Using naıve physics for odor localization in a cluttered ... - Springer Link

Auton Robot (2006) 20:215–230 DOI 10.1007/s10514-006-7102-3

Using na¨ıve physics for odor localization in a cluttered indoor environment G. Kowadlo · R. A. Russell

Published online: 8 June 2006 C Springer Science + Business Media, LLC 2006


Abstract This paper describes current progress of a project, which uses na¨ıve physics to enable a robot to perform efficient odor localization. Odor localization is the problem of finding the source of an odor or other volatile chemical. Most localization methods require the robot to follow the odor plume along its entire length, which is time consuming and may be especially difficult in a cluttered environment. These drawbacks are significant in light of potential applications such as search and rescue operations in damaged buildings. In this project a map of the robot’s environment was used, together with a na¨ıve physics model of airflow, to predict the pattern of air movement. The robot then used the airflow pattern to reason about the probable location of the odor source. This approach, based on na¨ıve physics, has successfully located odor sources in a simplified environment. This demonstrates that na¨ıve physics can be used to assist odor localization operations and indicates that similar techniques have great potential for allowing a robot operating in an unstructured environment to reason about its surroundings. This paper presents details of the na¨ıve physical model of airflow, the reasoning system, the experimental equipment, and results of practical odor source localization experiments. Keywords Naive physics . Odor localization . Mobile robotics

G. Kowadlo (
) · R. A. Russell Intelligent Robotics Research Centre, Monash University, Clayton, VIC 3800, Australia e-mail: [email protected] R. A. Russell e-mail: [email protected]

1. Introduction The ability to locate the source of an odor/chemical plume has many valuable applications. These applications include finding the source of dangerous substances such as airborne biological material, hazardous chemicals, gas and other pollutants, in industrial and other settings; detecting such things as plant matter and drugs in a customs or quarantine application; searching for survivors in earthquake-damaged buildings, landslides or avalanches; detecting fire in its initial stages; locating unexploded mines and bombs; and for interrobot communication, particularly in robotic swarms. Sniffer dogs are currently used in many situations to locate an odor source. However, dogs are unable to operate in some hazardous environments. For example, an airborne substance may be radioactive, pathogenic, poisonous or flammable, any one of which could adversely affect animals. Furthermore, they can become fatigued and injured, or may not be able to move around in confined spaces. Communication is also a problem. A dog cannot tell its human handler which chemical it has identified and as sniffing is reflexive, it will continue to sniff, misleadingly, even if impaired. To overcome these limitations, Russell (2001), suggested building a robotic sniffer dog that is capable of odor localization. While research into robotic odor localization only began in the early 90s, there are several research groups currently working in the area. The problem has been tackled using various methods, categorized below, depending upon the nature of the environment. 1.1. Diffusion dominated chemical dispersion The first category is methods that are designed for robots that operate where fluid motion is dominated by viscosity. In such situations, diffusion produces smooth variations in chemical Springer

216

concentration, which decreases radially from the source. Locating an odor source under these conditions has been tackled using a number of techniques, mainly focusing on above ground operation in air, but also on underground operation. For above ground conditions, a method that was created by Lilienthal et al. (2001), specifically to deal with a lack of airflow uses a ‘linear search’. The robot performs a onedimensional search along a straight line and identifies the closest position to the odor source. The suggested localization strategy for a two dimensional case, is to then move to the odor source position identified for the one dimension, turn at right angles and repeat the process. A ‘chemical gradient’ method has been used by several researchers (Rozas et al., 1991; Genovese et al., 1992; Buscemi et al., 1994). This technique involves a measurement of the chemical concentration at two spatially separated positions (using two chemical sensors simultaneously, or one sensor used twice), followed by the calculation of a chemical gradient. The robot then uses this information to move in the direction of increasing chemical concentration. ‘Swarms’ too have been used in this category (Buscemi et al., 1994). Each robot uses similar techniques, but together, the swarm has the advantage of redundancy and large coverage. The methods described assume that diffusion is the dominant short-term method of odor dispersal. For above ground robots, this assumption may lead to unreliable performance, as such conditions will only be encountered by extremely small robots (of a similar size to bacteria). However, these conditions hold, and are exploited by Russell (2004a, 2004b) to locate a chemical source underground. Russell uses a 3 dimensional search strategy to move a buried probe “following a path through the ground that corresponds to the edges of closely packed dodecahedra”, towards the point of highest concentration, which corresponds to the location of the odor source. 1.2. Turbulent fluid flow The second category is methods that are designed for robots that operate in the presence of a background fluid (air/water) flow in an open area free of obstacles. In this case, odor dispersal occurs through carriage by the fluid flow. A plume is formed downwind of the odor source, and turbulence causes the plume to spread out, become patchy, and meander. Within this category, Grasso et al. (2000) have used a standard ‘chemical gradient’ method, as described above for the category ‘Diffusion Dominated Chemical Dispersion’. The robot, a biomimetic lobster, measures the chemical concentration at two spatially separated positions, a chemical concentration gradient is calculated, and the robot lobster then moves in the direction of increasing chemical concentration. It is effectively a plume-edge following technique, which can be successful if the robot is facing in the correct diSpringer

Auton Robot (2006) 20:215–230

rection and in close proximity to the source, where the plume edge is well defined. Other researchers have used a ‘combination of plume acquisition and following’. They use absolute readings from wind direction and chemical sensors (Ishida et al., 1995; Ishida, 1996; Russell et al., 1995; Hayes et al., 2001), as well as enhancements by considering the transient behavior of the chemical sensor (Ishida et al., 2002). Another technique, that can be combined with the previous, and is a variation on the chemical gradient method, uses ‘directional chemical sensors’ to give directional information about the odor source. Within this sub category, both active and passive sensors have been used. Two reported active sensors are a stereo nose (fan with two nostrils) with a recurrent neural network (Duckett et al., 2001) that has had limited success; and a chemical sensor based on the silkworm moth that uses forced ventilation (Ishida et al., 1996). A passive sensor used by Ishida et al. (1995) Ishida (1996) and first described in Ishida (1994), was a probe that combines four anemometric and four odor sensors. Kazadi et al. (2000) reported a passive chemical sensor made from simple resistive polymer that is affected by flow direction. Finally, the use of ‘swarms’ has been investigated by Hayes et al. (2001). Each robot uses a version of plume acquisition and upwind following, and together they display increased efficiency. Like the swarms described in the previous section, they have the advantages of redundancy and large area coverage. Within this category is a method for controlling robots that operate in a corridor environment with airflow. In this environment there is little chemical gradient information unless the corridor walls absorb the airborne chemical. A method using upwind searching and knowledge of gross fluid dynamics effects in simple interconnected tunnels (Russell et al., 2000; Russell, 2001) has been shown to be successful. With the exception of the previous example, these methods largely focus on basing robot sensing and algorithms on the odor localizing behavior of some microbes, insects, and crustaceans, aiming for simplicity using reactive control schemes and local sensing. The main limitations are that robots must follow the plume along its entire length, which is time consuming and may not be possible. In addition, the effects of obstructing objects and walls on robot mobility are often neglected. Recently researchers have attempted to overcome these limitations in two ways. Firstly, by developing systems that go beyond purely reactive control, planning movements based on a model of the odor plume (Ishida et al., 1997; Farrell et al., 2003); and secondly, by exploiting more information from the environment additional to local wind and chemical readings (Ishida et al., 2004; Loutfi et al., 2004). Ishida modeled the plume by fitting a chemical distribution model to the sensor dynamics (Ishida et al., 1997). This is reliable for source detection. However, the robot must continually re-calculate and alter its heading, making it partially

Auton Robot (2006) 20:215–230

217

Fig. 1 Experimentally measured airflow (left) and odor concentration (right) averages. The odor source is shown with an arrow. See ‘Experimental Analysis’ for details

reactive, and the robot must travel to within 20 cm of the source. Farrell has used a Hidden Markov Model to map the chemical plume (Farrell et al., 2003). This approach appears to be effective, however it has not considered objects, walls, or been implemented on real robots. Ishida has incorporated color vision (Ishida et al., 2004), and Loutfi has used color vision and sonar (Loutfi et al., 2004). These sensors are used to infer information about the surroundings, which is used in conjunction with local sensing of airflow and chemical concentration. They form the inputs to behavior-based architectures. These methods effectively exploit more information than previous methods, enabling the robots to identify candidates thus dramatically diminishing the effective search space and greatly enhancing the ability to locate an odor source. However, they may break down in more complicated environments where fluid flow needs to be known. All of the reviewed methods would be unreliable in an indoor cluttered environment with low ceiling and thinly populated by objects that affect airflow; the type of enclosed space that may be encountered in a cave, air duct, sewer or crawl-way beneath a house. In this situation, distinct sectors of airflow form. The air and odor particles circulating within a sector are predominantly restricted to that sector. There is a small amount of mixing across boundaries due to turbulence, however in the case of chemical concentration the discrepancy is significant and easily discernable with a standard chemical sensor. As a result, odor mixes throughout the sector in which it was released, possibly with a local maximum distal from the source, and does not form a welldefined plume that can be easily followed reactively. This is demonstrated in Fig. 1, which shows the airflow and resulting chemical distribution in a typical environment for experiments reported in this paper. The details regarding these data are given in ‘4 Experimental Analysis’. Many applications of odor localization will require the ability to operate in these conditions, however to the authors’ knowledge they have not been tackled explicitly. We have implemented a solution for these environments that addresses

the aforementioned limitations by implementing a ‘sensemap-plan-act’ control scheme; drawing on the type of control that may be used by many higher-level animals, such as a human or dog, for similar sensor guided tasks. Information about the surrounding environment, not just local sensing, is used to model the airflow, creating an airflow map. Using this map, odor plumes detected at the robot are ‘projected back’ into the environment to identify probable locations for the source. A Reasoning System uses knowledge of the airflow from the airflow model to direct the robot to a set of target positions where it gathers information. The information is used to predict which objects are the likely odor sources. This method was first reported in Kowadlo and Russell (2003).

2. Airflow model The airflow-modeling algorithm must be capable of producing a map of airflow that captures the broad features. It must divide the room into sectors of airflow including information of wind direction at the least, and if possible, also give information on the velocity of the flow, however this is not essential. A conventional finite element model of fluid flow would be difficult to apply, because many boundary conditions are unknown, it would be very slow, the results would be more detailed than required, and in a non-intuitive form that makes it difficult to extract simple properties of the behavior. These problems are overcome by using a na¨ıve physics model of fluid flow to provide approximate information about patterns of airflow. Na¨ıve physics is essentially the use of common sense knowledge and physical intuition to model the environment, as opposed to the traditional approach of using differential equations. Generally speaking, this is a qualitative and not a quantitative approach. Take for example, the scenario of pouring liquid into a glass. We know what is going to occur to a level of detail that is sufficient for us to perform Springer

218

Springer

observe physical process

evaluate performance

make modifications

the operation effectively. We use na¨ıve physics—common sense and physical intuition. Predicting what will happen using classical physics would involve solving a series of nonlinear partial differential equations, and would have all the aforementioned limitations associated with a finite element analysis. The contemporary field of na¨ıve physics began with Hayes’ manifesto (Hayes, 1979; Hayes, 1985), which proposed a system for the large-scale formalization of common sense knowledge. This triggered other research under the banner of na¨ıve physics. One significant area that has emerged is termed ‘qualitative physics’ (Bobrow and Hayes, 1984; Weld and De Kleer, 1989; De Kleer and Williams, 1991; Faltings and Struss, 1992; Kuipers, 1994). In general this refers to dividing a variable into physically significant ranges and performing a type of qualitative mathematics to predict the behavior of a system. The representational methods of qualitative and classical physics were criticized by Gardin and Meltzer (1989). They argued that much richer structures are utilized for commonsense thinking usually “based on our perceptions of form and movement in the physical world”. They proposed a new type of ‘analogical representation’ to overcome this limitation, and demonstrated its usefulness with two simulations. In 1997 Davis published his ‘Na¨ıve Physics Perplex’, which proposed a new methodology for large-scale formalizations to rival Hayes’ proposal. Following on from Hayes’ vision, much of the effort has been intentionally spent on the theoretical aspects of formalization, and not on implementation (most notably a collection of papers that addressed a problem of cracking an egg, proposed by Davis in the formerly active web site ‘Common Sense 98 Problem Page’ (Davis, 1997)). These have concentrated on developing logical descriptions of objects and their relationships. To the author’s knowledge, there are no reports of a working na¨ıve physics model that has implemented these logical expressions with an algorithm for a practical application. We have developed a procedure for using such expressions, in our case in the form of na¨ıve rules, to create an executable algorithm. The procedure is illustrated in Fig. 2. The iterations are a cycle of preliminary testing of the rules for efficacy, and further intuitively based revisions. NaReM’s can model phenomena without knowledge of the mathematical relations; they require a form of supervised learning, and have the ability to generalize. Before an algorithm could be devised, the process of indoor airflows under these conditions needed to be understood with ‘ground truth’ examples of airflow. This was achieved by conducting fluid dynamics simulations. The validity of the simulations was verified with thorough experimental measurements for key examples, and by continually comparing with measured airflows during experiments.

Auton Robot (2006) 20:215–230

derive naive rules to describe physical process create algorithm that encapsulates the rules implement the algorithm test on new cases

no

is the model satisfactory? yes

NaReM (Naive Reasoning Machine)

Fig. 2 Creating a Na¨ıve Reasoning Machine

The simulations and subsequent algorithm development are limited to a constrained environment. This is defined as an area enclosed by walls and containing objects, both of which affect airflow that is introduced into and exhausted from the area via a number of openings. The environment is a predominantly rectangular room with one inlet, one outlet, up to four objects, and an area in the order of 5 m2 . The inlet and outlet are located on opposing walls. The airflow source blows air perpendicular to the wall in which it was mounted at a low velocity (approximately 0.5 m/s), and does not produce highly turbulent airflow. These airflow specifications are designed to match conditions in a number of indoor environments. For example, in crawl-ways under floors, in roof cavities and in tunnel systems. 2.1. Na¨ıve physics airflow model The algorithm essentially works by applying the intuitive na¨ıve rules displayed in Table 1, which were devised using the simulation results and a process of iterative refinement (illustrated in Fig. 2). The rules are briefly compared to the conventional method of computational fluid dynamics (CFD) based on the Navier-Stokes equations, illustrating the physical basis of the na¨ıve rules. Note that this process is not a direct contribution to a general large-scale formalization of commonsense knowledge as envisioned by Hayes (1979, 1985) and Davis (1997). In fact Hayes rejected attempts to solve practical problems using na¨ıve physics because of the unrealistic and potentially misleading simplifications that are inevitably made. He reasoned that these attempts would not ultimately contribute to a system that has potential for real applications. We feel that for this problem, it is possible to use the concepts of applying intuitive thinking to realize practical applications now. This is not intended solely as a means to

Auton Robot (2006) 20:215–230 Table 1

219

Na¨ıve rules

1 Air movement continues unless impeded 2 If impeded, the air stream bifurcates, and travels parallel to the obstructing object in both directions 3 Airflows themselves act as obstacles to other airflows. However, air travels parallel to the obstructing airflow only in the same direction as this airflow 4 Parallel airflows are conducted in a perpendicular direction (they spread out) 5 Air cannot be created or destroyed • This rule is loosely applied. It does not need to hold locally in the final airflow map. However, it must be observed when adding airflow arrows 6 Real airflows are affected by solid boundaries— there is a reflection effect due to rules 2 and 4 7 Opposing airflows are pseudo additive. Arrows of opposing directions cancel each other, and aligned arrows are retained 8 Objects create airflow shadows, i.e. they block the air leaving an area of low pressure behind the side that faces the wind

an end, however, we hope it is a study that can be built upon to further the field of na¨ıve physics. Furthermore, the methodology used excludes it from qualitative physics, as it does not create an airflow model by dividing the parameters into physically significant ranges, and then applying a type of computational fluid dynamics calculations. Qualitative physics is an area of na¨ıve physics that has seen attempts at modeling physical systems. In order to compare these rules to conventional CFD, a brief explanation is necessary. See (Alexandrou, 2001) for more detail. There are two main approaches to fluid dynamics: ‘finite control volume’ and ‘differential’. The former is concerned with average (bulk) behavior. The latter models the behavior at an infinitesimal dimensionless stationary point in the fluid (as opposed to classical mechanics, which looks at the movement of a particle or particles). This approach is based on the continuum assumption, which states that the conditions at a point are the local average of the molecules around that point. It allows a wider range and more detailed analysis of movement. Central to this approach is the concept of the stress tensor. The stress tensor, σ , represents the forces on this infinitesimal point. The stress tensor (simplified to two dimensions) along with its components is illustrated in Fig. 3. The Navier-Stokes equations express the movement of a fluid in differential form. They are used for CFD. The equations expressed in vector form in Cartesian coordinates for isothermal (constant temperature) laminar flow are:

σxx σyx

σyy

Stress tensor

where, u = velocity vector; ρ = the density of the fluid; g = acceleration due to gravity. Table 2 shows a comparison between the na¨ıve rules (described above) and conventional CFD, illustrating the physical basis of the na¨ıve rules. These rules were formulated into an algorithm, which is divided into two sections: a first and second pass. It was designed to operate on a rectangular room, with an inlet and outlet positioned on opposite walls at x = 0 and x = max. The room is divided into a grid with rows and columns in the y and x directions respectively. Each grid point has a type and a direction. The types can be one of: WALL Boundary to room VWALL Wind, which acts like a virtual wall. This has a direction SOLID An object EMPTY Nothing The direction, which applies only to type VWALL, can be horizontal (left or right), or vertical (up or down). τyy

=

P

τyx

P

P Hydrostatic Pressure

+

τxx

τxy τxx

τxy τyx

[

σxx σxy σ= σ σ yx yy

[

σxy

σxy

Conservation of momentum 
∂u i + u.∇u i = ρgi + (∇.σ )i ρ ∂t

P

σyx σxx

∇.u = 0,

[

σyy

Conservation of mass (continuity)

τyy

Viscous stresses

Fig. 3 Stress tensor Springer

220 Table 2

Auton Robot (2006) 20:215–230 Comparison between na¨ıve rules and conventional CFD

Na¨ıve rule

Conventional CFD—Physical basis

1 2

Conservation of momentum: There is no significant net force, so the fluid experiences no acceleration Conservation of momentum: Forces exerted by the obstruction (at the point of incidence) and the bulk fluid (hydrostatic force) result in flow along the obstruction Conservation of momentum: Viscous stress dominates, resulting in a net force in the direction of the airflow that is acting as an obstruction Conservation of momentum: Viscous stress causes adjacent air to experience a force in the same direction Conservation of mass This rule is based on rules 2 and 4 This is a fundamental concept in physics. Implicit in the conservation of momentum Eq. (2) This is a result rather than a rule. It is due to the conservation of mass and momentum

3 4 5 6 7 8

r Airflow is not continued on the other side of the object

The algorithm is briefly described below to give an overview of its operation. The instructions are listed, followed by a number indicating the na¨ıve rule on which they are based. The process is illustrated in Fig. 4. A detailed flow chart description of the algorithm is shown in Figs. 5 and 6. Within the flowcharts, sub processes are contained in rectangular boxes with double lines on each side. A selection of the sub processes are then expanded in subsequent flow charts or written explanations. The reference coordinates are shown in Fig. 4. First Pass

[Rule 8]. These steps are defined as a scan. The first pass begins with a scan, with the ‘scan direction’ set as from the inlet into the room. Then,

r Moving perpendicularly away from the initial scan, perform subsequent scans every grid spacing. Therefore, parallel scans are performed above and below the initial scan until the walls have been reached [Rule 4]. Each scan must initially extend the airflows that are perpendicular to the current scan (hence moving towards or away from the current scan) [Rule 1].

r Draw an arrow starting from the inlet into the room (setting the airflow at each of the covered grid points to this direction) [Rule 1]. r If an obstacle is encountered, draw an airflow arrow, the size of the grid spacing, parallel to the obstacle in the ‘sweep direction’. The ‘sweep direction’ is defined as perpendicularly away from this initial scan. This is the direction of subsequent scans, as explained further following this ‘First Pass’ description [Rule 2].

Second Pass

r Reflection from the walls: The same process is repeated, however the scan direction is reversed, and scans begin at the level of the walls and move towards the initial scan level of the input [Rule 6]. r Superimpose the initial and reflection scans: If they are of equal length, then they cancel out, otherwise the stronger airflow swamps the weaker airflow [Rule 7].

r Encountering an obstacle applies to the head and the tail of the airflow arrow. I.e. air must come from somewhere and go somewhere. Arrows are always added from head to tail [Rule 5]. r Obstacles can be a Wall, Virtual Wall [Rule 3] or Object.

An example of the results of the airflow modeling is shown in Fig. 12. The arrows show that the algorithm has predicted sectors similar to those measured in Fig. 1.

initial scan

secondary scan

Obstacle

tertiary scan x

y

Springer

θ

‘Sweep directions’ for scans on their respective sides of the initial scan.

Fig. 4 Examples of the airflow modeling algorithm procedure

Auton Robot (2006) 20:215–230

221

Fig. 5 Flowchart of airflow modeling algorithm—‘Model Air’

start

Set the scan level at inlet height 'initial level'.

Scan to boundary

start Set scan direction as inlet to outlet . conduct scan Scan to boundary: Perform scans from 'initial level' in direction of increasing y to the boundary. Store the results in Room 0

increment/ decrement scan position

Scan to boundary: Perform scans from 'initial level' in direction of decreasing y to the boundary. Store the results in Room 0

Reverse the scan direction.

no reached boundary/ 'initial level'?

yes

Scan to boundary: Perform scans from maximum y boundary to 'initial level'. (Reflection from boundary) Store the results in Room 1 end

Scan to boundary: Perform scans from minimum y boundary to 'initial level'. (Reflection from boundary) Store the results in Room 1

Superimpose Room 0 and 1

end

3. Reasoning system for source location prediction The Reasoning System is a ‘Sense-Map-Plan-Act’ control scheme that emulates common sense style behavior. It is comprised of four phases, of which phases 1 and 4 are optional, and phase 2 and 3, the heart of the system, have been implemented. The system is shown conceptually in Fig. 7. The first phase is the construction of a map of the environment. In many applications, a map is available, and in principle, there are a number of sensors that could be used in the mapping process, including laser rangefinders and

stereo-vision systems. Therefore, it was decided to provide the system with an a priori map. The second phase consists of modeling airflow in the environment; the third, the use of the model of airflow to predict the odor source—forming the Reasoning Algorithm, which will be described in detail; and the fourth, the verification of the odor source. The first phase will involve hypothesis testing using local sensing to confirm which openings are inlets and outlets, making it possible to operate without this type of prior knowledge, which may not be available. The fourth phase will involve the robot moving to and circumnavigating the predicted odor source. Due to Springer

222

Auton Robot (2006) 20:215–230

sense

start

map

Extend Last Scan: Extend the vertical vectors from previous scan.

obtain range data

generate a map of - the room - the airflow in the room

Phase 1 (an a priori map of the room can be provided.) Phase 2

generate a list of positions for information gathering for source prediction

Get Start & End: Calculate the start and end positions, xi and xf of the horizontal flow for this scan

plan simple heuristic search for the next position. Phase 3 Reasoning Algorithm

yes

xi or xf undefined?

- move to the next position - perform information gathering.

no

act Modify xi and xf to a grid spacing in from boundaries.

make prediction

no verify prediction

Phase 4

yes

do xi and xf overlap?

Fig. 7 Reasoning System

no

complex odor dispersal in this environment, verification will be based on both olfactory and visual sensing. In the event of negative verification, the next most probable odor source would be investigated.

object type at xi?

WALL, VWALL or SOLID

OUTLET

Add Arrow: One grid spacing long, perpendicular to scan, in direction opposite to 'sweep direction'

Undo last change to xi

OTHER

xf type ?

WALL, VWALL or SOLID

OUTLET

Add Arrow: One grid spacing long, perpendicular to scan, in 'sweep direction'

OTHER

Undo last change to xf

Add Arrow: Horizontal from xi to xf

end

Fig. 6 Flowchart of airflow modeling algorithm—Sub section ‘Scan’

Springer

3.1. Reasoning algorithm The reasoning algorithm (RA) uses the model of airflow to generate a list of positions that will provide information for prediction of the odor source location. It is in essence a cyclic search process. A simple heuristic is applied to determine the next position to move towards. The robot moves to this position where information is gathered, and the process is repeated. Information gathering consists of a further senseact sequence. The RA is described here as a written overview and a more detailed flowchart in Fig. 8. The RA must use the airflow model to firstly identify potential odor source candidates (one of several objects). It then predicts the potential odor paths from these candidates by tracing downstream from all sides of each candidate. This constitutes the list of positions to move towards. Each trace must be investigated, but any position on the trace is suitable (from 80 cm downstream of the object, which encourages collision free robot movements). The algorithm is tolerant to different choices of downstream trace length, but this will affect the resulting robot trajectories (robot movements are explained in this section). Due to the formation of circulating sectors, the trace will also circulate. Therefore if the trace is long enough, it will simply loop back on itself, and continues

Auton Robot (2006) 20:215–230

223

start

Project downstream odor trace from objects.

no

Detect odor ? yes Conduct upstream odor trace:

Record chemical concentration, and associate with every object in the vicinity of the upstream trace.

Find Scnet: Move towards the closest position on the closest predicted downwind odor trace, beginning from 80cm downstream of the object from which it originated.

yes

Have all predicted path been investigated? no Predicted source = the candidate with the highest associated concentration.

end

Fig. 8 Flowchart of Reasoning Algorithm—‘Top Level’

to accurately describes a potential region for the presence of odor particles. The algorithm then searches for the next target position, with the aim of minimizing the total distance that must be traveled to explore all the traces. This is a variant of the ‘traveling salesman’ problem, with the added complexity that the target positions are not localized, being distributed linearly along a trace. It is solved effectively (though nonoptimally) by applying the heuristic: find the closest position on a trace, out of all positions on all traces that have not yet been investigated. The robot then moves towards the target positions until every trace has been visited.

At each step the robot performs an ‘information gathering’ procedure as part of a ‘sense-act’ sequence. The chemical concentration is measured, and odor is considered to be significant if above a concentration threshold of 1.5 PPM, as derived from observation of measured values, see ‘4.2.2 Odor Concentration’. If significant, it conducts a virtual trace upstream to identify which candidate(s) were responsible. If no objects are detected, then the robot continues moving to the next target position. If one or more objects are detected, then they are the predicted odor source(s), the chemical concentration is recorded, and the robot continues to the next target position. After investigating every downwind trace, a list of predicted sources is presented, ordered by the chemical concentration that was measured for that source. The first source is the primary prediction. The purpose of the ‘downstream trace’ is to remotely test the candidates. However, once an odor is detected, it is necessary to ascertain if there are other possible candidates sourcing this position. The upwind trace determines other possible sources, as it conducts a more complete search (compared to the downstream trace) for other objects in the vicinity. This includes other traces, which are considered to be a source due to turbulent mixing. If a trace is detected in the vicinity, then the object from which it originated is considered to be the possible source. The upwind trace must also be used in the case of unpredicted odor detection (when not at a targeted position calculated as downwind from an object). The division is necessary as the downwind tracing algorithm is conducted from every object. Therefore, it must be a quick method of determining target positions. However, the upwind trace is the step after odor detection. It is conducted in one place, and must be a more efficient method of determining the possible odor sources for this position. Note that detecting a chemical is accomplished using absolute sensor readings. In future work the dynamic response will be incorporated in a similar fashion to Ishida et al. (2002). Move robot This section of the algorithm takes two parameters, direction and distance. It simply commands the robot to rotate to the target direction, and move the desired distance. If, during a move, an obstacle is detected, the robot will move backwards to clear the obstacle, rotate 90◦ and move forwards by 10 cm. This reactive movement is effective at bypassing objects and avoids complicated path planning. 4. Experimental analysis An enclosed environment was created and the following experiments conducted: The airflow and odor concentration

Springer

224

conditions were investigated, a conventional reactive odor localizing robot was tested, and the odor localization system reported in this paper was tested. The environment and experiments are explained in this section. 4.1. Enclosed environment An enclosed space with low ceiling was constructed to develop and test the robot system in an environment resembling a cave, crawl-way, air duct or tree canopy. The room had a floor area of 2820 × 1900 mm. A single layer of reconfigurable boxes 270 × 270 × 470 mm was used to build the walls and clear plastic, attached to a metal frame, formed the ceiling. Up to four boxes were placed in the room as potential odor sources. They had a height equal to the ceiling, and were large enough to disturb the airflow. One of the objects was an odor source, which the robot was required to locate. An air source introduced airflow into the room at a flow rate of up to 0.5 m/s. The air source consisted of six DC fans that forced air through an array of collimating straws (205 mm in length and 4.5 mm in diameter) to minimize the high levels of turbulence created by a large fan. The room and air source are shown in Fig. 9. The odor source was created by emitting ethanol vapor from one side of the box at a flow rate of 0.5 ml/s. The vapor was generated by bubbling air through a flask of methylated spirits, causing it to become fully saturated. 4.2. Investigating the environmental conditions The airflow characteristics and odor dispersal in this environment were investigated. Understanding of airflow patterns was required for the development of the airflow model (‘2 Airflow Model’), and the odor dispersal for the reasoning system (‘3 Reasoning System for Source Location’). An example of the environmental data gathered is shown in Fig. 1 and the experimental procedure is explained in this section.

Fig. 9 Picture of model room (left) and air source (right) Springer

Auton Robot (2006) 20:215–230

4.2.1. Airflow Airflow in this environment was simulated, and tested experimentally. Several key configurations displaying different features were compared to verify the validity of using the fluid dynamics simulation results to predict the airflow. In all test cases, the broad features were in correspondence. Fluid dynamics simulations. Fluid dynamics simulations were conducted with Flo++TM . The calculations were conducted for room temperature, 300 K. Laminar, steady state flow was assumed and air was modeled as having a constant density of 1.0 kg/m3 and a constant viscosity of 1.8 × 10−5 Ns/m2 . The constant density results in incompressible flow, decreasing solution time significantly. It is an accurate assumption due to the low velocity of the flow. The walls were defined as adiabatic as convective currents are not significant in these scenarios with forced ventilation. The environment is represented as 2D by defining a room with negligible height and with symmetrical boundary conditions between floor and ceiling. The calculations were terminated when the solution had converged sufficiently. Simulations were conducted for several room configurations. They were every combination of: three inlet positions, three outlet positions, zero to one object in one of nine segments (the room was divided into nine segments with equally spaced rows and columns). The room was divided into sectors in every scenario. A sector was considered to be a distinct region of circulating airflow—a broad feature that was extracted from the detailed information afforded by the simulation. An example of a simulated environment that corresponds to actual measurements is shown in Fig. 1. Experimental. The flow direction was measured by visual inspection of (a) cotton strands attached to the ceiling, and (b) a wind vane on the robot at regularly spaced positions within the environment. It became difficult to distinguish

Auton Robot (2006) 20:215–230

225

C B

F G

A H D

E

A: 15 minutes B: 20 minutes C: 10 minutes D: 9 minutes E: 5 minutes

= Odor source = Start Position

F: 12 minutes G: 5 minutes H: 10 minutes

= Odor source = Start Position

Fig. 10 Reactive odor localization

direction at low flow velocities. Flow directions were plotted, and used to draw the general airflow characteristics. Nothing is drawn where the flow velocity was too low to measure accurately. This was carried out for a range of scenarios covering those reported in this paper. Sectors that corresponded to fluid dynamics simulations were observed. An illustrative example is shown in Fig. 1. 4.2.2. Odor concentration Odor dispersal was measured by driving the robot around the environment, and taking chemical concentration measurements at regular locations. A measurement consisted of waiting 10 seconds for the sensor response to settle, and sampling the A to D converter (which measures the sensor voltage, see Eq. (2) 10 times with a pause of 0.25 seconds between readings. The average reading was deemed valid if the standard deviation was less than 2.5% of the average (it was almost always less than 0.5% of the average). The robot was driven to previously visited positions, where further measurements were taken in order to confirm that the odor concentration pattern was on average, steady. The results are shown for the same example used for airflow, and in the same figure, Fig. 1. For this experiment, the odor was measured at 44 positions distributed throughout the environment. Within the centre (approximately) of the airflow sector containing the odor source, three positions distributed equally along the length of the room were revisited with up to 10 min separation. At most the concentration, in PPM, fluctuated by a maximum of 20%. Furthermore, the delineation between sectors remained clear, with the fluctuations minimal within the area of low concentration, and higher at high concentrations, resulting in a relatively steady state pattern, with consistently high concentrations above 1.5 PPM. The contour plot of the data was drawn using MinitabTM . Through interpolation, this representation gives an indication

of the odor distribution. It shows clearly that odor is restricted to the sector in which it is released. 4.3. Reactive odor localization Previously reported reactive odor localization strategies were tested in this indoor environment. The robot described in Russell et al. (2003) was used, but with tin-oxide, rather than polypyrrole chemical sensors. In that study, the robot was used to implement the chemotaxis behavior of the Dung beetle, Silkworm moth, and a gradient based method. These and other algorithms involve large movements, and would therefore pose problems within this constrained environment. In addition, the Dung beetle and similar methods require movement to the border of the plume, which would also create problems in this case, as the odor distribution is more spread. Therefore, these algorithms are not feasible without severe modification. For the purposes of this study, the ‘Step-byStep’ algorithm using chemical gradient and airflow direction (Ishida, 1994) has been implemented. “The essence of this algorithm is that the robot selects a heading in the upwind direction as measured by the anemometric sensors and toward the centre of the odor plume as indicated by the gradient reading from the gas sensors” (Russell, 2001). Before indoor localization was attempted, this implementation was verified by testing it successfully over 10 trials in an open area with an odor source comprised of a lid full of ethanol, placed in front of the air source described in ‘4.1 Enclosed Environment’. For the indoor environment, sixteen trials were performed, of which eight typical results are shown in Fig. 10. In general the method can be successful as it constantly moves upwind, which in these simple environments, brings it to within the vicinity of the odor source. However, in most cases as typified by trials A and F, this involves a long route, which is very time-consuming, in this case 15 and 12 minutes respectively. The robot can also be successful in a relatively short amount of time, as shown with trial G (5 min). However, in areas where there is no strong wind direction (which often Springer

226

Fig. 11 A photograph of Roma, the odor source-locating robot

occurs in these types of environments), the robot is lost, and doesn’t move more than a few centimeters over a duration of 10–20 min, at which point the experiment was terminated (trials B and C). Additionally, due to low odor concentration gradients combined with strong airflow, or gradients that are directed away from the source, the robot can take an unsuccessful trajectory, as typified in trials D and H. The robot was unable to locate the source when released into the lower half of the room (corresponding to the identified airflow sector), as shown by trial E. In all successful trials, this method was very time consuming (more than 10 min); with the exception of one trial (which took 5 min). 4.4. Roma the robot—Construction and interfacing The robot used for locating the source of an odor and investigating the airflow and odor conditions is named Roma. Roma is a circular mobile two-wheeled robot with a 24 cm

Fig. 12 Screenshot of GUI Springer

Auton Robot (2006) 20:215–230

diameter base and approximately 15 cm high. It can be commanded to move forwards, backwards, to turn on the spot both clockwise and anticlockwise, and to report sensor readings (Fig. 11). Roma contains sensors, dc motors and gearboxes, support circuitry and a microcontroller. It is connected to a PC and a power source via a tether. The robot drive system comprises a passive castor and two wheels mounted along the diameter, each with an independent gearbox and motor for differential steering. This arrangement allows the robot to rotate on the spot about its centre. The sensors include a wind vane, a chemical sensor, bump sensors, and wheel encoders. The wind vane can be seen mounted on the top of the robot in Fig. 11. It senses the direction of airflow indicating one of eight 22.5◦ ranges. Optical encoder sensors are used for feedback of wheel rotation. The linear resolution of movement is 7.5 mm, and the angular resolution of turning 4.7◦ . A bump sensor ring surrounding the robot detects collisions. Chemical concentration is measured using a Figaro TGS2600 tin oxide chemical sensor. The sensor resistance RS is dependant on the concentration of reducing gasses such as methanol, ethanol and ammonia (and also varies with temperature and humidity). Over a limited range of concentrations, this relationship can be approximated by Watson (1984) R S = KR O C α

(1)

where α is the sensitivity, R is the sensor resistance on exposure to the target chemical, Ro is the sensor resistance in

Auton Robot (2006) 20:215–230

227

Start position. Odour detected. Conducted upwind trace and predicted object B as the odour source.

Trial A

Trial D

Trial C

Object A

Moving to predicted downwind trace from object A. Moving to predicted downwind trace from object B. Reached trace, move to other predicted downwind traces that have not been investigated.

Trial B

Object B Odour Source

Downwind trace: Predicted downwind odour path. Trial A, B, C, D Predicted Source: Object B Time taken: 40s, 30s, 35s, 32s

(a) Start position. Odour detected. Conducted upwind trace and predicted object B as the odour source.

Trial A

Object A

Moving to predicted downwind trace from object A.

Trial B

Moving to predicted downwind trace from object B.

Trial C

Reached trace, move to other predicted downwind traces that have not been investigated. Downwind trace: Predicted downwind odour path.

Object B Odour Source

Trial A, B, C Predicted Source: Object B Time taken: 30s, 37s, 33s

(b) Fig. 13 Experimental results (continued on next page)

clean air, K is the scaling constant, and C is the concentration. For these experiments, with the TGS2600 and ethanol, this relation holds between 1 to 100 PPM, with K = 0.67, and α = − 0.31. The sensor is used in a voltage divider arrangement, with a load resistor RL , and circuit voltage VC , so that the measured voltage, VRL is: 
RL · Vc (2) VR L = (R L + R S ) Roma’s C167 microcontroller is connected via a serial RS232 link to a PC. This microcontroller performs robot control which includes: controlling the robot’s movement, acquiring data from the sensors, receiving commands from

the Reasoning System that resides on the PC, and responding to those commands by either moving or sending data back to the PC. The control system for wheel movement was essentially a feedback proportional controller. A graphical user interface was constructed using QT from TrolltechTM to facilitate development and testing of the robot. A screen shot of the user interface is shown in the Fig. 12. The screen is divided into 4 columns. The contents of each column are described below from left to right. Column 1: Graphical display of room A graphical display of the room map including: Roma, airflows and predicted odor traces downwind from objects. Springer

228

Auton Robot (2006) 20:215–230

Start position. Conducted upwind trace and predicted object A as the odour source, for every step until cease to detect odour. Stopped detecting odour. Moving to predicted downwind traces from object A.

Trial B

Moving to predicted downwind trace from object B and C. Downwind trace: Predicted downwind odour path. Trial A, B Predicted Source: Object A Time taken: 1m 56s (116s), 1m 52s (112s)

(c) Start position. Trial B

Conducted upwind trace and predicted object C or D, as the odour source, for every step until cease to detect odour. Stopped detecting odour. Moving to predicted downwind traces from object A and B.

Trial A

Moving to predicted downwind trace from object C and D. Downwind trace: Predicted downwind odour path. Trial A, B Predicted Source: Object C or D Time taken: 65s, 73s

(d) Fig. 13 (Continued)

The arrows indicate direction of airflow. The bold section in the centre of the left wall is the inlet, and the corresponding symbol on the right is the outlet. The two large squares in the room area are obstacles, and the lines extending from the centre of all sides show the predicted downwind odor trace. The smaller square, with a small line extending out of one side, is the robot. The line represents the direction in which the robot is facing. Column 2: User control of mini world (predominantly) Column 3: Robot status/interface Column 4: Data log of reasoning algorithm steps.

Springer

4.5. System evaluation The robot was tested in a room with one inlet, one outlet and up to four objects (one of which was an odor source) over 18 trials. The configurations included variations in: the positions for the inlet and outlet; the position and orientation of objects; the shape and size of objects; and the robot (initial) position. The odor source was switched on prior to activation of the robot, in order to allow the chemical to circulate around the room and reach a relatively steady state pattern. Eleven of the trials, covering the range of scenarios and robot movements, are shown in Fig. 13(a–d). Each

Auton Robot (2006) 20:215–230

figure shows a room configuration in terms of inlet/outlet and box position, and the robot’s trajectory for multiple trials within that configuration. The inlet/outlets are shown as narrow rectangles with an arrow to indicate airflow direction, and the predicted downwind odor traces from the odor source candidates are shown with a dotted/dashed line (see legend). For each trial, the robot’s initial position is indicated with a large circle, the final position with a cross, and the path taken by a line, with each step delineated by a small circle. The circle is bold where a collision has occurred, and the subsequent movements numbered chronologically. The stages of the robot’s movement are indicated with line style and symbols, which are explained in the legends. The airflow patterns for the scenarios shown in Fig. 13(a) and (d) are displayed in the predicted airflow shown in the screenshot of Fig. 12, and the measured values shown in Fig. 1, respectively. In all of the eighteen trials performed, Roma successfully identified the correct object as being the odor source. In addition, this was achieved expediently, with a typical time of 30–50 seconds compared to 6 to 15 min, reported in the literature, for other single robot methods operating in the same or a smaller area. The success of the experiments is encouraging. It shows that this new approach to odor localization is feasible and can be effective. The Reasoning Algorithm is reliable; however a more general airflow-modeling algorithm is required to make this research practically useful. It is currently limited in its ability to predict the airflow accurately for large inlets/outlets offset from the centre of the wall, and for curved and non right angled walls. The experimental infrastructure is established, and a system combining a new airflow algorithm together with complementary visual sensing is currently being developed.

5. Conclusion There are many useful and humanitarian applications for robots that can locate the source of a chemical plume. Currently, the majority of work in this area uses reactive control schemes that track an odor plume along its entire length, which is slow and difficult in cluttered indoor environments. This research employs a higher-level control scheme. A na¨ıve physics model is used to model the airflow in the robot’s environment. Then a reasoning system uses the airflow model to propose the path of an odor plume and predict the most probable locations of the odor source. This approach has been shown to be effective for odor localization in a known indoor environment, without the need for the robot to travel to the source. With further development there is great potential for this approach to lead to many valuable applications

229

by generalization to a wider range of environmental configurations. In addition, this project is the first example of a na¨ıve physics model that has been successfully applied in a real robotics application. The current lack of practical applications is salient. If robots are to operate effectively in unstructured environments, they will need to be able to predict the consequences of their actions. In order to do this, they will need to have a commonsense reasoning ability. Developments in this area are felt to be one of the keys to producing robust useful robots. This research involves the application of na¨ıve physics to the specific task of modeling airflow and through this, odor localization. However, as the initial success in this application has shown, there is great potential for more general use of na¨ıve physics within the field of artificial intelligence, and more specifically, robotics. Future work will concentrate on two main areas of improvement: development of a more general airflow modeling algorithm, and implementation of a range data acquisition system to enable autonomous map building. The current airflow-modeling algorithm works for a limited set of room configurations. The Reasoning Algorithm cannot operate effectively if the airflow map is inaccurate. It is possible to continue refining the current airflow algorithm through the addition and implementation of further na¨ıve rules. Other major additions will be hypothesis testing to validate the airflow map, and a predicted odor source verification stage, whereby the robot circumnavigates the most probable odor source and ascertains if it really is the source. These improvements will enable the robot to move into unknown unstructured environments, perform airflow modeling more reliably, and confirm predicted odor locations. Acknowledgments The work described in this paper was supported by the Australian Research Council funded Centre for Perceptive and Intelligent Machines in Complex Environments as well as grants from the Monash University Harold Armstrong Memorial Fund and Monash Engineering Faculty Grant Scheme. The authors wish to thank Tony Brosinsky and Maurice Gay for their assistance with the mechanical work.

References Alexandrou, A. 2001. Principles of Fluid Mechanics. Prentice Hall. Bobrow, D.G., and Hayes, P.J. (eds.). 1984. Special Volume on Qualitative Reasoning About Physical Systems. Artificial Intelligence. Buscemi, L., Prati, M., and Sandini, G. 1994. Cellular robotics: Behavior in polluted environments. In 2nd International Symposium on Distributed Autonomous Robotic Systems, Saitama. Davis, E. 1997. The naive physics perplex. AI Magazine, 19(4):51–79. Davis, E. 1997. Physical reasoning: Cooking problem. The Common Sense 98 Problem Page, http://www.dcs.qmw.ac.uk/ conferences/CS98/CS98Problems.html.

Springer

230 De Kleer and J. Williams, B. (eds.), 1991. Special Volume on Qualitative Reasoning About Physical Systems II. Artificial Intelligence. Duckett, T., Axelsson, M., and Saffiotti, A. 2001. Learning to locate an odor source with a mobile robot. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA ‘01), Seoul. Faltings, B., and Struss, P. 1992. Recent Advances in Qualitative Physics. MIT Press. Farrell, J.A., Pang, S., and Li, W. 2003. Plume mapping via Hidden Markov methods. IEEE Transactions on Systems. Man, and Cybernetics, 33(6):580–863. Gardin, F., and Meltzer, B. 1989. Analogical representations of naive physics. Artificial Intelligence, 38(2):139–159. Genovese, V., Dario, P., Magni, R., and Odetti, L. 1992. Self organizing behavior and swarm intelligence in a pack Of mobile miniature robots In search of pollutants. In Proceedings of the 1992 lEEE/RSJ International Conference on Intelligent Robots and Systems (IROS ‘92), pp. 1575–1582. Grasso, F.W., Consi, T.R., Mountain, D.C., and Atema, J. 2000. Biomimetic robot lobster performs chemo- orientation in Turbulence using a pair of spatially separated sensors: Progress and challenges. Robotics and Autonomous Systems, 30:115–131. Hayes, A.T., Marinoli, A., and Goodman, R.M. 2001. Swarm robotic odor Localization. In Proceedings of the IEEE Conference on Intelligent Robots and Systems (IROS ‘01), Wailea, pp. 1073–1078. Hayes, P.J. 1979. The naive physics manifesto, In Expert Systems in the Microelectronic Age, D. Michie (eds.), Edinburgh University Press. Hayes, P.J. 1985. The second naive physics manifesto. In Formal Theories of the Commonsense World, J. Hobbs and R. Moore (eds.), Norwood, NJ., Ablex Publishing Corporation, pp. 1–36. Ishida, H. 1994. Study of autonomous mobile sensing system for localization of odor source using gas sensors and anemometric sensors. Sensors & Actuators, A 45(2):153–157. Ishida, H. 1996. Odor-source localization in the clean room by an autonomous mobile sensing system. Sensors & Actuators, B 33(1–3):115–121. Ishida, H., Hayashi, K., Takusaki, M., Nakamoto, T., Moriizumi, T., and Kanzaki, R. 1996. Odor-source localization system mimicking behavior of silkworm moth. Sensors and Actuators A, 51:225–230. Ishida, H., Kagawa, Y., Nakamoto, T., and Moriizumi, T. 1995. Odorsource localization in clean room by autonomous mobile sensing system. The 8th International Conference on Solid State Sensors and Actuators (Transducers ‘95/Eurosensors IX), pp. 783–786. Ishida, H., Nakamoto, T., and Moriizumi, T. 1997. Remote sensing and localization of gas/odor source and distribution using mobile sensing system. International Conference on Solid-State Sensors and Actuators, pp. 559–562. Ishida, H., Nakayama, G., Nakamoto, T., and Moriizumi, T. 2002. Controlling a gas/odor plume-tracking robot based on transient responses of gas sensors. In Proceedings of IEEE Sensors, 2002, pp. 1665–1670.

Springer

Auton Robot (2006) 20:215–230 Ishida, H., Tanaka, H., Taniguchi, H., and Moriizumi, T. 2004. Mobile robot navigation using vision and olfaction to search for a gas/odor source. In Proceedings of IEEE/RSJ International Conference on Intelligent Robots and Systems, Sendai, Japan. Kazadi, S., Goodman, R., Tsikata, D., and Lin, H. 2000. An autonomous water vapor plume tracking robot using passive resistive polymer sensors. Autonomous Robots, 9(2):175–188. Kowadlo, G., and Russell, R.A. 2003. Naive physics for effective odor localization. In Proceedings of the Australian Conference on Robotics and Automation, Brisbane. Kuipers, B. 1994. Qualitative reasoning: Modeling and simulation with incomplete knowledge. The MIT Press, Cambridge, London. Lilienthal, A., Zell, A., Wandel, M., and Weimar, U. 2001. Sensing odor sources in indoor environments without a constant airflow by a mobile robot. In Proceedings of the 2001 IEEE International Conference on Robotics and Automation (ICRA ‘01), Seoul. Loutfi, A., Coradeschi, S., Karlsson, L., and Broxvall, M. 2004. Putting olfaction into action: Using an electronic nose on a multi-sensing mobile robot. In Proceedings of IEEE/RSJ International Conference on Intelligent Robots and Systems, Sendai, Japan. Rozas, R., Morales, J., and Vega, D. 1991. Artificial smell detection for robotic navigation. 5th International Conference on Advanced Robotics, Robots in Unstructured Environments (ICAR ‘91), pp. 1730–1733. Russell, R.A. 2001. Survey of robotic applications for odor-sensing Technology. The International Journal of Robotics Research, 20(2):144–162. Russell, R.A. 2001. Tracking chemical plumes in constrained environments. Robotica, 19(4):451–458. Russell, R.A. 2004. Locating underground chemical sources by tracking chemical gradients in 3 dimensions. In Proceedings of IEEE/RSJ International Conference on Intelligent Robots and Systems, Sendai, Japan. Russell, R.A. 2004. Robotic location of underground chemical sources. Robotica, 22(1):109–115. Russell, R.A., Bab-Hadiashar, A., Shepherd, R.L., and Wallace, G.G. 2003. A comparison of reactive robot chemotaxis algorithms. Robotics and Autonomous Systems, 45(2):83–97. Russell, R.A., Kleeman, L., and Kennedy, S. 2000. Using volatile chemical to help locate targets in complex environments. In Proceedings of the Australian Conference on Robotics and Automation, Melbourne, pp. 87–91. Russell, R.A., Thiel, D., Deveza, R., and Mackay-Sim, A. 1995. A robotic system to locate hazardous chemical leaks. In Proceedings of 1995 IEEE International Conference on Robotics and Automation, pp. 556–561. Watson, J. 1984. The tin oxide gas sensor and its applications. Sensors and Actuators, A 5:29–42. Weld, D.S., and De Kleer, J. (eds.), 1989. Readings in qualitative reasoning about physical systems. Morgan Kaufman Publishers, Los Altos.