Elsevier

Neurocomputing

Volume 175, Part B, 29 January 2016, Pages 1019-1032
Neurocomputing

Synthesis of odor tracking algorithms with genetic programming

https://doi.org/10.1016/j.neucom.2015.09.108Get rights and content

Abstract

At the moment, smell sensors for odor source localization in mobile robotics represent a topic of interest for researchers around the world. In particular, we introduce in this paper the idea of developing biologically inspired sniffing robots in combination with bioinspired techniques such as evolutionary computing. The aim is to approach the problem of creating an artificial nose that can be incorporated into a real working system, while considering the environmental model and odor behavior, the perception system, and algorithm for tracking the odor plume. Current algorithms try to emulate animal behavior in an attempt to replicate their capability to follow odors. Nevertheless, odor perception systems are still in their infancy and far from their biological counterpart. This paper presents a proposal in which a real-working artificial nose is tested as a perception system within a mobile robot. Genetic programming is used as the learning technique platform to develop odor source localization algorithms. Experiments in simulation and with an actual working robot are presented and the results compared with two algorithms. The quality of results demonstrates that genetic programming is able to recreate chemotaxis behavior by considering mathematical models for odor propagation and perception system.

Introduction

Around the world, different environmental conditions and sometimes negligence produce hazard zones that endanger population. Such disaster areas affected by hurricanes, earthquakes, fires and nuclear catastrophes need to be restored as soon as possible without risking more lives. Rescue teams work intensively to diminish the effects, but occasionally they cannot reach the complete area due to toxic environments, the potential presence of explosive materials, collapses, or a simple circumstance like inadequate space. Rescue robots have received considerable attention in recent years, thus providing solutions for those scenarios where human rescue teams are unable to work. Today, a major trend in robotics research is to incorporate different sensor capabilities inspired from solutions of the natural world. The idea is that as soon as robots are able to see, hear and touch, technology will be able to emulate the human capabilities for searching, mapping, exploration and localization of different targets, such as lost or injured people, safe trajectories, gas leaks, explosives, to mention but a few. Nevertheless, these capabilities may not be robust enough for real disaster scenarios due, for example, to poor visibility generated by the presence of obstacles that increase the difficulty of reaching specific target zones. A promising research area that could tackle such limits is based on the inclusion of olfaction capabilities inspired by solutions found in the animal kingdom. Animals use the sense of smell for diverse tasks like inspection, recognition, mating and hunting, despite not always being its principal perception mechanism. For example, dogs are trained to accomplish search and rescue operations within disaster areas, airports and borders. In this way, they use primarily the sense of smell to localize drugs, explosives, chemicals, hazardous substances and even persons [1]. Moreover, perception of the environment composition (odor molecules and concentrations) through olfaction might develop into a set of strategies that could be implemented within a sniffing robot so that it could find the direction of odor trails and follow them until it reaches a saturated zone [2] to finally locate and detect toxic gas leaks, the origin of a fire, and so on.

Nevertheless, an optimal perception system like those encountered in simple organisms is not currently available since artificial sensors differ significantly from their biological counterparts. Moreover, the algorithms based on natural processes attempt to emulate the behavior of some animals, such as casting, and sweeping spiral [3] without reaching the same level of performance; i.e., the odor source is not located with high accuracy or it requires a lot of time to be reached. In our work, we believe that the difference between both systems may be due to the fact that the natural smell sense, unlike an artificial perception system, is evolved over many years until it acquires a way of locating the odor. Thus, the idea of synthesizing artificial odor tracking strategies will be developed through artificial evolution; in particular, the application of genetic programming in combination with our artificial nose. Next, we review the main works devoted to odor source localization.

In the literature, there are many algorithms that aim to increase the efficiency of odor source localization from the viewpoint of sensor usefulness. The techniques are generally classified through the following functions: chemotaxis, anemotaxis and fluxotaxis, depending on the environment and capabilities of odor sensors. This research centers on the development of an artificial nose and in particular this paper deals with the development of chemotaxis algorithms. Traditionally, the chemical gradient derived from certain chemicals in the environment is the basis for orientation and movement of an agent – mobile robot – and it forms the base of chemotaxis algorithms. In general, the approach mimics the perception of odor using single or multiple sensors placed at different positions while calculating gradient responses over time [2]. The onboard or remote computer is responsible for analyzing the signals and their variations with respect to time and space. The first robot charged with odor source localization was presented by Rozas et al. in 1991 [4]. The design consists of following odor gradients by taking two or more measurements by one sensor from different positions at different times. In this algorithm, the robot had to measure odor concentration at four different positions. Additionally, if the new measurement was smaller than the previous one, the robot returns to the last position. Through this routine, a robot takes a lot of time to get closer to the source. Later, in the mid 1990s the first sensor design used to obtain a measurement from two positions at the same time was presented by Ishida et al. [5], [6]. This odor compass requires rotating the probe 360°, a process that took 20 s to obtain a direction and about one minute to recover from its initial state. Results showed that the system points to the trail direction but not always to the source position. Afterwards, a new stereo architecture implemented on a Koala mobile robot used measurements at different times and positions to obtain a gradient [7]. Nevertheless, the robot needed to be very close to the source for detection. Then, a mechanical implementation was presented [8] in which motor speed on each tire was proportional to their averaged concentration issued from an array of sensors. Hence, the robot was forced to turn when it reached some virtual walls, thus staying near to the odor source. Again, the robot needed to pass close to the source to detect the odor. Recently, work was presented using an unmanned aerial vehicle and a pseudo-gradient algorithm [9]. The Airrobot AR100-B micro-drone was used with an autonomous routine based on wind information and chemical gradient, sensed around the environment. Due to turbulence generated by its propellers, the drone should stay in the same place for a long time between measurements.

The principal drawback of previous algorithms were related to the sensors processing time since it took a lot of time to be ready for a second measurement (more than one minute). Some strategies even required more time since they need to cover the whole area several times (more than 20 min). Also, the odor source was not always reached due to multiple local maxima placed near the odor source. This happens because vapors are volatile and tend to homogenize the whole area, but in the case of constant gas leaks, maximum concentration is always at the exit of odor source. Sometimes the difference between both nostrils is provided by an airflow that helps us to circulate the odor around the system, creating a trail at the robot׳s rear part, and, as a consequence, sensors are constantly saturated by the same odor. Moreover, some systems, besides using chemotaxis, also collect wind information (anemotaxis), which is the most popular technique for outdoor environments. Thus, while considering that perception of wind speed is imperceptible for humans and common anemometers beyond 0.1 m/s [10], and that anemotaxis techniques are not appropriate for indoor environments that have small air currents. In this work, we provide evidence that considering only the chemical gradient is enough to reach an odor source indoors. With this as a foundation, all techniques could be improved based on chemical gradient, wind speed, mass flux, alone or in combination. This includes the use of robot teams or swarms where each one of them can have different behaviors or cooperate to reach bigger zones that will decrease the tracking time.

Robots could be designed to learn to use their odor sensors while considering the limitations at the moment of perceiving the environment. Comparatively, some research about perception systems that learn to discriminate odors was presented in 1999 [11]. In this case, an artificial neural network simulates the olfactory sensory neurons; thus, enabling discrimination of organic vapors. A similar approach in 2001 imitates the olfactory bulb including rank-order filtering over artificial neural networks [12]. Later, in 2004 the mathematical model for all biological olfactive layers using artificial neural networks was achieved [13]. On the other hand, Continuous Time Recurrent Neural Networks (CTRNNs) presented on 2013 [14], evolved odor source localization with a simulated robot equipped with a single chemical sensor and wind direction sensor. Schaffernicht and coworkers in 2014 [15] modeled and mapped the distribution of gas events, as well as detection and non-detection of a target gas using Bayesian Spatial Event Distribution. Recent work by Zhang and colleagues discusses localizing several odor sources [16], implementing a method based on niching Particle Swarm Optimization (PSO).

Moreover, to accomplish odor source localization, it is fundamental to consider the characteristics of the sensor with respect to desaturation time, concentration difference between sources, and reaction time. In our work, these features and their mathematical models are the basis of a learning perception system, which is used to derive a new technique that offers better results. The goal of this research is to obtain an algorithm that validates the use of chemical sensors to track and locate odor sources, especially when other sensors are limited or unavailable. Thus, the method relies on the perception of chemical odors while avoiding other sensors such as anemometers, cameras, sonars, and so on. In particular, the aim is to design algorithms, based only on olfaction (chemotaxis), that are able to follow straighter paths, thus reaching the source faster in comparison with current techniques. The algorithm for indoors obtained by genetic programming considers the mathematical models for odor propagation, and the perception system implemented into an unmanned ground vehicle (UGV) that looks for an optimal way of achieving the task in environments with imperceptible air currents for humans, while considering the limitations and advantages of the implementation.

Our paper is organized as follows: the problem statement is presented in Section 1.2 followed by a summary of our contributions in Section 1.3. In Section 2 the perception systems for chemical gradient are detailed. In particular, we describe the odor propagation model in 2.1, the chemotaxis technique in Section 2.2, and the artificial nose in Section 2.3. Section 3 details the chemotaxis algorithms. In Section 4 the description of our proposed methodology is comprehensively addressed. The simulation framework in Section 4.1, and the genetic programming method in Section 4.2. Then, in Section 5 the experimental setup and results are discussed. Finally, Section 6 provides the conclusions together with possible future work.

In the scenarios mentioned above, and others, a difficult problem is measuring and locating the odor source position caused by airflow patterns related to the phenomenon of advection and diffusion [17]. A solution of such real-world problems should consider a quick detection of the leaking gas origin since the faster the identification of an odor׳s trail direction the better. While anemotaxis is commonly used outdoors, since the vector component of advection is much bigger than that of diffusion; when indoors we present a solution based only on chemotaxis since airflow is mostly imperceptible. Hence, perception of the environment using an optimal signal processing model becomes crucial to understand and develop good control behaviors in the form of odor source localization algorithms that are simple to implement on a mobile platform [18], [19], [20]. In fact, the algorithm responsible to achieve the planning task depends not only on the environmental conditions, but also relies on the perception system and the physical characteristics of the robot; for example, desaturation time, measurement delays, sensors’ position, robot׳s mechanics, and so on. Thus, the analysis should focus on the definition of the environment with the aim of characterizing the odors dynamic behavior, since different environments could require different kinds of sensors and architectures to adequately perceive the signals and provide useful information for future analysis.

In this research, Genetic Programming (GP) is used as the optimization tool for training an UGV in such a way that it could be seen as an odor source localization platform, while taking into account the capabilities and limitations of the real-working robot, whose physical implementation of the perception system is capable of obtaining the source direction from where an odor is coming. A simulation environment was created to test different algorithms; it considers the mathematical models for both the spread of odor and the properties of the perception system. In this way, the modeling and simulation of its physical properties are used to obtain an algorithm that is evolved with the GP-approach to produce an optimal localization strategy. Thus, in order to validate the proposal against different techniques of chemotaxis at the state-of-the-art, the algorithms are compared in simulation and in practice with a real-working robot by deploying the artificial nose in a specially conceived environment in such a way to measure the chemicals in the room and testing all studied algorithms under the same circumstances.

The work presented in this paper extends the results of the perception system previously proposed [21], [22], [23] in such a way as to diminish the common drawbacks outlined before. It implements a bio-inspired nose system with the capability of determining the direction of a source placed at a given distance by using a pair of nostrils divided by a septum. In the inhalation process, the artificial nose concentrates the odor molecules near the sensorial system, and at exhalation the nose desaturates the sensors. The proposed design complements the sensor model detailed before [7], [24] by including the cyclic behavior of an artificial nose placed into a chamber. Thus, genetic programming for odor source localization is a novel technique that produces a precise algorithm while considering the capabilities and mathematical models of a given perception system and mobile robot. The secondary contributions of this research are listed below:

  • The simulation was developed on Netlogo which is general enough to run any kind of experiments. It proved to be useful for visualizing and characterizing the mathematical models of the environment and the physical implementation.

  • Odor source localization is achieved using only olfaction by chemotaxis for an unstructured indoor environment without using other sensors as anemometers and cameras.

  • The algorithm obtained by GP is better in terms of task achievement in comparison with common techniques for odor source localization that are based on chemotaxis.

  • This algorithm was implemented on a real working system at an indoor environment with small airflows and the odor source was reached faster than common techniques due to the tracking of straighter paths.

Section snippets

Perception of chemical gradient

Different techniques for odor source localization have been applied to approach several tasks while considering the environmental conditions and the kind of sensors that are used to design the perception system. Today, there are several proposals based on the design of sensor configurations and their deployment within a mobile robot. In particular, we will refer to them as openly exposed or cased. The first proposal is illustrated by sensors that are directly exposed to the environment without

Basic algorithms for mimicking chemotaxis

When we talk about odor source localization techniques three stages are identified. The first one is based on the exploration of the area to find an odor trail while the robot is actually doing other tasks. The second one relies on the act of following the trail until the robot strikes a specific target while considering time limit, physical frontier and visual information. The last one is the validation of the source that could be corroborated through the combination of other sensor readings

GP-trained odor source localization technique

It is well known that evolutionary algorithms are powerful tools to tackle difficult optimization problems. Such techniques are inspired from the Darwinian theory of natural selection, which is mainly based on concepts such as inheritance, mutation, crossover, selection and so on. In our work, genetic programming uses a population of computer algorithms for odor source localization that is evolved along a number of generations. It means that, generation by generation, GP stochastically

Experiments and results

Evolutionary computation has proved to be a successful tool for solving problems in robotics. In this work, we show that genetic programming can be seen as a viable methodology for synthesizing odor source localization techniques and the results provide evidence about its usefulness within such robotics problem. Thus, to validate our claim, a comparison against three different algorithms including two of the state-of-the-art and one designed with the proposed technique were tested in simulation

Conclusions

This paper presented simulation and experimental results of three different algorithms mimicking chemotaxis as performed by a robot. The algorithms for odor source localization were implemented in a simulation and successfully deployed in a real-world environment. The results in simulation match those achieved during the real-world test and therefore we can say that it is possible to design solutions to the OSL problem by the proposed technique. In fact, a better solution was obtained using GP.

Acknowledgements

This work has been supported by the National Council of Science and Technology of Mexico (Consejo Nacional de Ciencia y Tecnología - CONACyT), the “Laboratorio de Robótica del Área Noreste y Centro de México” founded by CONACyT, the Focus Group on Robotics at Tecnológico de Monterrey, as well as by the EvoVisión Team of CICESE and the project 155045 - “Evolución de Cerebros Artificiales en Visión por Computadora”. First author supported by scholarship 32081 from CONACyT. In particular we want

B. Lorena Villarreal graduated with Honorific Mention the Bachelor degree as a Mechatronics Engineer from the Tecnológico de Monterrey, Campus Monterrey in 2008. She also took courses on automotive engineering and design at the Fachhochschule Braunschweig/Wolfenbutel, in Wolfsburg, Germany and on Lean Manufacturing endorsed by the Institute on Industrial Engineers. She obtained her Ph.D. degree in 2014 from Tecnológico de Monterrey and was invited as a visiting researcher to collaborate with

References (34)

  • J. White et al.

    Odor recognition in an artificial nose by spatio-temporal processing using an olfactory neuronal network

    Neurocomputing

    (1999)
  • T. Roppel et al.

    Biologically-inspired pattern recognition for odor detection

    Pattern Recognit. Lett.

    (2000)
  • C. Browne et al.

    The use of scent-detection dogs

    Ir. Vet. J.

    (2006)
  • G. Kowadlo et al.

    Robot odor localizationa taxonomy and survey

    Int. J. Robot. Res.

    (2008)
  • N.J. Vickers

    Mechanisms of animal navigation in odor plumes

    Biol. Bull.

    (2000)
  • J. Rozas, R. Morales, D. Vega, Artificial smell detection for robotic navigation, in: 5th International Conference on...
  • H. Ishida, T. Nakamoto, T. Moriizumi, Study of odor compass, in: IEEE/SICE/RSJ International Conference on Multisensor...
  • T. Nakamoto, H. Ishida, T. Moriizumi, Active odor sensing system, in: ISIE ׳97 – Proceedings of the IEEE International...
  • A. Lilienthal, T. Duckett, A stereo electronic nose for a mobile inspection robot, in: Proceedings of the IEEE...
  • A. Lilienthal, T. Duckett, U. Nunes, A. deAalmeida, A. Bejczy, K. Kosuge, J. Macgado, Experimental analysis of smelling...
  • P.P. Neumann et al.

    Gas source localization with a micro-drone using bio-inspired and particle filter-based algorithms

    Adv. Robot.

    (2013)
  • D. Agdas et al.

    Wind speed perception and risk

    PLoS One

    (2012)
  • I. Valova et al.

    An oscillation-driven neural network for the simulation of an olfactory system

    Neural Comput. Appl.

    (2004)
  • G. de Croon, L. O׳Connor, C. Nicol, D. Izzo, Evolutionary robotics approach to odor source localization, Neurocomputing...
  • E. Schaffernicht et al.

    Bayesian spatial event distribution grid maps for modeling the spatial distribution of gas detection events

    Sens. Lett.

    (2014)
  • J. Zhang, D. Gong, Y. Zhang, A niching pso-based multi-robot cooperation method for localizing odor sources,...
  • J. Crank, The Mathematics of Diffusion, 2nd Ed., Clarendon Press, Bristol, England,...
  • Cited by (31)

    • A Deep Q-Network for robotic odor/gas source localization: Modeling, measurement and comparative study

      2021, Measurement: Journal of the International Measurement Confederation
    • Location of natural gas leakage sources on offshore platform by a multi-robot system using particle swarm optimization algorithm

      2020, Journal of Natural Gas Science and Engineering
      Citation Excerpt :

      At the same time, the proposed framework is evaluated based on the measurement error, and the effectiveness of the proposed scheme is verified. B. Lorena Villarreal (Villarreal et al., 2016) et al. developed a heuristic robot odor source location algorithm based on genetic algorithm. The quality of results demonstrates that genetic programming is able to recreate chemotaxis behavior.

    • Improving the prediction of material properties of concrete using Kaizen Programming with Simulated Annealing

      2017, Neurocomputing
      Citation Excerpt :

      ML techniques such as ANNs and Support Vector Machines (SVMs [11]) produce models that are considered black-boxes, i.e., they are hard (if not impossible) to understand. Therefore, EC algorithms such as Genetic Programming (GP, [12,13]), have been employed to generate smaller models by means of Symbolic Regression (SR [14–16]) that are easier to interpret [3,10,17,18] In an attempt to develop an SR method that requires a smaller number of objective function evaluations and still produces high-quality, interpretable models, De Melo proposed a hybrid approach named Kaizen Programming (KP, [21]), based on the Kaizen [22] and the Plan-Do-Check-Act (PDCA [23]) methodologies.

    View all citing articles on Scopus

    B. Lorena Villarreal graduated with Honorific Mention the Bachelor degree as a Mechatronics Engineer from the Tecnológico de Monterrey, Campus Monterrey in 2008. She also took courses on automotive engineering and design at the Fachhochschule Braunschweig/Wolfenbutel, in Wolfsburg, Germany and on Lean Manufacturing endorsed by the Institute on Industrial Engineers. She obtained her Ph.D. degree in 2014 from Tecnológico de Monterrey and was invited as a visiting researcher to collaborate with the EVOVision Group at the computer department of CICESE in Baja California. She received the MITs Technology Review award called Innovators under 35 México Edition in 2014. Recently, the Royal Academy of Engineering in collaboration with the University of Oxford gave her the opportunity to participate in a training course on technology commercialization as part of the Leaders in Innovation Fellowship. Her research interests are autonomous robotics and artificial intelligent systems.

    Gustavo Olague received the Ph.D. degree in Computer Vision, Graphics and Robotics from INPG and INRIA. He is currently a Professor in the Computer Science Department at CICESE in Ensenada. Professor Olague has written over hundred conference and journal papers and co-edited two special issues in Pattern Recognition Letters and Evolutionary Computation, as well as served as co-chair of the Real-World Application track at the Genetic and Evolutionary Computation Conference. Dr. Olague has received numerous distinctions such as the Talbert Abrams award offered by the ASPRS; best paper awards at major conferences like GECCO, EvoIASP, and EvoHOT; and received two times the Bronze Medal at the Human-Competitive awards at GECCO. He is the author of the book Evolutionary Computer Vision published by Springer.

    J.L. Gordillo graduated in Industrial Engineering from the Technological Institute of Aguascalientes, Mexico. He obtained both the D.E.A. degree and the Ph.D. in Computer Science from the National Polytechnic Institute of Grenoble, France, in 1983 and 1988, respectively. From 1989 to 1990 he was an Assistant Professor at the Department of Automatic Control of the Center for Advanced Studies and Research of the National Polytechnic Institute of Mexico (CINVESTAV-IPN).

    Currently he is Director and Professor at the Center for Robotics and Intelligent Systems and (CRIS) at the Tecnológico de Monterrey (ITESM). He has been a Visitor Professor in the Computer Science Robotics Laboratory at Stanford University (1993), at the Project Sharp of INRIA Rhone-Alpes in France (2002 and 2004), at LAAS-CNRS in Toulouse, France (2007–2008), and at some other universities and research institutes. His research interests are in computer vision for robotics applications, in particular autonomous vehicles, and the development of virtual laboratories for education and manufacturing. He participated and leaded R&D projects with industry like Honewell Bull in France, Sun Microsystems, Peñoles and TV Azteca; and government entities like the Mexican Army, the French-Mexican Laboratory for Computer Science (LaFMI), the Institute of the Water of the Nuevo Leon state (IANL), and the National Council for Science and Technology in Mexico (CONACyT). In particular, he promoted and leaded the National Net on Robotics and Mechatronics (RobMec). Actually he leads the Robotics National Laboratory, founded by CONACyT and ITESM, and the Robotics Research Focus Group at the ITESM.

    View full text