Behavioral Mo dels of the Praying Mantis as a 1 Basis for Rob otic Behavior a a b Ronald C. Arkin Khaled Ali Alfredo Weitzenfeld b Francisco Cervantes-Perez a Col lege of Computing, Georgia Tech, Atlanta, GA, 30332-0280 U.S.A. b Depto. Academico de Computacion, Instituto TechnologicoAutonomo de Mexico Mexico City, MX Abstract Formal mo dels of animal sensorimotor b ehavior can provide e ective metho ds for generating rob otic intelligence. In this article we describ e howschema-theoretic mo dels of the praying mantis derived from b ehavioral and neuroscienti c data can b e implemented on a hexap o d rob ot equipp ed with a real-time color vision system. This implementation incorp orates a wide range of b ehaviors, includin g obstacle avoidance, prey acquisition, predator avoidance, mating, and chantlitaxia b ehaviors that can provide guidance to neuroscientists, ethologists, and rob oticists alike. The goals of this study are threefold: to provide an understanding and means by which elded rob otic systems are not comp eting with other agents that are more e ective at their designated task; to p ermit them to b e successful comp etitors within the ecological system and capable of displacing less ecient agents; and that they are ecologically sensitive so that agent-environment dynamics are well-mo deled and as predictable as p ossible whenever new rob otic technology is intro duced. Key words: Schema-based control; b ehavior-based rob otics; praying mantis b ehavior 1 Ecological Rob otics The study of sensory guided b ehaviors in living animals has b ecome signi - cant not only for scientists working in neuroscience and computational neuro- 1 This researchwas conducted under funding provided by the NSF under Grant IRI-9505864 and CONACyT under Grants 546500-5-C006A and 546500-5- C018A Preprint submitted to Elsevier Preprint 29 Septemb er 1998 science, but also for those studying rob otics and distributed arti cial intelli- gence who are using functional principles generated from the study of living animals as mo dels to build computer-based automata that display complex sensorimotor b ehaviors. The research rep orted in this article, which follows these lines, is tied together by a common theoretical framework: schema the- ory. We strive to gain a b etter understanding of the relationship an agentmust maintain with its surroundings. Ecological rob otics refers to the incorp oration of asp ects of the relationship a rob ot maintains with its environmentinto its control system i.e., its ecology. One means for developing such a control sys- tem is by exploiting mo dels of b ehavior develop ed by ethologists or neurosci- entists. Although considerable research has b een conducted in the mo deling of neural controllers based on animal mo dels e.g., [8,15,47], incorp oration of environmental interactions has b een far less studied within the rob otics community e.g., [50]. Although some work has b een undertaken within the arti cial life arena [39,36], almost all of this work has b een conducted in sim- ulation or at b est on primitive rob otic implementations. All to o often, these approaches lack b oth a strong biological basis for their working assumptions and any formal underpinnings neural, b ehavioral, and computational for the results they obtain. It is our contention, that the use of schema theory [2] and neurophysiological and ethological mo deling metho ds can provide credible, generalizable, and useful results in this domain. Most of our previous research has considered the b ehavioral pro cess dynamics within an agent, and in some limited ways, collective b ehavior among similar agents [12]. In so doing wehave neglected signi cant asp ects of the environ- ment that can and should b e incorp orated into a systemic view of a rob otic agent's place within the world. Wenow fo cus on this broader view of rob otics, to gain a fuller understanding of how an agent participates with its environ- mental pro cesses. McFarland, for some time, has advo cated the concept of an agent's ecological niche [48,49]. This view mandates that in order to have a successful rob otic implementation, a rob ot must nd its place within the world, i.e., its niche. This niche will enable it to survive and successfully comp ete with other agents. This p ersp ective holds not only for rob otic systems but organizations as well - the novelty lies in its application to rob otic systems. McFarland's work has to date heavily emphasized economic pressures, but of course there are also many others. An in-depth understanding and dynamic mo deling of the relationship a rob ot has with its environment i.e., the overall ecology is imp ortant to ensure that elded rob otic systems are not comp eting with other agents that can do the task more e ectively and hence prove themselves useless; b e successful com- 2 p etitors within the ecological system and can p otentially displace less ecient agents; and b e ecologically sensitive so that agent-environmental system dy- namics are well-mo deled and as predictable as p ossible whenever new rob otic technology is intro duced. This article examines how such an understanding can b e develop ed through the use of biological mo dels of b ehavior that are p orted onto rob otic control systems. It is not the intention that these rob ots directly displace their biological counterparts, but rather that they b ecome capable of ultimately nding a unique niche in the world within which they can prosp er. In this article, we present b oth simulation studies and physical results obtained on the implementation of a mo del of praying mantis b ehavior on a rob otic hexap o d equipp ed with a real-time color vision system. As we are working with mo dels generated by animal scientists, we hop e that not only will these results havevalue within the rob otics community in terms of providing a path for generating intelligent b ehavior in machines, but that they may also serve as a basis for feedback for stimulation, regeneration, and re nement of the animal mo dels themselves. 2 Background and Motivation The relationships b etween the interdisciplinary areas in this research are de- picted in Figure 1. Biological data are used to generate abstract schema mo dels that can either b e directly imp orted into the our rob otic software control sys- tem generator MissionLab [43,44], or abstracted further into the context of neural networks NSL and then translated to abstract b ehavioral schemas ASL prior to imp ortation into a sp eci c rob ot control program. These soft- ware to ols MissionLab, ASL, NSL that emb o dy our notion of schema theory are describ ed further in Section 2.2. First, however, we present the biological motivation for our system. 2.1 Neuroscience and Ethology On the biological side, wehave b een studying visuomotor co ordination phe- nomena in amphibia toad and insects praying mantis. These animals live within a three dimensional environment, richinmultiple mo des of sensory signals, but their b ehavior is mainly guided by visual information. From an ecological p oint of view, these animals react to visual environmental domains of interaction which can b e classi ed into two groups: moving and non-moving ob jects. Diverse stationary ob jects may in uence the animal's next action which, in general, is directed to improve the animal's survival chances. For 3 Fig. 1. Framework for the study of mo dels of biological organisms as a basis for rob otic control. example, frogs movetowards zones in the visual eld where blue is prep on- derant, a situation that might b e asso ciated with the presence of prey to eat, and of water to maintain its b o dy humidity [37]. In the case of the praying mantis Fig. 2, when it is placed in an op en eld with no mobile ob jects around, it executes several motor actions that conform to what wehave called the chantlitaxia i.e., in search of a prop er habitat b ehavior. Fig. 2. Overhead outdo or view of a praying mantis. Di erentmoving ob jects may elicit a sp eci c b ehavior from these animals. For example: During the mating season, the presence of a female frog in the male's vi- sual eld yields an orienting resp onse towards the female, followed byan approaching action if the female is far away, or a clasping b ehavior if the female is within reaching distance in the frontal part of the visual eld. A predator-like stimulus may yield one of several avoidance b ehaviors de- p ending up on its parametric comp osition. In amphibia, a large ying stim- 4 ulus close to the animal releases a ducking resp onse [30,31,38,41], whereas, in the mantis, a similar stimulus elicits a deimatic b ehavior i.e., the man- tis stands up, and op ens the wings and forearms displaying a p osture that demonstrates a much bigger size than it actually has [46]. The presence of p otential prey may elicit one of several actions, dep ending on the spatio-temp oral relationship b etween the prey and the animal i.e., amphibia or insect. These include an orienting resp onse towards the part of the visual eld where the prey is lo cated, followed by an approaching b ehavior when the prey is lo cated far a eld in the frontal part of the visual eld.