswere chosen to be consistent with 5 expected sources (0.004).The grid cell size for these maps was selected as 10 cm × 10 cm.Eqs. (10a) and (10b) were used in standard occupancy gridmapping algorithm as inverse sensor model (Fig. 9).

Multiple sources of different sizes are present, diffusing ethanolin a turbulent environment. Moreover, the plumes generated bythe different sources can coalesce. Nevertheless, our algorithmsucceeds in building maps that are easily interpretable in termsof location and number of likely sources. There are clearly threeseparated candidate areas, one for each source. The areas are nearthe sources and are a bit to the ‘‘north-east’’ with respect to thetrue locations. This suggests the presence of a weak wind (furtherevidence appears in the raw data of Fig. 6, where the samplingsof ‘‘experiment 1’’ are reported). This little gap between declaredpositions and true ones derives from the fact we did not use thewind information in our model given the difficulty in having areliable measurement.

One merit factor that we consider in this work is the expectednumber of sources given themeasurements, E(#sources | z1:t). It is


Fig. 8. Posteriors in log odds form (top) and posteriors (bottom) for one trial using IP and standard algorithm. The priors are equal to 410−3 (consistent with 5 expectedsources). The maps are created from ‘‘experiment 2’’ dataset. The standard algorithm presents slight overestimated posteriors with respect to IP and the peaks related to thethree different sources are less uniform as value than in maps produced with IP (Note the different gray scales used).

computed as the sum over all the cells of themap of the posteriors,that is,

E(#sources | z1:t) =

−i

p(mi | z1:t). (15)

By comparing the posterior expected number of sources to theactual number of sources we can gauge the accuracy of theposteriors in the entire map. Accurate posteriors are necessary ifdecisions would be made on the basis of the map.

Figs. 7 and 8 compare the maps produced by the IP and thestandard algorithm applied to two sets of experimental data:while these maps appear to be qualitatively similar there aredifferences. The first difference is in the estimated number ofexpected sources given themeasurements: IP gives a result of 5.84and5.44 and it is an essentially correct number, standard algorithmproduces 13.62 and 13.78 (overestimated numbers). Secondly,the IP produces a map with peaks of increased probability moreuniformly distributed if compared to those produced by thestandard algorithm. For example, in Fig. 8, we can notice thatfor standard algorithm the spot related to S3 contains cells withprobabilities higher if compared to the spots related to S1 and S2.

Even if the grid cell size of 10 cm × 10 cm gives good resultsproducing smooth maps, an interesting point to investigate iswhat happens to the produced maps if the size of the grid cells

is incremented. This might be necessary when a large indoorenvironment has to be mapped and the robot is not providedwith enough memory resources. The data of the ‘‘experiment 1’’were used to create a map with the grid cell size of 30 cm ×

30 cm (see Fig. 9). The used prior is consistent with an expectationof 5 sources in the experimental area. Even if the producedmaps are less smooth they are useful to localize the sources. TheE(#sources | z1:t) in this case are 5.97 for the IP and 7.53 for thestandard. So far, it appears from the showed maps that standardalgorithm overestimates the number of expected sources slightlywith respect to the IP algorithm. Thus far we have utilized thecorrect prior: for ‘‘correct’’ we mean that the prior is compatiblewith the true number of sources present in the search area.However, the problem with the standard algorithm is its strongdependence on priors and thresholds.

This is evident from Fig. 10 where the posteriors produced bythe algorithmswith priors equal to 4×10−5 are reported (this is thecase where the prior underestimates the true number of sources).The data from which the maps are created are once again those of‘‘experiment 1’’. It is evident that while the IP algorithm continuesto produce a good map, the standard one totally overestimates theposteriors. This is evident in the number of expected sources giventhe measurements. While for IP algorithm E(#sources | z1:t) = 7and this is still a good estimate, the standard one results in 195estimated sources. It is clearly a completely unrealistic number.


Fig. 9. Posteriors in log odds form (top) and posteriors (bottom) created from ‘‘experiment 1’’ dataset using IP and standard algorithm. The priors are equal consistentwith 5 expected sources. The cell size is 30 cm × 30 cm. The maps appear less smooth, but always usable to localize the sources. The standard algorithm presents slightlyoverestimated posteriors.

When viewed in log odds form the peaks produced by the standardalgorithm are still in the right position; however it is unclear howto base decisions on these values.

The dependence of the two algorithms on the priors and threedifferent thresholds is shown in Fig. 11where the expectednumberof sources given the posteriors versus different used priors isreported. On the graph, the results of the two algorithms withthe thresholds of 0.6, 0.8 and 1µl/l are reported. The averagesof the results of the datasets from 15 experiments are reported.While IP algorithm proves very robust to the priors decreasing,the standard one starts to overestimate the number of sourcesas priors decrease. The two algorithms tend to have similarbehaviors only when the priors grossly overestimate the truenumber of sources present in the searching area. When the priorsgrossly overestimate the true number of sources neither algorithmproduces posteriors consistent with the true number of sources.

One delicate issue for map creation in either algorithm is thedecision of the threshold to discriminate if a measurement is adetection or not: a too low threshold could consider too manymeasurements to be detections while they could be due to thesaturation of the environment not bringing useful informationabout source position, while a too high one might generate mapsthat miss some sources. The threshold depends clearly on thequantity of odor present in the air and that is function of different

factors: wind, source dimensions/ejecting rates and time sourceswere free to eject before the robot starts its survey. However, tochoose the threshold, firstly, we relied on the fact that sourceshad being ejecting for at least 10 min before the survey startedgiving a certain time to the gas to diffuse; secondly, on the fact thateven if the odor sources are different in shapes, they are all of thesame order of magnitude. The robustness of the chosen thresholdhas been investigated using also 0.6 and 1 µl/l as thresholds, and,once again while for IP the maps do not change significantly, forstandard algorithm the posteriors of the produced maps are moreinfluenced by the used threshold.

The comparison between IP and standard shows that IPmethodis much less sensitive than standard algorithm to both priors andthresholds changing.

However, when the priors grossly overestimated the truenumber of sources the two algorithms are not able to build amap consistent with the true number of sources present in thesearching area.

The maps created by IP algorithm appear to be suitable tomap different sources ejecting a chemical. The probable positionsproduced by the maps, as told, are a bit to the ‘‘north-east’’ of thesource due to a light airflow blowing toward that direction. Toquantify the errors between true source position and estimatedone, we report in Fig. 12 the averages on the experimental trials of


Fig. 10. Posteriors in log odds form (top) and posteriors (bottom) from ‘‘experiment 1’’ dataset using IP and standard algorithm. The priors are equal to 4 × 10−5 (priorunderestimates the true number of sources). While IP continues to produce maps coherent with the true number of sources, the standard algorithm overestimates theposterior (Note the different gray scales used).

100

10-5 10-4 10-3 10-2 10-1 100

101

102

103

Log (priors)

Lo

g (

E[#

sou

rces

z1–

t ])

Expected number of sources given the measurements vs. different priors and different thresholds(plotted data are the averages of the results of using standard

and IP algorithms on 15 different experimental datasets)

IP (th = 0.6)IP (th = 0.8)IP (th = 1)atd (th = 0.6)atd (th = 0.8)atd (th = 1)

Fig. 11. Expected number of sources given the posteriors versus different used priors and different used detection threshold. The values are the averages of the results ofapplying IP and standard algorithm on 15 different experimental datasets. IP results are much less sensitive to both thresholds and priors variations, though both algorithmsare incapable of producing a map consistent with the true number of sources when the priors grossly overestimate the true number of sources.

the distances between real and estimated position of the sources.The probable source position for the different sources was selectedas the center of mass of the probabilities of the three spots onthe maps that present an increased probability to contain an

occupied cell. The average of the errors was in all three caseslower than 50 cm and the S3 is very well localized (average errorof about 18 cm). The relatively low standard deviations suggestthat map creation with our method is quite insensitive to the gas


10

0

20

30

40

50

60

70

Dis

tan

ce (

cm)

S1 S2 S3Sources

Errors between the real and estimated positions of the sources(average and standard deviation computed on 15 trials)

45.0733.63

18.52

Fig. 12. Errors between estimated and real source position. The reported values arethe averages and the standard deviations of the 15 experimental performed trials.

distributions sensed in different experiments (this is also deduciblefrom Figs. 7 and 8 where the maps are very similar). The methodfilters the information contained in the different gas acquisitionsto estimate the most likely sources’ positions.

It is interesting to remark that the proposed method can buildreliable maps in a relatively short time and in an online manner.In our experiments an area of 6 m × 2.1 m was mapped using64 measurements (each gas acquisition lasted 3 s). The final timefor one trial is about 5.5 min. With non-detections, moreover, theexplored areas that are not likely to contain a source are explicitlyshowed. This allows a hypothetical human operator not only toknow in real time where the sources likely are, but also to knowareas that are candidates to be empty, so to make decisions whilethe robot is mapping the area without waiting for the end of themapping process.

6. Conclusions and future works

In this paper, we address the problem to localize, with amobile robot, multiple sources ejecting a chemical in an indoorenvironment with no strong airflow. These conditions represent arealistic domestic or industrial scenario. Without a strong airflowthe task to localize a source gets harder because no informationabout the airflow can be used to infer the source direction, andbecause the low energy turbulent mixing produces a patchy andintermittent gas concentration. In our setting the presence ofmultiple sources ejecting a chemical is considered. The presenceof multiple sources increases the difficulties of the problem:reactive/behavior based algorithms offer no clear extension toenvironments with multiple sources. To overcome these problemswehave described amapbuilding basedmethod aiming at creatinga map of the environment: the survey area is divided into cellsand each cell presents the probability to contain a source. Evenif our method does not explicitly localize the source (the robotmoves on a predefined path), it can give useful information to ahuman operator: areas where sources are likely to be, areas thatare probably empty and areas not explored are indicated.

To create this kind of map we use occupancy grid mappingmethods. We show that in our particular setting (low priorwith which the cells are initialized) the well known artifacts oftraditional occupancy grid mapping techniques [18] caused byindependently estimating the state of each cell, are amplified [16]producingmapswith posteriors overestimating the true number ofsources and hardly usable for our needs. To solve this, we adopt analgorithm (IP) that is an extension [16] to the traditional Bayesiangrid algorithm. This IP algorithm is essentially based on a ‘‘revisedprior’’ in order to build a linkage between previous measurementsthat resulted in detections of the gas and the more recent ones.A particular sensor model is assumed taking into account theintensities of the measured gas. Our method has been tested in a

corridor of CRIM Lab and has been showed to create reliable mapsquite insensitive to the specific dataset used.

Our method presents the following characteristics:• it does not need to assume the presence of only one source to

work (multiple sources were present in the corridor);• it does not rely on information about the airflow to work;• it can provide suitable maps after a short time (a 6 m × 2.1 m

area mapped in about 5 min and a half);• it can work in an online fashion.

All these features make it suitable to be used in real indoormonitoring task.

In future works we will study possible solutions in combiningthe use of this map building method with reactive algorithms. Themaps should be useful to tell the robot where to explore for newsources and avoid to refind the previously localized sources.

References

[1] B.I. Shraiman, B.I. Siggia, Scalar turbulence, Nature 405 (6787) (2000)639–646.

[2] J. Murlis, J.S. Elkinton, R.T. Cardè, Odor plumes and how insects use them,Annual Review of Entomol. 37 (1992) 505–532.

[3] A. Lilienthal, T. Ducket, Building gas concentration gridmaps with a mobilerobot, Robotics and Autonomous Systems 48 (2004) 3–16.

[4] S. Kazadi, R. Goodman, D. Tsikata, D. Green, H. Lin, An autonomouswater vaporplume tracking robot using passive resistive polymer sensors, AutonomousRobots 9 (2000) 175–188.

[5] C. Lytridis, E.E. Kadar, G.S. Virk, A systematic approach to the problem of odoursource localization, Autonomous Robots 20 (2006) 261–276.

[6] R.A. Russell, A. Bab-Hadiashar, R.L. Scheffer, G. Wallace, A comparison ofreactive chemotaxis algorithms, Robotics and Autonomous Systems 45 (2003)83–97.

[7] A.T. Hayes, A. Martinoli, R.M. Goodman, Distributed odor source localization,IEEE Sensors Journal 2 (3) (2002) 260–273.

[8] H. Ishida, G. Nakayama, T. Nakamoto, T. Moriizumi, Controlling a gas/odorplume-tracking robot on transient responses of gas sensors, IEEE SensorsJournal 5 (3) (2005) 537–545.

[9] H. Ishida, Robotic systems for gas/odor source localization: gap betweenexperiments and real-life situations, in: Proceedings of Workshop on RoboticOlfaction of IEEE International Conference on Robotics and Automation ICRA2007, Rome, Italy, 2007.

[10] G. Ferri, E. Caselli, V. Mattoli, A. Mondini, B. Mazzolai, P. Dario, SPIRAL: a novelbiologically-inspired algorithm for gas/odor source localization in an indoorenvironment with no strong airflow, Robotics and Autonomous Systems 57(4) (2009) 393–402. 30.

[11] H. Ishida, H. Tanaka, H. Taniguchi, T. Moriizumi, Mobile robot navigation usingvision and olfaction to search for a gas/odor source, Autonomous Robots 20(2006) 231–238.

[12] A. Lilienthal, D. Reiman, A. Zell, Gas source tracing with a mobile robotusing an adapted moth strategy, Autonome Mobile Systeme (AMS) Karlsruhe(December 4–5) (2003) 150–160.

[13] A. Lilienthal, T. Duckett, Experimental analysis of smelling Braitenbergvehicles, in: Proceedings of the IEEE International Conference on AdvancedRobotics (ICAR 2003), Coimbra, Portugal, 2003, pp. 375–380.

[14] H.P. Moravec, Sensor fusion in certainty grids for mobile robots, AI Magazine9 (1988) 61–74.

[15] A. Elfes, Occupancy grid: a probabilistic framework for robot perception andnavigation, Ph.D. Thesis, Department of Electrical and Computer Engineering,Carnegie Mellon University, 1989.

[16] M.V. Jakuba, Stochastic mapping for chemical plume source localization withapplication to autonomous hydrothermal vent discovery, Ph.D. Thesis, WHOI-MIT, 2007.

[17] S. Thrun, Learning occupancy grid maps with forward sensor models,Autonomous Robots 15 (2003) 111–127.

[18] S. Thrun, W. Burgard, D. Fox, Probabilistic Robotics, MIT Press, Cambridge,Massachusetts, 2005.

[19] H. Ishida, T. Nakamoto, T. Moriizumi, Remote sensing of gas/odor sourcelocation and concentration distribution using mobile systems, Sensors andActuators B 49 (1998) 52–57.

[20] A.H. Purnamadjaja, R.A. Russell, Congregation behaviour in a robot swarmusing pheromone communication, in: Proceedings of the AustralasianConference on Robotics and Automation, Sydney, Australia, 2005, pp. 1–7.

[21] L. Marques, A. Martins, A. de Almeida, Environmental monitoring withmobile robots, in: Proceedings of the 2005 IEEE/RSJ International Conferenceon Intelligent Robots and Systems (IROS 2005), Edmonton, Canada, 2005,pp. 3624–3629.

[22] P. Pyk, S. Bermudez i Badia, U. Bernardet, P. Knusel, M. Carlsson, J. Gu,E. Chanie, B.S. Hansson, T.C. Pearce, P.F.M.J. Verschure, An artificialmoth: chemical source localization using a robot based neuronal modelof moth optomotor anemotactic search, Autonomous Robots 20 (2006)197–213.

[23] C. Stachniss, C. Plagemann, A. Lilienthal, Learning gas distribution modelsusing sparse Gaussian process mixtures, Autonomous Robots 26 (2–3) (2009)187–202.


[24] J.A. Farrell, S. Pang, W. Li, Plume mapping using hidden Markov methods,IEEE Transactions on Systems Man and Cybernetics Part B: Cybernetics 33 (6)(2003) 850–863.

[25] S. Pang, J.A. Farrell, Chemical plume source localization, IEEE Transactionson Systems Man and Cybernetics Part B: Cybernetics 36 (5) (2006)1068–1080.

[26] J.A. Farrell, J. Murlis, X. Long, W. Li, R.T. Cardé, Filament-based atmosphericdispersion model to achieve short time-scale structure of odor plumes,Environmental Fluid Mechanics 12 (2002) 143–169.

[27] D. Kortenkamp, R.P. Bonasso, R.Murphy, AI-basedMobile Robots: Case Studiesof Successful Robot Systems, MIT Press, Cambridge, Massachusetts, 1998.

[28] S. Thrun, Learning metric-topological maps for indoor mobile robot naviga-tion, Artificial Intelligence 99 (1) (1998) 21–71.

[29] J. Borenstein, Y. Koren, The vector field histogram – fast obstacle avoidancefor mobile robots, IEEE Journal of Robotics and Automation 7 (3) (1991)278–288.

[30] J. Buhmann, W. Burgard, A.B. Cremers, D. Fox, T. Hofmann, F. Schneider, J.Strikos, S. Thrun, The mobile robot Rhino, AI Magazine 16 (1) (1995).

[31] K. Konolige, Improved occupancy grids for map building, Autonomous Robots4 (1997) 351–367.

[32] H.P. Moravec, D.W. Cho, A bayesian method for certainty grids, in Workingnotes of AAAI 1998 Spring Symposium on Robot Navigation, Stanford, CA,1989, pp. 57–60.

[33] A. Lilienthal, A. Loutfi, T. Duckett, Airborne chemical sensing with mobilerobots, Sensors 6 (2006) 1616–1678.

Gabriele Ferri received the Laurea degree (M.S.) inComputer Engineering (with Honors) from the Universityof Pisa, Italy, in 2003. From 2003 to 2005, he workedas a Software/System Consultant Engineer for WASScompany (a Finmeccanica group company) in Livorno,Italy, working on the design of a new System of Controland Guidance for a light-weighted torpedo. He earned aPh.D. in Biorobotic Science and Engineering jointly fromIMT Advanced Studies in Lucca, Italy and Scuola SuperioreS.Anna, Pisa, Italy in 2008. He was a visiting researcherfor one year at Woods Hole Oceanographic Institution,

Woods Hole, Massachusetts, USA. Currently, he is a Postdoctoral Investigator at theBioRobotics Institute,MarineRobotics Lab, Scuola Superiore Sant’Anna.His researchinterests include robotic systems for environmental monitoring/exploration, robotnavigation and localization.

Michael V. Jakuba is presently a postdoctoral fellow atthe Australian Centre for Field Robotics at the Universityof Sydney. His research interests revolve around thedevelopment of underwater robots to facilitate oceanscience, particularly through high resolution and longterm autonomous benthic mapping. Dr. Jakuba earned hisdoctorate from the MIT/WHOI Joint Program in AppliedOcean Physics and Engineering in 2007. His thesis workdeveloped novel mapping algorithms for the autonomouslocalization of hydrothermal venting.

Alessio Mondini received his Laurea Degree in electronicengineering from the University of Pisa, Italy in 2002. In2003, he worked in the national energy company (ENEL)with a stage program focused on data acquisition systemand sensor conditioning for physical measures into thegeothermal wells. From May 2004, he is working at CRIMLaboratory of Scuola Superiore Sant’Anna. He receiveda Ph.D. in Microsystem Engineering from University ofRoma Tor Vergata in 2009. His main research interestsinclude microsystems technologies, microelectronics andminiaturized data acquisition system for sensor networks.

Virgilio Mattoli received his Laurea Degree in chemistry(with Honors) from the University of Pisa and the Diplomain Chemistry (with Honors) from the Scuola NormaleSuperiore of Pisa in 2000. He joined the CRIM Lab of ScuolaSuperiore Sant’Anna in 2001 with a research programfocused on the control and integration of miniaturizeddevices for environmental and biomedical application.He received his Ph.D. in bio-engineering (with Honors)in 2005. He has been a visiting researcher at StanfordUniversity, California, USA, and at Waseda University,Tokyo, Japan. From May 2008, he obtained a temporary

position of assistant professor in bio-engineering at SSSA. Since November 2009,he has been a Team Leader of the Smart Materials Platform in the Center forMicro-BioRobotics IIT@SSSA. His main research interests include: thin film sensor

microfabrication, sensor conditioning, miniaturized acquisition system, wirelesssensor networks, mechatronics and biorobotics. He is currently involved in severalresearch projects on these topics.

Barbara Mazzolai graduated in Biology (with Honors)at the University of Pisa, Italy, in 1995. From 1994 to1998, she worked at the Institute of Biophysics of theNational Research Council on environmental topics. In1998, she received her Master degree in Eco-managementat Scuola Superiore Sant’Anna (SSSA), Pisa. She had aresearch assistant position at CRIM Lab of SSSA from 1999to 2004, where she was the founder of a research areaon ‘‘Microtechnologies and robotics for environmental andagrofood applications’’. From July 2004 to October 2009,she had a position of Assistant Professor in Biomedical

Engineering at SSSA. Shewas co-founder of the ResearchCentre on Sea Technologiesand Marine Robotics of the Scuola Superiore Sant’Anna in Livorno. From November1, 2009 to February 15, 2011, she was Team Leader at the Center for Micro-BioRobotics IIT@SSSA of the Istituto Italiano di Tecnologia (IIT) of Genoa, Italy.She is currently the Coordinator of the IIT@SSSA Center. Her research interestsare in the field of biorobotics and micro-biorobotics, and service robotics. She isproject manager of many European projects in the field of biorobotics and servicerobotics. She is author of more than 90 appeared in international journals, books,and conference proceedings. She is Guest Co-Editor of a Special Issue of two SpecialIssues of Applied Bionics and Biomechanics on ’’Biologically inspired robots andmechanisms’’. She has been amember of panels of the EuropeanCommissionwithinthe Seventh Frame Program in the field of Robotics and ICT, and of the PortugueseFoundation for Science and Technology in the field of Environment and Health. Sheis a member of the IEEE, of the Engineering in Medicine and Biology Society, and ofthe Robotics & Automation Society.

Dana R. Yoerger received the SB, SM, and Ph.D. Degrees inMechanical Engineering from the Massachusetts Instituteof Technology. He is currently a Senior Scientist at theWoods Hole Oceanographic Institution, where he designsand implements robotic control systems for teleroboticand autonomous underwater vehicles. He supervises thethesis research of students enrolled in the WHOI/MITjoint program in Oceanographic Engineering, in the areasof control, robotics, and design. Dr. Yoerger has gone tosea on over 50 oceanographic expeditions, including the1985 Titanic discovery cruise. He was a member of the

team that developed the telerobotic control system for the Jason ROV and hasserved as Jason navigator on many cruises. For operations of the fully autonomousABE vehicle, his responsibilities include science liaison, control system designand implementation, mission programming and testing, and generation of dataproducts. He has also participated in a number of education and outreach programsfeaturing deep submergence technology and seafloor exploration. These includefive Jason Project expeditions and two REVEL programs. Dr. Yoerger is a memberof WHOI’s Deep Submergence Laboratory (DSL) which was founded by Dr. RobertBallard. DSL features a unique team of engineers, technicians, administrators,and students who perform engineering research and development, and alsoconduct extensive oceanographic field work under the auspices ofWHOI’s NationalDeep Submergence Facility. He also works with engineers from WHOI’s AppliedEngineering Laboratory and the Alvin Group.

Paolo Dario received the Dr. Eng. Degree in mechanicalengineering from the University of Pisa, Pisa, Italy, in 1977.He is currently a Professor of Biomedical Robotics at theScuola Superiore Sant’Anna, Pisa, Italy. He also teachescourses at the School of Engineering of the Universityof Pisa, and at the Campus Biomedico University, Rome,Italy. He has been a Visiting Professor at Brown University,Providence, RI, at the Ecole Polytechnique Federale deLausanne (EPFL), Lausanne, Switzerland, and at WasedaUniversity, Tokyo, Japan. He was the founder of theAdvanced Robotics Technologies and Systems (ARTS)

Laboratory and is currently the coordinator of the Center for Research inMicroengineering (CRIM) Laboratory of the Scuola Superiore Sant’Anna, where hesupervises a team of about 70 researchers and Ph.D. students. He is also the Directorof the Polo Sant’AnnaValdera and aVice-Director of the Scuola Superiore Sant’Anna.His main research interests are in the fields of medical robotics, mechatronics,and micro/nanoengineering, and specifically in sensors and actuators for the aboveapplications. He is the coordinator of many national and European projects, theeditor of two books on the subject of robotics, and the author of more than 200scientific papers (75 in ISI journals). He is Editor-in-Chief, Associate Editor, andMember of the Editorial Board of many international journals. Prof. Dario servedas President of the IEEE Robotics and Automation Society during 2002–2003, andhe is currently Co-Chair of the Technical Committees on Biorobotics and of Robo-ethics of the same society. He is a Fellow of the European Society on Medical andBiological Engineering, and a recipient of many honors and awards, such as theJoseph Engelberger Award. He is also a Member of the Board of the InternationalFoundation of Robotics Research (IFRR).

mapping multiple gas/odor sources in an uncontrolled indoor environment using a bayesian occupancy...

Documents