DustCart, a Mobile Robot for Urban Environments: Experiments of Pollution Monitoring and Mapping during Autonomous Navigation in

Urban Scenarios Gabriele Ferri1, Alessio Mondini1, Alessandro Manzi2, Barbara Mazzolai3, Cecilia Laschi2, Virgilio Mattoli3, Matteo

Reggente4, Todor Stoyanov4, Achim Lilienthal4, Marco Lettere5, Paolo Dario1,3

e-mail : g.ferri@sssup.it

1 CRIM Lab 2 ARTS Lab 3 Center in Micro-BioRobotics 4 Learning Systems Lab,AASS 5 Synapsis SRL

Scuola Superiore Sant’Anna Scuola Superiore Sant’Anna IIT@SSSA Italian Institute of Technology

Örebro University Piazza Dante 19/20

Viale Rinaldo Piaggio 34, 56025 Pontedera (Pisa), Italy

Viale Rinaldo Piaggio 34, 56025 Pontedera (Pisa), Italy

Viale Rinaldo Piaggio 34, 56025 Pontedera (Pisa), Italy

S-70182, Örebro, Sweden 57121 Livorno, Italy

Abstract – In the framework of DustBot European project, aimed at developing a new multi-robot system for urban hygiene management, we have developed a two-wheeled robot: DustCart. DustCart aims at providing a solution to door-to-door garbage collection: the robot, called by a user, navigates autonomously to his/her house; collects the garbage from the user and discharges it in an apposite area. An additional feature of DustCart is the capability to monitor the air pollution by means of an on board Air Monitoring Module (AMM). The AMM integrates sensors to monitor several atmospheric pollutants, such as carbon monoxide (CO), particular matter (PM10), nitrogen dioxide (NO2), ozone (O3) plus temperature (T) and relative humidity (rHu). An Ambient Intelligence platform (A mI) manages the robots’ operations through a wireless connection. AmI is able to collect measurements taken by different robots and to process them to create a pollution distribution map. In this paper we describe the DustCart robot system, focusing on the AMM and on the process of creating the pollutant distribution maps. We report results of experiments of one DustCart robot moving in urban scenarios and producing gas distribution maps using the Kernel DM+V algorithm. These experiments can be considered as one of the first attempts to use robots as mobile monitoring devices that can complement the traditional fixed stations.

Index Terms – mobile robots, urban robots, gas mapping, navigation.

I. INTRODUCTION

ITH the developing and the increasing maturity of research in mobile robotics, autonomous agents are

starting to go out of the laboratories, where most of the research has been carried out in the past, and are going to operate in real-world outdoor scenarios. This objective poses new challenges to the researchers’ community due to the intrinsically dynamic nature of outdoor environments: many techniques for localization and navigation working well in indoor do not scale well outdoor. For example, traditional scan matching and features extraction are susceptible to failures since typical structured elements like walls and corners might be rare or beyond sensors range and parked cars could be mistakenly associated with known features or can cause the

discarding of the range measurements losing possible useful information. On the other hand, technologies using GPS signal can be used to achieve an estimate of the robot position. Despite these difficulties, the development of autonomous agents able to move in outdoor environments can provide new systems to fulfill useful tasks in the near future such as cleaning, surveillance, transportation and environmental monitoring. In the framework of DustBot European project [1], [2] we have developed a multi-robot system for urban hygiene management. One of the two kinds of robot created during the project is DustCart. DustCart is a two-wheeled robot, based on the commercial Segway RMP200 platform [3]. The robot is designed to provide an automatic door-to-door garbage collection service in the pedestrian areas of the historical centers of towns where other larger vehicles could face difficulties of movement. In a typical operational scenario, a user requests garbage removal by placing a call to the automated DustBot call center. A robot is then dispatched to the predefined user home address. The robot interacts with the user through a touchscreen and receives a garbage bag. Then, it moves to a discharging site where it deposits the bag in different locations based on the type of garbage the user chose on the touchscreen. The robot operations are managed by an Ambient Intelligence platform (AmI), which receives the phone calls, selects the robot to move and manages the robots’ navigation. The capability of DustCart to autonomously navigate in real urban environments avoiding static and dynamic obstacles suggested us to endow it with additional features: the possibility of providing useful information about the city through the touchscreen and the capability of monitoring the air quality using different environmental sensors. Different kinds of sensors were integrated in a robot’s module called Ambient Monitoring Module (AMM) connected to the main robot’s controller board (see Sec. 2) including NO2, O3, CO, PM10, temperature and humidity sensors. Vehicular traffic, industrial exhaust and the fossil fuels combustion have a well known impact on the health of people, especially for people living in cities (about 50% of the total world population [4]) so an accurate and continuous monitoring of these chemicals in urban areas is of common interest. Nowadays, these

W

pollutants are monitored by stationary stations located in strategic positions of the cities. The monitoring stations send the measured pollutant levels to a central station for data processing. To increase the spatial resolution of the monitoring, a possible solution is the use of autonomous robots equipped with environmental sensors. This idea was pursued in the DustBot project: DustCart can act as a mobile monitoring station in a network of “moving sensors”. The robots move around the city to fulfill their tasks sending the acquired measurements to AmI. AmI, in this way, can create pollutant dispersion maps based on data coming at the same time from several locations of the city offering a synoptic view of the environmental conditions of a given area. Due to power consumption, dimensions and cost constraints, the sensors that can be mounted on the robots are less accurate and sensitive than those deployable in fixed monitoring stations. The basic idea is that the robots do not substitute the fixed monitoring stations, but they operate as complementary systems providing more pervasive information about the air quality of our cities. Research on using chemical sensors in robotic systems started in early 90s essentially focusing on the development of two typologies of robots: the first typology is able to follow odor trails marked on the ground as ants follow a pheromone trail [5]; the second one is able to track odor trails formed in fluid media (air or water) to find a chemical source [6], [7], [8]. Complementary to these approaches, the creation of chemical distribution maps can provide an invaluable help in localizing a pollutant source and also gives information about the actual distribution of a gas in an area of interest. Building gas distribution models is a challenging task given that, in many realistic scenarios, gas dispersion is dominated by turbulent advection. Turbulent flows create packets of gas following chaotic trajectories [9]. In principle, CFD (Computational Fluid Dynamics) models can be applied. However, CFD models are computationally very expensive and become intractable for sufficiently high resolution in typical real world settings. Most importantly they depend sensitively on the accurate knowledge of the state of the environment (boundary conditions), which is not available in practical situations. The simplest solution to build a map of the average gas concentrations is to use measurements acquired by a grid of fixed sensors [10]. This solution offers the advantage of reduced time consumption to build a map but it needs much effort in sensors calibration and, furthermore, lacks in scalability because, increasing the area, the number of sensors needs to be incremented with problems of cost and maintenance. The use of mobile robotic agents to produce gas distribution maps offers more flexibility. Hayes et al. [11] proposed to build a two-dimensional histogram over the number of “odor hits” (sensed values above a certain threshold) received in the corresponding area by six robots moving along a random walk path. The robots have to cover completely all the area of interest to produce a reliable map and the “odor hits”, given their binary nature, strongly depend on the chosen threshold. In general, during the mapping task, the

robots have to scan the area using a number of sampling locations relatively small not to make the task too time-expensive: an extrapolation process is therefore needed. Proposed methods in literature have built maps of concentrations using different kinds of extrapolation [12], [13], [14], [15]. Another approach is to create maps of probabilities a source is present in a certain location [16], [17], [18]. In [12] the measured gas intensities are convolved with a Gaussian kernel and experiments were carried out in an uncontrolled indoor environment. In [13] gas measurements taken by 5 robots in a ventilated room are fused together with wind velocity in an advection-diffusion model (a non-turbulent transport model was assumed) using a Kalman Filter. In [14] experiments with a mobile robot driven on a zig-zag trajectory in a ventilated environment are reported. Since the measurement locations were not equally distributed, triangle-based cubic filtering was used for interpolation. The more jagged distribution with respect to that produced in the same environment from the averaged values sensed by a stationary sensor [14] shows a flaw of the interpolation. There is no means of averaging out the fluctuations of the gas concentration measured at the sampling locations. A different approach aiming at addressing these problems was proposed by Stachniss and co-workers [15] in which a sparse Gaussian process mixture model is used to represent the smooth background signal and to model sharp boundaries around patches of high concentrations.

Techniques to estimate the probability a source is in a certain position have been proposed in [16] (hidden Markov methods) and in [17], [18] (Bayesian methods) . The methods proposed in [16] and [17], based on flow measurements and on binary detections and non detections of chemical effluent, update the probability that some cells on a grid contain a source. These methods were applied in underwater scenarios and are based on the key assumption that a single source is present in the search area. In [18] a Bayesian occupancy grid mapping method, not relying on the assumption of a single source present in the monitored area, was used in a multi-source indoor environment with no artificial ventilation to produce the probability of each discrete cell in a grid contains an active plume source. So far, research on robots using gas sensors has been conducted in simplified indoor environments while only recently few examples of robots in outdoor experiments have been reported [19], [20]. To address the gas distribution mapping issue in outdoor environments different models were developed for atmospheric dispersion [21], [22]. RIMPUFF, for example, is a Gaussian puff model used to calculate the dispersion of airborne materials at the mesoscale under the condition of a moderate topography [22]. However, such models cannot capture all the relevant aspects of gas propagation with a sufficient level of detail. High resolution models are required particularly at small scales and in typical complex indoor and outdoor settings where critical gas concentrations often have a local character. In this paper we show the gas distribution maps created

using the Kernel DM+V algorithm [23]approach to extrapolate and build twodistribution models from a sequence of localized measurements acquired by DustCart during in urban environments. This algorithm fitsof high resolution gas distribution modeling, thanks to the application of the kernel DM+V [23]sensor measurements as random variables. This modelnon-parametric in the sense that it makes no assumptions about a particular functional form of the modelin this way the limitation of a perfect boundary conditions, as in the case of fluid dynamics model. The learned model is represented as a pair of 2grid maps, one representing the distribution mean and theother one the corresponding predictive variance per grid cell. In contrast to the covariance of the mean, which only decreases as new observations are processed, our estimate of the variance will adapt to the real variability of gas readings at each location. In [19] and [20]been extended taking into account the by modeling the information content of the gas sensor measurements as a bivariate Gaussian kernel whose shape depends on the measured wind vector, and evaluated in different uncontrolled environments, from a big room to an outdoor scenario.

II. DUSTCART ROBOT

The locomotion solution adopted for DustCart (height 150 cm, weight 85 kg) (see Fig. 1) is based on a Segway RMP200 platform. This platform guarantees a high maneuverability in the tight urban spaces. A metallic frame (see Fig.1- b) was built on the RMP200 to contain robot’s electronics and the bin (77 literscan be opened/closed by the robot. A head (that can rotate) with LEDs in the eyes is placed on top of the structure and it is used for human-robot interaction. LEDs on the eyes, the head direction and acoustic signals or recorded vocal messages provide information to the user about the state of the robot. The system architecture is shown in Fig. 2. The core part of DustCart is the Supervisor software running on an industrial PC with a Pentium M 1.7 GHz processor, 512 MB RAM, with a Linux (Ubuntu 8.04) OS. The Supervisor manages all the robot’s operations: it acquires the data coming from the sensors, communicates with AmI through a wireless connection (using Internet), receives the commands (tasks and destination points) and then communicates with theother boards (all based on ARM9 STR912FSTMicroelectronics) using a CAN bus as a communication channel. In particular, the AMM modulecollects raw measurements from the sensors at different frequencies, processes them and sends the dataSupervisor PC through the CAN bus (see Sec 2.1).

2.1) Ambient Monitoring Module (AMM)

The Ambient Monitoring Module (AMM) integrates several environmental sensors for pollution monitoring. The selection of the sensor types was done on the base of several characteristics such as accuracy, sensibility,

[23], a statistical approach to extrapolate and build two-dimensional

from a sequence of localized measurements acquired by DustCart during the navigation

This algorithm fits the requirement of high resolution gas distribution modeling, thanks to the

[23], that treats the measurements as random variables. This model is

makes no assumptions form of the model, avoiding

knowledge of the , as in the case of fluid dynamics

The learned model is represented as a pair of 2-d grid maps, one representing the distribution mean and the other one the corresponding predictive variance per grid

the covariance of the mean, which only observations are processed, our estimate

adapt to the real variability of gas [20] this method has

the wind information the information content of the gas sensor

bivariate Gaussian kernel whose shape , and evaluated in

uncontrolled environments, from a big room to

The locomotion solution adopted for DustCart (height 150 ) is based on a Segway

This platform guarantees a high in the tight urban spaces. A metallic b) was built on the RMP200 to contain

77 liters of volume) that can be opened/closed by the robot. A head (that can rotate) with LEDs in the eyes is placed on top of the

robot interaction. LEDs on the eyes, the head direction and acoustic signals or

sages provide information to the user about the state of the robot. The system architecture is

. The core part of DustCart is the Supervisor software running on an industrial PC with a

GHz processor, 512 MB RAM, with a The Supervisor manages all the

acquires the data coming from the sensors, communicates with AmI through a wireless connection (using Internet), receives the commands (tasks nd destination points) and then communicates with the

STR912F boards from STMicroelectronics) using a CAN bus as a

In particular, the AMM module collects raw measurements from the sensors at different frequencies, processes them and sends the data to the

bus (see Sec 2.1).

) Ambient Monitoring Module (AMM)

The Ambient Monitoring Module (AMM) integrates several environmental sensors for pollution monitoring.

done on the base of such as accuracy, sensibility, time

(a) (b) Fig. 1. DustCart robot pictures. (a) Robot with the navigation sensors. (b) Robot with the open bin. (c) Touchscreen and interaction with the user.

Fig. 2. The system architecture. Dashed rectangles represent distinct hardware components on the robot, solid rectangles representprocesses. The AMM module is showeddata from the sensors at different frequencies, processes them and sends them to the Supervisor PC through the CAN bus.of response and power consumption. Our choice was mainly driven by the working rangeconsidering the pollution limits imposed by the European Commission. In particular, regarding pollution, in 1996, the Environment Council adopted Framework Directive 96/62/EC on ambient air quality assessment and management. This Directiverevision of previously existing legislation and the introduction of new air quality standards for previously unregulated air pollutants, setting the development of daughter directives pollutants including nitrogen dioxide (NOmatter (PM10-PM2.5), ozone (O(CO), which we selected to monitorachieve a good compromise between accresponsiveness more than onewere integrated in the module (usually faster sensor are less accurate). The core of the Air Monitoring Module (AMM) is an ARM9 microcontroller from STmicrocontroller is devoted to communicate with the Supervisor and the sensors. The AMM is connected to the Supervisor by a CAN bus, as the otherrobot. The AMM is connected I2C modified bus. Fig. 3 showsAMM system. The sensors used in the DustCartoutput signal, measured physical quantity, supply voltage, sensitivity, accuracy, etc. For these reasons,be integrated directly in the AMM, but

(b) (c)

obot with the navigation sensors. (b) Robot with the open bin. (c) Touchscreen and interaction with the user.

The system architecture. Dashed rectangles represent distinct

he robot, solid rectangles represent software The AMM module is showed in a dashed ellipse: it collects

data from the sensors at different frequencies, processes them and sends them to the Supervisor PC through the CAN bus.

consumption. Our choice was working range of the sensors, limits imposed by the European

. In particular, regarding the atmospheric pollution, in 1996, the Environment Council adopted the

Directive 96/62/EC on ambient air quality assessment and management. This Directive covers the revision of previously existing legislation and the introduction of new air quality standards for previously unregulated air pollutants, setting the schedule for the development of daughter directives about a range of

nitrogen dioxide (NO2), particulate ozone (O3), and carbon monoxide

to monitor. Furthermore, to achieve a good compromise between accuracy and

e sensor for each pollutant integrated in the module (usually faster sensor are

The core of the Air Monitoring Module (AMM) is an ARM9 microcontroller from STMicroelectronics. This

to communicate with the The AMM is connected to the

upervisor by a CAN bus, as the other sub-modules of the he AMM is connected to the sensors through an

shows a block diagram of the

DustCart are different in terms of output signal, measured physical quantity, supply voltage,

For these reasons, they cannot the AMM, but a first stage of

conditioning is necessary. Furthermoredesigned to be “open”, providing the capabilitynew sensors (or substituting old sensors with new calibrated ones) at any time with no need of modifyingconfiguration parameters in the robot. To perform these actions the Universal Smart Sensor Interface (USSI) used [24]. This intermediate hardware interface provides two important features: the capability to the configuration of analogical circuitry parameters configuration and the possibilitysensors plug & play. These characteristicthe use of a communication protocol compliant tfamily standard 1451 and by the usedescriptor, the Transducer Electronic Data Sheet (TEDS)[25], [26]. By the USSIs 8 different sensors wereintegrated (Fig. 3). Fig. 4 shows the developed prototype integrated in the DustCart robot.

2.2) Navigation and localization in urban environments

Navigation: The navigation strategy implemented for the Dustrobots is conceptually structured in thr(see Fig. 5). In a typical use case, upon a call by one registered user, AmI is responsible for the allocation, monitoring and successful execution of a garbage collection task. In the context of robot navigation, maintains a geo-referenced map of the city, along with the locations of known users, available roads and important locations. AmI identifies the user, selects an idle robot in the vicinity and plans a path using a Dijkstra algorithm in a network of predefined waypoint locationsspaced about 20 m. Once the DustCart robot is assigned toa task, an iterative procedure guides it through the city. The robot receives the coordinates of the next goal point, together with a map of the relevant portion of the city. Using the geographical coordinates of the goal point and map center, the robot transforms all locations into a local coordinate system and plans a path to the goal location. At this point the problem is essentially reformuclassical motion planning problem and solved using the wavefront path planning algorithm generates a sequence of waypoints that have to by the robot. An obstacle avoidance strategy is used to guide the robot to the next waypoint, while avoidcollisions with obstacles not present in the map. The Vector Field Histogram (VFH+) algorithmchosen as a suitable obstacle avoidance and following approach. Sonar and laser range measurementsare fused and used to compute the safest heading and speed for the robot. For this purpose, DustCart uses data from the Hokuyo UTM-30LX laser range finder and five SRF10 sonarsfrom Devantech (see Fig.1). Once DustCart reaches the goal point, AmI sends the next one and eventually another part of the map. In this way AmI can change dynamically the maps and the path of the robot if some changes occur (e.g.: parked cars change the possible path the rocover). DustCart navigation software is based onopen-source framework Player [30].

. Furthermore, the platform is capability of adding

old sensors with new need of modifying the

figuration parameters in the robot. To perform these actions the Universal Smart Sensor Interface (USSI) was

This intermediate hardware interface provides the capability to simply change

circuitry by software and the possibility to make the

characteristics are achieved by communication protocol compliant to the

the use of the sensor Transducer Electronic Data Sheet (TEDS)

he USSIs 8 different sensors were shows the developed prototype

2.2) Navigation and localization in urban environments

The navigation strategy implemented for the DustCart robots is conceptually structured in three control layers

upon a call by one is responsible for the allocation,

monitoring and successful execution of a garbage robot navigation, AmI

referenced map of the city, along with the locations of known users, available roads and important

identifies the user, selects an idle robot in using a Dijkstra [27]

in a network of predefined waypoint locations art robot is assigned to

an iterative procedure guides it through the city. The robot receives the coordinates of the next goal point, ogether with a map of the relevant portion of the city. Using the geographical coordinates of the goal point and map center, the robot transforms all locations into a local coordinate system and plans a path to the goal location. At

is essentially reformulated into a classical motion planning problem and solved using the

ront path planning algorithm [28]. Wavefront ence of waypoints that have to be reached

idance strategy is used to guide the robot to the next waypoint, while avoiding collisions with obstacles not present in the map. The Vector Field Histogram (VFH+) algorithm [29] was chosen as a suitable obstacle avoidance and path following approach. Sonar and laser range measurements are fused and used to compute the safest heading and

For this purpose, DustCart uses data from the Hokuyo 30LX laser range finder and five SRF10 sonars

Once DustCart reaches the goal point, AmI sends the next one and eventually another part of the map. In this way AmI can change dynamically the maps and the path of the robot if some changes occur (e.g.: parked cars change the possible path the robot can

navigation software is based on the

Fig. 3. Scheme of AMM

Fig. 4. Sensors integrated in the

Fig. 5. Scheme of the Navigation system.

Localization: The DustCart localization system hasdue to the urban scenarios in which the robot operate. The urban areas can large squares where GPS signal can be used but where the walls of the buildings are too far to be detected by robot’s exteroceptive sensors or narrow streets where GPS is not reliable. Furthermore, the presence of peoaround the robot has to be taken into considerationAiming at developing a robust localization systemvariable conditions we do not rely on the detection of natural features in the environment, since such detection is often subject to failure and not very accuratemay not be present (curbs, lanes demarcation), or they may be too far to be detected or obstructed by people walking. Furthermore, the use of vision is subjected to the problem of changes in light conditions given has to work day and night. The proposed system is based on the position/heading estimates coming from deadreckoning based on wheels encoders and a magnetic

AMM module architecture.

Sensors integrated in the AMM module.

Scheme of the Navigation system.

DustCart localization system has strict requirements due to the urban scenarios in which the robot has to

eas can present different features: large squares where GPS signal can be used but where the walls of the buildings are too far to be detected by robot’s exteroceptive sensors or narrow streets where GPS is not reliable. Furthermore, the presence of people walking

to be taken into consideration. ng a robust localization system in these

do not rely on the detection of natural features in the environment, since such detection is

ilure and not very accurate: they may or may not be present (curbs, lanes demarcation), or they may be too far to be detected or obstructed by people walking. Furthermore, the use of vision is subjected to the

in light conditions given that DustCart has to work day and night. The proposed system is based on the position/heading estimates coming from dead-reckoning based on wheels encoders and a magnetic

compass. To limit the drift of the dead-reckoning estimate error we use different methods: GPS data when fixes with good merit factors are produced (this may be rare in urban areas with narrow streets and high buildings);artificial landmarks easily detectable by robot’s laser scanner; and, when particular precision is needed (i.e.during docking tasks), a network of ultrasonic beacons are used to trilaterate robot’s position. All the different sensor measurements are fused together using an Extended Kalman Filter (EKF) [31] based on a differential drive kinematics model of the robot

where x,y represent the coordinates of the robot in a reference system aligned with north and east, robot heading with respect to north, v isspeed, is the process noise vector antime equal to 0.05 s. The filter is run every the “correction” step data coming from different sensors are used when present: different data can come from encoders, magnetic compass, GPS, laser scanneroptical beacons and ultrasound receivers for the acoustic beacons (not described in this paper).

Encoders and compass: The encoders of the wheels are read every 10 ms. Every 5 data read from wheels encoders a step of the filter is run. An electronic magnetic compass (TCM2.6 from PnI Corps) is used (8 Hz frequency) to improve the estimate of heading angle. Experiments on a rectangular path of 4 m x 6 m with a robot’s speed of 0.8 m/s show an average error of 0.56 m in covered distance on 10 trials, error in heading that can be considered lower than 3 degrees. The performance of the robot’s deaddepends on a large number of factors such as theof the floor and the dynamics of the movement. However, during these tests, the floor was a typical urb(similar to asphalt) and the robot performed some tight turns. For these reasons the results can give a reasonable estimate of an average error drift: approximately 1 m every 20 covered meters. GPS: A low-cost commercial Superstar 2 GPS receiNovAtel is used to provide a latitude/longitude measurement at 1 Hz frequency. The GPS receiver performance was characterized through several experiments in different environments showing errors strongly depending on the GPS reported figure(horizontal figures of merit). The GPS fixes are inserted in the filter only when some conditions on the figure of merits are respected. Artificial beacons for localization: The used landmarks are a combination of an optical beacon and a radio module (see Fig. made from plexiglass tubes (1m high, 5 cm diameter)

reckoning estimate hods: GPS data when fixes with

good merit factors are produced (this may be rare in urban row streets and high buildings); optical

artificial landmarks easily detectable by robot’s laser when particular precision is needed (i.e.

a network of ultrasonic beacons are ll the different sensor

are fused together using an Extended based on a digitalized

of the robot:

represent the coordinates of the robot in a

reference system aligned with north and east, ψ is the is the translational

is the process noise vector and ∆T the sampling The filter is run every 50 ms and for

the “correction” step data coming from different sensors are used when present: different data can come from

tic compass, GPS, laser scanner for the s and ultrasound receivers for the acoustic

are read every 10 ms. Every 5 data read from wheels encoders a step of the filter is run.

TCM2.6 from PnI to improve the estimate

of heading angle. Experiments on a rectangular path of 4 m x 6 m with a robot’s speed of 0.8 m/s show an average

n covered distance on 10 trials, with an ing that can be considered lower than 3

The performance of the robot’s dead-reckoning umber of factors such as the typology dynamics of the movement. However,

the floor was a typical urban pavement (similar to asphalt) and the robot performed some tight turns. For these reasons the results can give a reasonable

error drift: approximately 1 m

cost commercial Superstar 2 GPS receiver from NovAtel is used to provide a latitude/longitude measurement at 1 Hz frequency. The GPS receiver performance was characterized through several experiments in different environments showing errors strongly depending on the GPS reported figures of merit

of merit). The GPS fixes are inserted in the filter only when some conditions on the figure of

ombination of an optical Fig. 6). Beacons were

tubes (1m high, 5 cm diameter)

wrapped with reflective material, topped with a radio module CC2430 from Texas Instruments. The reflective surface registers as a high intensity point when hit by a laser beam from the robot's scanner. Thus, it is possible to identify reliably the beacons when in line of sight and up to a distance of about 7 meters. When the robot observes a laser endpoint with high reflectivity, the position of that point is calculated using the current robot position and heading estimate. associate the observed point to one particular beacon, DustCart checks if the distance between the estimated point position

Bp and the known position of one of the

beacons i , Bip , is lower than a certain threshold

γ<− BBi pp ˆ

Once this relation is satisfied by the beacon observed point is associated with the position of that beacon and an estimate of the robot position is computed based on the difference between the known and the estimated position of the beacon.position fix is then used as an observation in the EKF with a maximum update frequency of 10 Hz. precision of the laser in measuring the bearing and the distance to a beacon, the errordepends on the accuracy of heading estimatelinearly with the distance to the beacon. Thewhich a beacon is observedassociated measurement noise in the EKF.results show the maximum error is less than 40 cm at a distance of 7 m. The system appears to give good position estimates and one beacon can be seen from a relalarge area (a 7 m radius circle).However, two issues have to be carefully considered for a safe navigation: the spacing with which the beacons have to be placed and the robustness to external disturbances (parked and moving bicycles and cars actingpositives). Concerning the distance, the results shown in the previous sections (not considering the possible help given by GPS) suggests a drift of about 1 m in position estimate every 20 m of covered distance: for this reasonwe placed a visible beacon every 20 m to keep the estimate error sufficiently low for a safe navigation(< 3 m). The disturbance rejection is addressed changing the threshold γ during the robot navigationradio communication with the beacons. A lower threshold aims at not considering as possible beacons reflective objects close to the actual beacons.increases (the drift in position error increments) thethreshold is raised to not miss a true beacon.

Fig. 6. An optical beacon used as an artificial landmark for robot navigation. In the image the pole (1 m high, 5 cm diameter) wrapped with reflective material and a CC2430 module for wireless communication.

wrapped with reflective material, topped with a radio module CC2430 from Texas Instruments. The reflective

h intensity point when hit by a laser beam from the robot's Hokuyo UTM30-LX laser

Thus, it is possible to identify reliably the of sight and up to a distance of about

When the robot observes a laser endpoint with h reflectivity, the position of that point is calculated

using the current robot position and heading estimate. To associate the observed point to one particular beacon, DustCart checks if the distance between the estimated

and the known position of one of the

is lower than a certain threshold γ, that is,

(2)

Once this relation is satisfied by the beacon i, the observed point is associated with the position of that

acon and an estimate of the robot position is computed based on the difference between the known and the estimated position of the beacon. The so obtained x-y position fix is then used as an observation in the EKF with a maximum update frequency of 10 Hz. Considering the

laser in measuring the bearing and the error of the estimate essentially

heading estimate and increases th the distance to the beacon. The distance to beacon is observed is used to modify the

associated measurement noise in the EKF. Experimental results show the maximum error is less than 40 cm at a distance of 7 m. The system appears to give good position estimates and one beacon can be seen from a relatively large area (a 7 m radius circle). However, two issues have to be carefully considered for a safe navigation: the spacing with which the beacons have to be placed and the robustness to external disturbances (parked and moving bicycles and cars acting as false positives). Concerning the distance, the results shown in the previous sections (not considering the possible help given by GPS) suggests a drift of about 1 m in position estimate every 20 m of covered distance: for this reason

e beacon every 20 m to keep the estimate error sufficiently low for a safe navigation(< 3 m). The disturbance rejection is addressed changing the

the robot navigation and performing a radio communication with the beacons. A lower threshold aims at not considering as possible beacons reflective objects close to the actual beacons. As the covered space

sition error increments) the is raised to not miss a true beacon.

. An optical beacon used as an artificial landmark for robot

navigation. In the image the pole (1 m high, 5 cm diameter) wrapped with reflective material and a CC2430 module for wireless

Once the threshold reaches the maximum valuethe robot sees a high reflective point, CC2430 module if there are beacons nearby. If one beacon is present, it answers with its identifinumber allowing in this way an association (the range of the radio module is opportunely tuned). In this way, even if the robot gets lost (high error in localizatsees a beacon it can reset the error and can reitself. The method proved to be robust even with a high concentration of false positives in Örebro (Sweden).

III. MAPS CREATION

3.1) Mapping algorithm

In this section we briefly describe the bKernel DM+V algorithm, a detailed description could be found in [23]. The general gas distribution mapping problem addressed is to learn a predictive twodimensional model p(r|x,x1:n,r1:n) for the gas readingcoming from one sensor at location xtrajectory x1:n and the corresponding measurements The central idea of kernel extrapolation methods is to understand gas distribution mapping as a density estimation problem addressed using conGaussian kernel N. The first step in the algorithm is the computation of weights ωi

(k), which intuitively represent

the information content of a sensor measurement cell k:

ωi(k)(σ0 )=N(|xi -x

(k)|, σ0 )

The weights are computed by evaluating a Gaussian kernel N at the distance between the location of the measurement xi and the center x(k) of cell the kernel width. Using (3), weights ωi

(k

readings ωi(k)·r i , and weighted variance contributions

ωi(k)·τi are integrated and stored in temporary grid maps:

Ω(k)(σ0 )=∑i=1:n ωi(k)(σ0 ),

R (k )(σ0 )=∑i=1:n ωi(k)(σ0 )·r i

V (k)(σ0 )=∑i=1:n ωi(k) (σ0 )·τi

where, τi =(r i -r ik(i))2 is the variance contribution of

reading i and rk(i) is the mean prediction of the cell closest to the measurement point xi. From the integrated weight map Ω(k) we compute a confidenceindicates high confidence for cells for whichnumber of readings close to the center of the grid cell is available. The confidence map is computed as

α(k) (σ0 )=1- exp(-Ω(k)(σ0)2/ σ Ω 2)

where σΩ is a scaling parameter that defines a soft marginwhich decides whether the confidence in the estimate for acell is high or low. By normalising with the integrated weights Ω(k) and linear blending with a best guess for the case of low confidence, we finally obtain the map estimate of the mean distribution r(k) and the corresponding variance map v(k):

r(k)(σ0 )= α(k) R(k)/ Ω(k)+ 1 - α(k) ·r*

Once the threshold reaches the maximum value, whenever a high reflective point, checks via a

here are beacons nearby. If one beacon is present, it answers with its identification number allowing in this way an association (the range of the radio module is opportunely tuned). In this way, even if the robot gets lost (high error in localization), when it

it can reset the error and can re-localize The method proved to be robust even with a high

rebro (Sweden).

In this section we briefly describe the basic ideas of the a detailed description could be

The general gas distribution mapping problem addressed is to learn a predictive two-

) for the gas reading r x, given the robot

and the corresponding measurements r1:n. The central idea of kernel extrapolation methods is to understand gas distribution mapping as a density estimation problem addressed using convolution with a

. The first step in the algorithm is the , which intuitively represent

the information content of a sensor measurement i at grid

(3)

uted by evaluating a Gaussian at the distance between the location of the

of cell k, where σ0 is k), weighted sensor

, and weighted variance contributions

are integrated and stored in temporary grid maps:

(4)

is the variance contribution of

prediction of the cell k(i) From the integrated

we compute a confidence map α(k) which indicates high confidence for cells for which a large number of readings close to the center of the respective grid cell is available. The confidence map is computed as

(5)

is a scaling parameter that defines a soft margin which decides whether the confidence in the estimate for a

By normalising with the integrated and linear blending with a best guess for the

confidence, we finally obtain the map estimate and the corresponding

(6)

v(k)(σ0 )= α(k) R(k)/ Ω(k)+ 1 - α(k)

The second terms in these equations contain the best estimate, which is used for cells with a low confidence, i.e. for cells for which we do not have sufficient information from nearby readingvalue of α(k). As the best guessr* we use the average over allestimate vtot of the distribution variancemeasurement points is computed as thevariance contributions. The parameters to be tuned for thefollowing: the kernel width σconfidence scaling parameter σof extrapolation on individual readings and the cell size determines the resolution at which difcan be made. σΩ defines a soft margin on the confidence estimate, which is used to decide whether sufficient information is available to esmean and variance for a given grid cell.obtained for different parameters are

Fig. 7. Weight map Ω(k) (top row) and the corresponding confidenceα

(k) (bottom row) obtained using the parand σ0=100cm, c=10 cm (right).

3.2) Maps creation process

To create the maps, Supervisor (every 2 seconds) the data coming from the AMM andcorrelates them with time and the by the Localization Module. The created data packet is transmitted to the AmI infrastructure Communication Module (this module uses a WiFi, UMTS or GPRS channel according to the current signals strength). The data are collected by AmI and stored in a database; when a user makes a query for some data, the AmI platform performs a call to the Pollution Model Server where the data are results are sent back in image, graphic or table format.the data exchanged between DustCart Supervisor PC are embeddeddocument format.

Fig. 8. Environmental data management architecture

·vtot (7)

The second terms in these equations contain the best which is used for cells with a low confidence,

for which we do not have sufficient readings, indicated by a low

. As the best guess of the mean concentration we use the average over all sensor readings. The

of the distribution variance in regions far from measurement points is computed as the average over all

parameters to be tuned for the algorithm are the σ0, the cell size c and the σΩ. σ0 governs the amount

of extrapolation on individual readings and the cell size determines the resolution at which different predictions

defines a soft margin on the confidence estimate, which is used to decide whether sufficient information is available to estimate the concentration mean and variance for a given grid cell. The Ω(k ) and α(k)

different parameters are showed in Fig. 7.

(top row) and the corresponding confidence map

(bottom row) obtained using the parameters σ0=15cm, c=10 cm (left)

, Supervisor acquires periodically 2 seconds) the data coming from the AMM and

correlates them with time and the position data produced . The created data packet is

AmI infrastructure through the (this module uses a WiFi, UMTS

ng to the current available The data are collected by AmI and

stored in a database; when a user makes a query for some data, the AmI platform performs a call to the Pollution

where the data are processed and then the sent back in image, graphic or table format. All exchanged between the AmI core and the

are embedded into XML

data management architecture.

This system enables AmI to build mapsame time data from different robots, allowinpossibility of using also mapping algorithmsconcurrently from multiple robots.

IV. RESULTS

During the DustBot project we demonstrated the systemseveral public demos in uncontrolled urban environments: Pontedera (Italy), Peccioli (Italy), Massa (Italy),(Sweden) and Bilbao (Spain). We report here the gas distribution mapping experiments carried out in a pedestrian area of Örebro (Sweden) (see Fig.9robot covered a path for a collecting task: it started from a docking station (blue spot in Fig. 9), wenthouse to collect garbage (red spot) and thendischarging area before coming back to the docking station. During the movement, the robot acquired pollution level measurements from the sensors frequency of 0.5 Hz. The robot movedspeed of 1 m/s. A controlled fire was placed in the square as a pollution source (orange spot in Fig. 9

Fig.9. Top left: Aerial view of the experimental pedestrian area (Stortorget, Örebro – Sweden). Top right: DustCart in the experimental scenario. Bottom: The path covered by the robotrobot’s localization module). The pollution sourcethe map (orange spot): the source was a controlled fire placed in the square. Finally the user house location (red spot) and location (green square) are showed.

Fig. 10 shows an example of the weight map the corresponding confidence map α(k) (right) model parameter σ0=150 cm and grid cell siThe two red spots in the weight map are in correspondence of two stops of the robot, wherereadings for the same location were gathered. One of the stops occurred when the robot interacted with the customer and one when discharged the garea in the confidence map represents“high” confidence, where the predictions frextrapolation are considered reliable. Fig.11 shows the raw sensors response movement in the inspected area for the PM10 pollutantand the limit imposed by the law (red line). The peakexceeding the law limit is sensed when the robot passesclose by the gas source. The learned model is

This system enables AmI to build maps receiving at the same time data from different robots, allowing the

mapping algorithms using data

During the DustBot project we demonstrated the system in uncontrolled urban environments:

Pontedera (Italy), Peccioli (Italy), Massa (Italy), Örebro We report here the gas

distribution mapping experiments carried out in a (see Fig.9). DustCart

path for a collecting task: it started from a went to the customer

and then moved to the back to the docking

the robot acquired the from the sensors at a

moved at a maximum controlled fire was placed in the square

(orange spot in Fig. 9).

. Top left: Aerial view of the experimental pedestrian area

DustCart in the experimental The path covered by the robot (blue line) (output of

pollution source position is showed on a controlled fire placed in the (red spot) and the discharging

shows an example of the weight map Ω(k) (left) and (right) created with

grid cell size 10 x 10 cm. The two red spots in the weight map are in

f two stops of the robot, where many readings for the same location were gathered. One of the stops occurred when the robot interacted with the customer and one when discharged the garbage. The red

represents locations with “high” confidence, where the predictions from

shows the raw sensors response during the the PM10 pollutant

and the limit imposed by the law (red line). The peak exceeding the law limit is sensed when the robot passes

The learned model is then

represented as a pair of 2-d grid maps, one repthe distribution mean (6) and thethe corresponding predictive variance per grid cellcontrast to the covariance of the mean, which onlydecreases as new observations are processed, our estimate

Fig. 10. Weight map Ω(k) (left) and the corresponding confidenceα (k) (right) obtained using the model size=10 cm x 10 cm.

Fig. 11. Raw response from the DustTrak 8520 sensor

of the variance will adapt to the real vareadings at each location. Fig.12 shows the gas distribution maps obtained by DustTrak 8520 sensor. The left part shows the mean distribution and the right part the variance distribution. Both the two maps have a maximum region closesource location. Here it is interesting to note that the variance distribution has a maximum in the exact location of the source. This is in accordance with empirical knowledge that we have from previous work, where we observed that the variance diinformation about the source location. Fig.13 shows the gas distribution maps obtained by CO sensor from Applied Sensors. The left part showsmean distribution of the CO pollutant, the right part variance distribution. Also in this case the variance map shows high value close to the gas source, as well as the mean map, where the high concentrations levels spread in a larger area.

V. CONCLUSIONS

In this paper we addressed the problem of mapping chemicals in real urban environmentsagents. To this aim, we equipped DustCart robot, a two wheeled robot developed in the framework of DustBot European project [1], with different environmental sensors. We developed and tested different autonomous agents developed) and an Ambient IntelligenceAmI manages the robots’ movements andenvironmental data produced byAmI can process measurementsrobots to produce chemical distribution maps.

d grid maps, one representing and the other one representing

the corresponding predictive variance per grid cell (7). In the covariance of the mean, which only

observations are processed, our estimate

) and the corresponding confidence map

model parameters σ0=150cm, grid cell

Raw response from the DustTrak 8520 sensor (PM10 pollutant).

adapt to the real variability of gas

shows the gas distribution maps obtained by the rak 8520 sensor. The left part shows the mean

distribution and the right part the variance distribution. Both the two maps have a maximum region close to the source location. Here it is interesting to note that the variance distribution has a maximum in the exact location of the source. This is in accordance with empirical knowledge that we have from previous work, where we observed that the variance distribution has major information about the source location.

shows the gas distribution maps obtained by the CO sensor from Applied Sensors. The left part shows the

CO pollutant, the right part the o in this case the variance map

the gas source, as well as the the high concentrations levels are

ONCLUSIONS

addressed the problem of mapping vironments using autonomous

e equipped DustCart robot, a two wheeled robot developed in the framework of DustBot

, with different environmental and tested a system composed by

(2 DustCart robots were and an Ambient Intelligence infrastructure.

AmI manages the robots’ movements and collects environmental data produced by the robots. In this way

measurements produced by different chemical distribution maps. We showed

results of tests of the whole system and the using the Kernel DM+V [23] algorithmDustCart autonomously navigating in garbage colltasks. Results show the system can workrobots can be used as moving monitoring stations to provide a synoptic and pervasive view of the air quality in a city not achievable by traditional fixed stations.

Fig. 12. Gas Distribution Map from the DustTrak 8520 sensor. Left: Mean distribution. Right: Variance distributionrepresents the source location.

Fig. 13. Gas distribution map from the CO sensorLeft: Mean distribution. Right: Variance distributionrepresents the source location.

REFERENCES

[1] www.dustbot.org [2] B. Mazzolai, V. Mattoli, C. Laschi, P. Salvini, G. Ferri, G.

Ciaravella, and P. Dario, “Networked and Cooperating RobotUrban Hygiene: the EU funded DustBot ProjectConference on Ubiquitous Robots and Ambient Intelligence (URAI 2008), Seoul, Korea, November 20-22, 2008.

[3] www.segway.com [4] http://www.unfpa.org [5] R. A. Russell, “Laying and sensing odor markings as a strategy for

assisting mobile robot navigation tasks”, IEEE Robot. Mag., Vol. 2, pp. 3–9, 1995.

[6] H. Ishida, K. Suetsugu, T. Nakamoto, and T. Moriizumi, “Study ofautonomous mobile sensing system for localization of odor sourceusing gas sensors and anemometric sensorsVol. 45, pp. 153–157, 1994.

[7] A. Lilienthal, A. Loutfi and T. Duckett, sensing with mobile robots”, Sensors, Vol. 6

[8] G. Ferri, E. Caselli, V. Mattoli, A. Mondini, B. Mazzolai and P. Dario, “SPIRAL: a Novel Biologically-Inspired Algorithm for Gas/Odor Source Localization in an Indoor Environment with no Strong Airflow”, Robotics and Autonomous SystemsIssue 4 , pp. 393-402, April 2009.

[9] B. Shraiman and E. Siggia, “Scalar Turbulencepp.639–646, 8 June 2000.

[10] H. Ishida, T. Nakamoto, and T. Moriizumi, “Remote sensing of gas/odor source location and concentration distribution using

and the maps produced algorithm using one

garbage collecting can work robustly and the

robots can be used as moving monitoring stations to provide a synoptic and pervasive view of the air quality in a city not achievable by traditional fixed stations.

DustTrak 8520 sensor. Left:

Mean distribution. Right: Variance distribution. The white circle

sensor from applied Sensors.

iance distribution. The white circle

B. Mazzolai, V. Mattoli, C. Laschi, P. Salvini, G. Ferri, G. Networked and Cooperating Robots for

Urban Hygiene: the EU funded DustBot Project”, 5th International Conference on Ubiquitous Robots and Ambient Intelligence (URAI

22, 2008.

A. Russell, “Laying and sensing odor markings as a strategy for IEEE Robot. Automat.

H. Ishida, K. Suetsugu, T. Nakamoto, and T. Moriizumi, “Study of r localization of odor source

sensors and anemometric sensors”, Sens. Actuators A,

A. Lilienthal, A. Loutfi and T. Duckett, “Airborne chemical 6, pp. 1616-1678, 2006.

, E. Caselli, V. Mattoli, A. Mondini, B. Mazzolai and P. Inspired Algorithm for

Gas/Odor Source Localization in an Indoor Environment with no Robotics and Autonomous Systems, Volume 57 ,

Scalar Turbulence “, Nature, 405,

T. Moriizumi, “Remote sensing of gas/odor source location and concentration distribution using

mobile systems”, Sensors and Actuators B1998.

[11] A.T. Hayes, A. Martinoli and R.M. Goodman, Distributed odor source localization, IEEE Sensors Journal, Vol.2, No.3273, 2002.

[12] A. Lilienthal, and T. Ducket, “Building gas concentration gridmaps with a mobile robot”, Robotics and Autonomous Systemspp. 3-16, 2004.

[13] L. Marques, A. Martins, and A. de Almeida,monitoring with mobile robotsConference on Intelligent Robots and Systems (IROS 2005)Edmonton, Canada, pp. 3624 -3629

[14] P. Pyk, S. Bermudez i Badia, U. Bernardet, P. Knusel, M. Carlsson, J. Gu, E. Chanie, B.S. Hansson, T.C. PearceVerschure, “An artificial moth: chemical source localization using a robot based neuronal model osearch”, Autonomous Robots, Vol. 20, pp.

[15] C. Stachniss, C. Plagemann, and A. Lilienthal,distribution models using sparse Gaussian process mixturesAutonomous Robots, Volume 26 ,2009.

[16] J.A. Farrell, S. Pang, and W. Li, Markov methods”, IEEE Transactions on Systems, Man and Cybernetics Part B: Cybernetics, Vol. 33, No.6

[17] S. Pang, and J.A. Farrell, “Chemical plume source lIEEE Transactions on Systems, Man and Cybernetics Part B: Cybernetics, Vol.36, No.5, pp.1068

[18] G. Ferri, M. V. Jakuba, E. CaselliYoerger, and P. Dario, “Localizing multiple gas/odor sources in an indoor environment using Bayesian occupancy grid mapping”, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2007), San Diego, USA, pp. 566

[19] M. Reggente, and A.J. Lilienthalfor gas distribution mapping in outdoor emobile robot”, Proceedings of IEEE Sensors,

[20] M. Reggente, and A. J. Lilienthalkernel DM+V/W algorithm for building gas distribution maps in uncontrolled environments”, Procedia Chemistry484, 2009.

[21] H. R. Olesen, P. Løfstrøm, R. Berkowicz, and M. Ketzel., “Regulatory odour model development: survey of modelling tand datasets with focus on building eNERI, Denmark, 2005.

[22] S. Thykier-Nielsen, S. Deme, and T. Mikkelsen,the atmospheric dispersion module RIMPUFFRODOS(WG2)-TN(98)-02, Risø National LaDenmark, 1999.

[23] A. J. Lilienthal, M. Reggente, M. Trincavelli, J. L. Blanco, and J. Gonzalez, “A statistical approach to gas distribution modelling with mobile robots – the kernel DM+V aInternational Conference on Intelligent Robots and Systems (IROS09) , pp. 570-576, 2009.

[24] V.Mattoli, A.Mondini, B. Mazzolai, G. Ferrichip Universal Smart Sensor Interface based on IEEE 1451 standard”, submitted to Sensors and Actuators

[25] IEEE Std 1451.1-1999, Standard for a Smart Transducer Interface for Sensors and Actuators - Processor (NCAP) Information Model, Institute of Electrical and Electronics Engineers, Inc., Piscataway, New Jersey 08855, June 25, 1999.

[26] IEEE Std 1451.3 -2003, IEEE Standard for a Smart Transducer Interface for Sensors and Actuators Transducer Electronic Data Sheet (TEDS) Formats for Distributed Multidrop Systems.

[27] E.W. Dijkstra, “A note on two problems in connexion with graphs”, Numerische Mathematik

[28] S. M. LaValle, Planning algorithmsCambridge, UK, 2006.

[29] I. Ulrich, and J. Borenstein, “VFH+: Reliable obstacle avoidance for fast mobile robots”, IEEE International Conference on Robotics and Automation, ICRA98,1577, May 1998.

[30] http://playerstage.sourceforge.net/[31] A. Gelb, Applied optimal estimation

Massachusetts, 1974.

and Actuators B, Vol. 49, pp. 52-57,

A.T. Hayes, A. Martinoli and R.M. Goodman, Distributed odor source localization, IEEE Sensors Journal, Vol.2, No.3, pp. 260-

Building gas concentration gridmaps Robotics and Autonomous Systems, Vol. 48,

and A. de Almeida, “Environmental monitoring with mobile robots”, 2005 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2005),

3629, 2005. P. Pyk, S. Bermudez i Badia, U. Bernardet, P. Knusel, M. Carlsson, J. Gu, E. Chanie, B.S. Hansson, T.C. Pearce, and P.F.M.J.

An artificial moth: chemical source localization using a robot based neuronal model of moth optomotor anemotactic

Vol. 20, pp.197-213, 2006. and A. Lilienthal, “Learning gas

distribution models using sparse Gaussian process mixtures”, Volume 26 , Issue 2-3 , pp. 187-202, April

and W. Li, “Plume mapping using hidden IEEE Transactions on Systems, Man and

, Vol. 33, No.6, pp. 850-863, 2003. Chemical plume source localization”,

IEEE Transactions on Systems, Man and Cybernetics Part B: 1068-1080, 2006.

E. Caselli, V. Mattoli, B. Mazzolai, D. R. , “Localizing multiple gas/odor sources in an

indoor environment using Bayesian occupancy grid mapping”, IEEE/RSJ International Conference on Intelligent Robots and

, San Diego, USA, pp. 566-571, 2007. J. Lilienthal, “ Using local wind information

for gas distribution mapping in outdoor environments with a Proceedings of IEEE Sensors, pp. 1715-1720, 2009.

J. Lilienthal, “Statistical evaluation of the kernel DM+V/W algorithm for building gas distribution maps in

Procedia Chemistry, Vol. 1, pp. 481-

m, R. Berkowicz, and M. Ketzel., odour model development: survey of modelling tools with focus on building effects”, Technical Report 541,

Nielsen, S. Deme, and T. Mikkelsen, “Description of odule RIMPUFF”, Technical Report

Risø National Laboratory, Roskilde,

A. J. Lilienthal, M. Reggente, M. Trincavelli, J. L. Blanco, and J. A statistical approach to gas distribution modelling

the kernel DM+V algorithm”, IEEE/RSJ International Conference on Intelligent Robots and Systems

B. Mazzolai, G. Ferri, and P. Dario, “Single chip Universal Smart Sensor Interface based on IEEE 1451

Sensors and Actuators A. 1999, Standard for a Smart Transducer Interface

Network Capable Application Processor (NCAP) Information Model, Institute of Electrical and Electronics Engineers, Inc., Piscataway, New Jersey 08855, June

2003, IEEE Standard for a Smart Transducer rs and Actuators -Digital Communication and

Transducer Electronic Data Sheet (TEDS) Formats for Distributed

A note on two problems in connexion with Numerische Mathematik, Vol. 1, pp.269–271, 1959.

Planning algorithms, Cambridge University Press,

J. Borenstein, “VFH+: Reliable obstacle avoidance IEEE International Conference on

Robotics and Automation, ICRA98, Leuven, Belgium, pp.1572-

http://playerstage.sourceforge.net/. stimation, MIT Press, Cambridge,