Journal of Medical and Biological Engineering, 32(6): 453-456
453
Technical Note
An Approach for Robot-Based Odor Navigation Andy Blanco Rodríguez1
Alejandro R. Garcia Ramirez2
Amarilys Lima Lopez2
Edson Roberto De Pieri3
Alejandro Durán Carrillo de Albornoz1,*
1
Institute of Science and Technology of Materials, University of Havana, Havana 10400, Cuba Department of Computation Engineering, University of Vale de Itajai, Itajaí-SC 88302-202, Brazil 3 Department of Automation and Systems, Santa Catarina Federal University, Florianópolis-SC 88040-900, Brazil 2
Received 24 Apr 2011; Accepted 4 Nov 2011; doi: 10.5405/jmbe.924
Abstract Environment pollution has pushed an increasing interest in the development of automated and intelligent systems for monitoring and analysing environmental variables. Odor sensing is generally applied to security, elderly care, food quality control, and domestic gas leakage or fire detection. In this field, recent applications of mobile robotics systems are being considered, more specifically, the detection of an odour source by a mobile robot. In this paper, an ethanol odor field, coming from a source, is measured using Metal Oxide Semiconductors sensors. An explore- exploit-based algorithm, which makes use of this environment information, is implemented in a mobile robot by combining the spiral and the plume-centered upwind search algorithms. Simulation results validate the robot performance during an odor detection task. Keywords: Odor detection, Plume tracing, Infotaxis, Mobile robot
1. Introduction The goal of odor localization is to find the source of a chemically volatile substance in the environment, which is vital for many life forms on Earth [1]. Some common applications are finding the source of dangerous substances, detecting drugs, searching for survivors in natural disasters, detecting fires in their initial stages, and locating unexploded mines and bombs. Mobile robotics are increasingly used in medical applications, such as robotized wheelchairs [2]. Humanoid robots have been developed to help educate and socialize children with special needs, and mobile assistive devices have been developed for monitoring, coaching, and motivating prescribed exercise therapy [3]. Powered exo-skeleton robots have been employed for medical care, in particular to help nurses lift and carry patients. Step rehabilitation robots can reduce the number of therapists needed for training [4,5]. They also allow the training process to be customized for each patient. Odor sensing is generally applied to security, elderly care, food quality control, and domestic gas leakage or fire detection [3]. A survey device with odor sensors that can detect the degree of human fatigue has been reported [6]. Odor localization is not as simple as sensing variations in a chemical concentration due to diffusion [7]. Several problems * Corresponding author: Alejandro Durán Carrillo de Albornoz Tel: +537-878-3969 E-mail:
[email protected]
concerning mobile robot odor detection in various environments (i.e., diffusion- dominated fluid flow, turbulencedominated fluid flow, and turbulence-dominated weak fluid flow) have been reported [1]. Numerous algorithms have been developed for finding an odor source and for determining a robot’s path. These algorithms commonly mimic the chemotactic behaviour of biological organisms, for example, zigzag or spiral [8]. Some algorithms also consider other parameters besides chemical information, such as air flow or temperature. Other tracing techniques are based on analytical odor distribution models. Ishida et al. [9] presented a mobile robot comprising chemical and air-flow sensors. Starting from the sensory data and knowing the position of the odor source in advance, the behavior of the plume was described with an odor distribution model. Marques et al. [10] computed a goal vector for guiding a robot towards the odor source using a concentration gradient. This gradient was obtained from an analytic gas distribution model and the measured upwind airflow direction. Lilienthal et al. reported an odor localization method based on an analytical model [11,12]. They used an experimental grid map which was obtained by measuring concentration values. Their method is suitable for turbulence-dominated weak fluid flow environments. An infotaxis navigation approach was proposed by Martin-Moraud [13]. This method considers the odor plume not as a continuous cloud, but as comprising intermittent odor patches which are dispersed by the wind. Far from the source,
J. Med. Biol. Eng., Vol. 32 No. 6 2012
454
the probability of finding one of these patches vanishes. In such conditions, a recommended strategy is to first explore the environment and to gather information, which is then used to lead the robot to the estimated source location. The results of a simulation study show that an infotaxis algorithm produces trajectories similar to those observed in the flight of moths attracted by a sexual pheromone [14]. In an earlier work, [15], LEGO hardware and the LabVIEW platform were used for an odor detection task. The navigation algorithm is based on a Braitenberg vehicle [16], and it does not consider the plume features. Therefore, it failed in 60% of experimental tests performed during an odor localization task. However, when the experiment was conducted by presenting alternately ethanol samples to sensors located at both sides of the robot, excellent reactive behavior (98% effectiveness) was reported [15]. To better understanding the scent distribution process, the present study develops an infotaxis-based technique. The odor plume is studied in two dimensions, with the gas concentration as a function of distance away from the source (exploration stage). The model assumes isotropic homogeneous turbulence and a constant unidirectional fluid flow across the entire area. An odor dispersion profile is built using two Metal Oxide Semiconductors MOS ethanol sensors. A plume-centered upwind search algorithm [9] is used for plume tracking (exploitation stage) using the output from the odor diffusion model. Simulation results show the viability of the proposed technique.
2. Materials and methods 2.1 System set-up Two ethanol gas sensors (TGS2620, Figaro, Japan) based on SnO2 semiconductors were employed for measuring the concentration of ethanol vapors. The sensors were attached directly to a robot, as shown in Fig. 1. A virtual instrument was developed using the LabVIEW 7.1 (National Instruments, U.S.A.) platform for acquiring data from the sensors.
placed 5 cm in front of the pot. The table surface was mapped into 56 square cells (10 cm × 11 cm). The outer line of the grid was chosen as the goal line. It was drawn parallel to the shorter side of the table and nearest to the fan. A 10 cm × 11 cm grid resolution was chosen. An odor map was obtained at the plane parallel to the grid and at the height of the sensors. After the fan was powered, the pot was shaken and 10 s later the program was started and the
Figure 2. Experimental set-up.
sensor output was recorded. Initial conditions were restored between experiments by removing the pot for one minute. 2.2 Theoretical odor profile An ethanol concentration two-dimensional graph profile was built from the coded sensor information and positions (X, Y). The sensor information was coded as a voltage level, which was converted into concentration values using: k C Rs
1
(1)
where C is the gas concentration value, Rs is the sensor resistance when exposed to a gas, and α is the sensor sensitivity, which depends on the surface material of the sensor and the target gas [17]. Hinze developed a distribution model based on turbulent diffusion [18]. Several researchers have employed this model to track the plume in an indirect reactive way. In the model, the time-averaged gas concentration of a point source is described as: C ( x, y)
l V r x exp 2 r CDT 2 CDT
(2)
where C(x,y) is the concentration value at point (x,y), CDT is the turbulent diffusion coefficient, l is the release rate of odor, V is the wind speed, and r is the distance between the odor source and the point (x,y), defined as: r Figure 1. Photograph of robot.
The robot was placed on the shorter side of a table (117 cm × 73 cm) and a plastic pot containing ethanol was place 115 cm away from it, as shown in Fig. 2. A DC fan was
xs x 2 ys y 2
(3)
x is the projection of the displacement with respect to the source ((xs–x), (ys–y)) in the upwind direction. x is zero in the cross-wind direction, and most significant in the downwind direction.
Robot-based Odor Navigation
x xs x cos ys y sin
(4)
where (xs,ys) is the location of the odor source, and is the angle of the upwind direction counter-clockwise from the x-axis. This model introduces some uncertainty in the estimation of the odor source position because it assumes a constant wind direction. This assumption is not invalid when obstacles and turbulence are present. From Eq. (1), when r →0, C(x,y) reaches its maximum value, which means that the odor source has been reached. If the wind direction changes, the shape of the odor plume will change. In this approach, the ethanol concentration values were measured at (x,y) points on the plane (Fig. 2) to make an odor concentration profile. The experimental data were represented using one dependent variable (concentration) and two independent variables (X and Y). The nonlinear surface fitting method, which allows multiple independent variables, was performed under OriginPro 8 software (OriginLab Corporation, U.S.A.). Also, a robotic simulation study was performed, in which the corresponding odor plume was traced starting from the experimental data obtained from the sensors. MatLab (MathWorks, U.S.A.) software was used for visualizing and processing the data.
455
difference in the amplitude responses of the sensors. This may be due to variations in the sensors or air flow direction generated by the fan. The odor profile in Fig. 5 exhibits Gaussian behavior in the two axis (X, Y position and C concentration), in accordance with turbulent diffusion. Figure 6 shows a theoretical curve,
Figure 4. Average concentration map obtained from sensors.
3. Results and discussion The sensitivity value of the sensor (α) was obtained from its data of sensitivity, which were supplied by the manufacturer (Fig. 3). With the values of Rs/R0 corresponding to 100 and 1000 ppm of gas concentration, a value of α = 0.69 was calculated, which is in the range reported by the manufacturer [19]. Constant k, which depends on the sensor material [19], was then calculated from Eq. (1), with R0 as the experimental sensor resistance obtained in clear air (4.22 kΩ). The resulting value was k = 20 × 104 Ω/ppm.
Figure 5. Experimental odor profile.
Figure 6. Theoretical odor profile.
Figure 3. Normalized sensor sensitivity characteristics.
Figure 4 shows the odor map obtained by averaging the values from the two sensors using MatLab functions imagesc and contourf. Concentration values were in the range of 0 and 1024. Higher concentration values have red tones and lower concentration values have blue tones. The odor distribution profile deviates on the left side of the table, which indicates a
where tones represent odor concentration levels (as explained above). These concentration values were computed from equations of the Hinze model. Comparing both experimental (Fig. 5) and theoretical (Fig. 6) profiles, the graphs show similar concentration shapes. However, the behavior of the concentration for ethanol in two dimensions (experimental curve) was not described by the model of Hinze (theoretical curve), despite of the visual similarity of both curves. This may be due to a deviation in the odor profile introduced by carrier air, the surface method employed for fitting, the turbulence effect due to robot geometry, the small size of the measurement table, and the insufficient
456
J. Med. Biol. Eng., Vol. 32 No. 6 2012
number of measured concentration values (X,Y points). Two kinds of fluid, air and the target gas, based on the platform proposed by Liu and Lu were simulated [20]. A source introduces the target gas into the air, causing a plume. At first, a threshold value, defined using the gas model, triggers a search spiral algorithm [3] which drives the robot to find a plume vestige. Then, to generate velocity commands, a plume-centered algorithm [4] leads the movement towards the odor plume. This approach requires only one chemical sensor (or the arithmetic mean of chemical sensor data). The implemented navigation algorithm used only the measured concentration. Figure 7 shows three paths followed by the robot in the simulated environment. In all cases, the robot successfully traced the plume. Note the path in the center and right snapshots. Concentration values increase as the robot moves towards the odor source. Gradient information can be extracted when the robot is moving at a relatively low speed because the measurement system generally filters out (up to a scale of several seconds) the fastest concentration fluctuations. This filtering is mainly due to the slowness of the sensor responses compared to the steepness of turbulent fluctuations. Thus, a robot equipped with such sensors can navigate in a much weaker turbulent odor plume. In such cases, isolated patches are replaced by a smoother concentration field, which fluctuates around a non-zero mean [21]. Sensor, control, and actuator delays were not modeled in this study.
concentration. Good simulation results were obtained. Experimental tests will be performed in future works to improve the fitting output.
Acknowledgements The authors thank PhD. Liliam Becherán and Brazil-Cuba 069/09 CAPES-MES Project: Development of a mobile robot and an electronic nose platform for environmental applications.
References [1] [2] [3]
[4]
[5] [6]
[7]
[8] [9]
[10] [11]
Figure 7. Odor tracing simulation snapshots.
4. Conclusion A mobile robot navigation scheme was developed for odor localization to better understand the mechanisms of biological odor detection. The environment is explored to map the gas concentration measured using chemical sensors. The robot is then directed to trace and find the odor source. In a simulation, the data did not strictly converge due to variations in the odor profile, the surface method employed for fitting, the turbulence effect, the small size of the measurement table, and the insufficient number of measured concentration values (X,Y points). Due to the local characteristics of gas sensor measurements, spatial coverage, and a certain amount of temporal time averaging during measurement, time transient concentration peaks might not be sufficiently averaged out, and thus these peaks remain in the concentration map as minor deviations from the smooth course of the distribution, i.e., regions of low
[12]
[13]
[14] [15]
[16] [17]
[18] [19]
[20]
[21]
G. Kowadlo and R. A. Russell, “Robot odor localization: a taxonomy and survey,” Int. J. Robot. Res., 27: 869-894, 2008. R. C. Simpson, “Smart wheelchairs: A literature review,” J. Rehabil. Res. Dev., 42: 423-436, 2005. H. Ishiguro, M. Asada, S. C. Shapiro, M. Thielscher, C. Breazeal, M. J. Mataric and H. Ishida, “Trends and controversies: human-inspired robots,” IEEE Intell. Syst., 21: 73-85, 2006. Laboratory of Biomechanical Engineering, Faculty of Engineering Technology, University of Twente: LOPES project (LOwer-extremity Powered ExoSkeleton). Available: http:// www.bw.ctw.utwente.nl/research/projects/lopes.doc/index.html Hal Project. Available: http://www.halproject.com T. Oyabu, A. Okada, O. Manninen and D. D. Lee, “Proposition of a survey device with odor sensors for an elderly person,” Sens. Actuator B-Chem., 96: 239-244, 2003. J. A. Farrell, J. J. Murlis, X. Long, W. Li and R. T. Cardé, “Filament-based atmospheric dispersion model to achieve short time-scale structure of odor plumes,” Environ. Fluid Mech., 2: 143-169, 2002. N. J. Vickers, “Mechanisms of animal navigation in odor plumes,” Biol. Bull., 198: 203-212, 2000. H. Ishida, T. Nakamoto and T. Moriizumi, “Remote sensing of gas/odor source location and concentration distribution using mobile system,” Sens. Actuator B-Chem., 49: 52-57, 1998. L. Marques, U. Nunes and A. de Almeida, “Olfaction based mobile robot navigation,” Thin Solid Films, 418: 51-58, 2002. A. Lilienthal, F. Streichert and A. Zell, “Model-based shape analysis of gas concentration gridmaps for improved gas source localisation,” IEEE Int. Conf. Robot. Autom., 3564-3569, 2005. A. Lilienthal and T. Duckett, “Gas source localisation by constructing concentration gridmaps with a mobile robot,” Proc. Eur. Conf. on Mobile Robots, 1-6, 2003. E. Martin Moraud and D. Martinez, “Effectiveness and robustness of robot infotaxis for searching in dilute conditions,” Front. Neurorobot., 4: 1-8, 2010. D. Martinez, “On the right scent,” Nature, 445: 371-372, 2007. A. R. García, A. Blanco, A. O. López, D. Bertol and A. Durán, “Environmental odor perception: an evaluation of a platform based on LabVIEW and the LEGO NXT,” Proc. Biosignals and Biorobotics Conf., 209-213, 2010. V. Braitenberg (Ed.), Vehicles: Experiments in Synthetic Psychology, Boston: MIT Press, 1984. N. Yamazoe and K. Shimanoe, “Theory of power laws for semiconductor gas sensors,” Sens. Actuator B-Chem., 128: 566-573, 2008. J. Hinze, Turbulence, New York: McGraw-Hill, 1975. M. A. Pérez, J. C. Á lvarez, J. C. Campo, F. J. Ferrero and G. J. Grillo (Eds.), Instrumentación Electrónica, Madrid, Spain: Thomson, 2006. Z. Liu and T. Lu, “A simulation framework for plume-tracing research,” Proc. Australas. Conf. on Robotics and Automation, 1-7, 2008. D. Martinez, O. Rochel and E. Hugues, “A biomimetic robot for tracking specific odors in turbulent plumes,” Auton. Robot., 20: 185-195, 2006.
