Tracking scents and odor sources is a major challenge in robotics, with applications to the localization of chemical leaks, drugs, and explosives (Russell 1999). Nowadays, animals are commonly used in safety and security tasks because of their excellent smell detection capabilities. Examples include dogs and honeybees (Rains et al. 2008). However, using animals to sniff specific odors related to bombs or explosives has several drawbacks. On top of the hazards of such endeavors, animals like dogs show behavioral variations and changing moods. They get tired after extensive work and require frequent retraining as their performance decreases over time. As an alternative, could we envision using olfactory robots to advantageously replace animals for these tasks?
At short distances from the source, olfactory search methods inspired by bacterial chemotaxis provide acceptable solutions for navigating a robot. Bacteria like Escherichia coli direct toward chemo-attractants by climbing a concentration gradient. They alternate between periods of straight swims called runs and random reorientations called tumbles (Berg 2003). Such biased random walks have been implemented on real robots (Lytridis et al. 2006; Russell et al. 2003; Marques et al. 2006), yet with a limited success, mainly because of the use of a single odor sensor. Unlike bacteria, other animals may use instantaneous gradient information assessed by comparing the responses of spatially separated chemosensors. This claim is supported by experiments showing that unilateral lesions, like the blockage of the nasal airflow in one nostril in rats or the ablation of one of the antennae in crayfish, impair odor source localization (Kraus-Epley and Moore 2002; Rajan et al. 2006; McMahon et al. 2005). Autonomous olfactory robots using bilateral comparison include the robotic “lobster” for tracking saline plumes in water (Consi et al. 1995; Grasso et al. 1997; Grasso 2001), the Braitenberg olfactory robot (Lilienthal and Duckett 2003), and our robot (Hugues et al. 2003; Martinez et al. 2006), whose trajectory curvature was constantly modulated by the difference in concentration between left and right sensors (Figure 5.1a). A prerequisite to all the aforementioned chemotactic robots is the existence of a relatively smooth concentration gradient. The experiments we performed revealed that a concentration gradient can effectively be measured (see Figure 5.4 , left in Martinez et al. 2006), but only when the robot moves slowly (2.5 cm/s) and near to the source (search area limited to 3 m2). As chemotactic search strategies are applicable only in the vicinity of the source, we considered the possibility of exploring the environment by using vision, in addition to olfaction (Martinez and Perrinet 2002). An important limitation nevertheless is that odor source candidates need to be identifiable from visual features.
Far from the source, the concentration landscape of an odor, called a plume, is very heterogeneous and unsteady, and consists of sporadically located patches (Weissburg 2000; Roberts and Webster 2002). Even at moderate distances (order 10 m), detections become sporadic and only provide cues intermittently. Given this discontinuous flow of information, how then can we efficiently navigate a robot toward the source over moderate or large distances (order 100 m)? It is well known that insects such as male moths successfully locate their mates over distances of hundreds of meters. To do so, male moths adopt a typical behavior (for reviews see, e.g., Murlis et al. 1992; Kaissling 1997; Vickers 2006). Upon sensing a pheromone patch, they surge upwind, and when the odor information vanishes, they perform an extended crosswind casting until the plume is reacquired. This strategy has the advantage of being purely reactive; i.e., actions are completely determined by current perceptions, that is, surge upwind upon sensing a pheromone patch and cast crosswind when odor information vanishes. Such reactive methods have been simulated or implemented on robots in various forms (Kuwana et al. 1999; Pyk et al. 2006; Balkovsky and Shraiman 2002). An efficient variant is the spiral-surge strategy (Hayes et al. 2002; Lochmatter et al. 2008) that combines upwind surge in the presence of the odor with spiraling in its absence (Figure 5.1b). Yet performance of reactive strategies at distances beyond 100 m, when the reacquisition of the plume becomes very unlikely, is unclear. Reactive casting-surge methods address the search problem only from an imitation perspective; i.e., they mimic the choices performed by animals through a rule-based approach, regardless of the underlying mechanisms from which the behavior emerges. This biomimetic approach raises the question of how well reactive strategies may be adapted to new environmental conditions as those occurring when the distance from the source increases.
For those conditions, a more sophisticated method, infotaxis, was proposed recently (Vergassola et al. 2007; Martinez 2007). Infotaxis relies on Bayesian inference to maximize information gain, and exploits the expected distribution of odor encounters. It involves a period of exploration during which the searcher builds a probabilistic map of the source location. As the searcher accumulates information, the map becomes sharper and its entropy—which reflects the uncertainty about the location of the source—decreases. Because the expected search time is determined by the uncertainty of the belief, the searcher moves so as to maximize the expected reduction in entropy, and therefore the rate of information acquisition. Maximizing information gain entails a competition between two actions, exploitation and exploration. The former drives the searcher toward locations where the probability of finding the source is high. The latter favors motion to regions with lower probabilities of source discovery but high rewards in terms of information gain. Infotaxis is a cognitive strategy in the sense that an internal model of the world is built from past detections and actions so that memory and learning play a crucial role (Figure 5.1c).
Advantages, disadvantages, limitations, and biological plausibility of cognitive and reactive search strategies are largely unknown and remain to be quantified. In this chapter, we review the two approaches and report comparisons based on simulations and robotic experiments. To our knowledge, only two studies have considered a robotic implementation of infotaxis: Lochmatter (2010) and Moraud and Martinez (2010).
© 2013 by Taylor & Francis Group, LLC.