MIRO: An Embedded Distributed Architecture forBiologically inspired Mobile Robots

Alfredo WeitzenfeldComp Eng Dept, ITAM

Rio Hondo 1, San Angel TizapanMexico City, MEXICO, 01000

Email: alfredo@itam.mx

Sebastian Gutierrez-NolascoSchool of Inf and Comp ScienceUniversity of California, IrvineIrvine, CA 92697-3430, USA

Email: seguti@ics.uci.edu

Nalini VenkatasubramanianSchool of Inf and Comp ScienceUniversity of California, IrvineIrvine, CA 92697-3430, USA

Email: nalini@ics.uci.edu

Abstract— Nature has always been a source of inspirationin the development of robotic systems. As such, the study ofanimal behavior (ethology) and the study of the underlying neuralstructure responsible for behavior (neuroethology) have inspiredmany robotic designs. In general, neuroethological based systemstend to be more complex than ethological ones thus being moreexpensive to compute, a common problem to both simulationand robotic experimentation. To overcome this problem, it isnecessary either to incorporate very powerful hardware or,particularly in the case of mobile robots, embed the robot viawireless communication into remote distributed computationalsystem where expensive computation can take place. Whilethe first approach simplifies the overall robotic architecture itresults in bulky and expensive robots. The second approachresults in smaller and less expensive robots, although involvingmore complex architectures. The work presented in this paperdiscusses the second approach that of embedding mobile robots todistributed computational systems. We describe our current workin conducting neuroethological robotic experimentation using theMIRO (Mobile Internet Robotics) system linked to the NSL/ASLneural simulation system. In optimizing overall system perfor-mance, communication between the robot and the computingsystem is managed by an Adaptive Robotic Middleware (ARM).

I. INTRODUCTION

Many different approaches have been proposed in recentyears in controlling autonomous robots. Lately, one of mostpopular has been that of behavioral based robotics [1], both interms of technological as well as biologically inspired robotics,such as those imitating animal ethology. In addition to thestudy of animal behavior neuroethological intends to modelneural structure as related to behavior. It should be notedthat there exist many robot architectures that do incorporatesome kind of neural processing, although most of them areof the artificial neural type involving non-biological trainingcapabilities [2]. Yet, there are important motivations behind thedesign of neuroethological robots. One reason is in providinginspiration for future robotics architectures, as has happenedbefore with neural architectures. Another important reasoninvolves neuroscientific experimentation where currently mostwork is done in terms of simulation. By providing with anexperimentation platform many issues that over simplifiedin simulation can be further analyzed by providing withembodiment.

One important concern with neuroethological robotic ex-perimentation involves how to achieve real-time performanceconsidering the expensive nature of neuroscientific processing.One approach to overcoming this challenge is to have ”super-robots” in analogy to supercomputers, something that usuallyresults in prohibitively expensive and bulky robotic systems. Asecond approach is to incorporate simpler and less expensiverobotic hardware although embedding it to an inexpensivenetwork of computers. Under such a computing architecturetime-consuming processing will be done remotely outsidethe robotic hardware, with the robot sending sensory inputand receiving motor commands via wireless communication.Such an approach reduces the robot’s physical size, powerrequirements as well as cost. A number of robotic architecturesembedded into the Internet have already been proposed [3]involving a large number of applications [4]. These efforts,most of them involving teleoperation, have highlighted thepotential of the Internet when linking remote robotic devicesto humans or other computational resources in a distributedfashion. Yet, to take advantage of such embedded architec-tures it is first necessary to overcome restrictions in wirelesstransmission bandwidth, unreliable communication or evencomplete failures.

In this paper we discuss our current work on embeddedrobotics, where (1) at the application level biologically in-spired neural based behaviors make it possible to experimentwith neuroethological robot architectures, while, (2) at thesystems level adaptive middleware support the embeddedrobotic system in a transparent fashion.

II. BIOLOGICALLY INSPIRED MOBILE ROBOTS

Through experimentation and simulation scientists are ableto get an understanding of the underlying biological mech-anisms involved in living organisms. These mechanisms,both behavioral and structural, serve as inspiration in thedevelopment of neural-based autonomous robot architectures.Some examples of animals having inspired robotic systemsare frogs and toads [5], praying mantis [6], cockroaches [7],and hoverflies [8] among others. To address the underlyingcomplexity in building such biologically inspired neural basedrobotics systems we usually distinguish among two differentlevels of modeling, behavior and neural networks [9].

At the behavioral level, neuroethological data from liv-ing animals is gathered to generate single and multi-animalsystems to study the relationship between a living organ-ism and its environment, giving emphasis to aspects suchas cooperation and competition between them. Examples ofbehavioral models include the praying mantis Chantlitaxia(”search for a proper habitat”) [10] and the frog and toad(rana computatrix) prey acquisition and predator avoidancemodels [11]. We describe behavior in terms of perceptual andmotor schemas [12] decomposed and refined in a recursivefashion. Schema hierarchies represent a distributed model foraction-perception control. Behaviors, and their correspondingschemas, are processed via the Abstract Simulation LanguageASL [13]. For example, in Arkin et al. [14] we describe apraying mantis prey-predator model as a basis for ecologicalrobotics, designed and implemented exclusively at the behav-ior level using finite state automata [15].

At the structural level, neuroanatomical and neurophysio-logical data are used to generate perceptual and motor neuralnetwork models corresponding to schemas developed at thebehavioral level. These models try to explain the underlyingmechanisms for sensorimotor integration. Examples of neuralnetwork models are tectum and pretectum-thalamus responsi-ble for discrimination among preys and predators [10], the preyacquisition and predator avoidance neural models [16] and thetoad prey acquisition with detour behavior model involvingadaptation and learning [17]. Neural networks are processedvia the Neural Simulation Language NSL [18]. Models thatinvolve neural networks are usually limited in scope as in[19], while more complex models [20] are simplified in termsof their inherent neural complexity. For example, let us con-sider the toads ”prey-predator” visuomotor coordination modeldescribed in [21], with schema and neural level componentsshown in Figure 1. The diagram shows two levels of modelinggranularity. At the schema level, blocks correspond to schemasor behavior agents representing animal or robot behavior.At the neural level, blocks represent neural networks, somehaving a direct correspondence to brain regions [22]. One ofthe main concerns with neural networks has been the expensivenature of computation. For example, a ”typical” retina model[23] may consist of more than 100,000 neurons and half amillion interconnections requiring many hours of simulation.This has lead to a number of distributed neural processingarchitectures [24].

III. EMBEDDED DISTRIBUTED ARCHITECTURE

As part of our current work in the design of embeddeddistributed architecture we have developed the MIRO (MobileInternet Robotics) system as shown in Figure 2. The architec-ture consists of multiple robots, each one connected to its ownparticular copy or instance of the neural computational systemwhere communication is done in a wireless fashion. Processingis distributed among the actual robotic hardware and theremote computational system. Although it would be possible inprinciple to share robot ”intelligence” among multiple robots,we keep a fully autonomous robot architecture in providing

Fig. 1. Toad’s prey-predator visuomotor coordination model archi-tecture with schema and neural level modules

with truly neuroethological experimentation. Other applicationcould easily take advantage of information sharing (see [25]for a discussion on distributed versus centralized roboticsystems). Under our MIRO architecture: (i) time-consumingprocesses are carried out in the (neural) computational system,implemented using the distributed NSL/ASL system while (ii)sensory input, motor output and other limited tasks are carriedout in the robot hardware.

A typical computation cycle involves the robot initiallysending sensory input (visual and tactile) data to the neu-ral computational system. The neural computational systemwould then process the sensory input cycling through itsneural modules while finally sending motor output back tothe robot. These cycles continue indefinitely or until somespecific task is completed. In such a way, the computationalsystem provides the robot’s ”intelligence”, while the robotdoes limited processing. The major challenge in the distributedarchitecture relates to the always-changing network and envi-ronment conditions (such as transient failures, disconnections,or reduced connectivity). For such purpose we have developedan Adaptive Robotic Middleware (ARM) framework managingcommunication between the robot and neural computationalsystem in adapting to changing conditions, primarily that ofcommunication, a major concern when video is involved.

The great advantage of a middleware approach is that itprovides with transparent mechanisms in enhance applicationresponse at run-time [26] [27] [28]. Most current middle-ware frameworks dynamically add and remove componentsat run time without interrupting system operation with com-munication services usually tailored to static conditions [29][30] [31] [32]. This approach is not well suited for highlymobile environments, where resource and power constraints,together with security issues (authentication, authorization andcommunication secrecy or integrity) pervade the application.In such environments, the communication framework must beable to automatically reconfigure itself in order to respond tochanges in the communication environment, a critical aspectin an embedded real-time architecture.

In such a way, the middleware architecture allows specifi-cation of communication requirements in a high level mannerthat can be later associated with low-level specific architecturalimplementations using a comprehensive set of basic communi-cation protocols. The middleware is responsible in determininghow, when and what information should be modified in orderto match communication fluctuations. For example, bandwidthadaptation enables information delivery in a manner sensitiveto the resources available, and may entail the use of techniquessuch as media conversion and compression to achieve thedesired results.

Fig. 2. MIRO embedded robotic architecture consisting of multipleautonomous robots linked to their own instance of the distributedneural computational system. All such instances are connected toInternet for remote monitoring

IV. EXPERIMENTS AND RESULTS

We have prototyped the MIRO robot architecture with anumber of experiments involving prey acquisition predatoravoidance. For example, in Figure 3 we show three differentexperiments involving a toad and a barrier in front of a prey,where fencepost gaps interposed [33] together with simulationresults (Figure 4 for the corresponding experiments. Theoriginal simulations were developed in the NSL C++ systemand then ported to the newer NSL Java system, currently linkedto the MIRO robotic architecture. To monitor system results,Internet-linked aerial cameras as well as the robot cameraswere included, as show in Figure 5 (top). Note that one ofthe key advantages of the MIRO distributed architecture isthat neural behaviors can be visualized at the same time asthe actual experiments it performing, as shown in Figure 5(down). Obviously there is an additional penalty to pay inperformance but it is well worth during model developmentor fine-tunning. In Figure 6 we show sample output for oneof the experiments, involving prey acquisition with a 10cmbarrier showing direct detour. The experiment was carried outon a single Lego-based robot connected in a wireless fashionto the MIRO system. A wireless camera was added on top ofthe robot transmitting video in a wireless fashion to remotevideo capture devices. Initial experiment control, experimentmonitoring and model visualization we all carried out from

Internet via a client-server architecture involving applets andservlets.

Fig. 3. A. Approach to prey with single 10cm barrier with immediatedetour. B. Approach to prey with single 20cm barrier: first trialwith toad in front of 20cm barrier (numbers indicate the successionof the movements). The toad directly approaches de center of thebarrier requiring successive trials to manage the detour around it. C.Approach to prey with single 20cm barrier. After 3 trials the toaddetours directly around the 20cm barrier. Arrowheads indicate theposition and orientation of the toad following a single continuousmovement after which the toad pauses.

Fig. 4. In diagrams D-F we see corresponding simulated results ofthe experiment.

V. DISCUSSION

The work presented in the paper overviews the challengesand complexity in modeling autonomous robots inspired bybiological systems in terms of both behavior and neuralstructure. One of the motivations behind this work is to provideneuroscientists with robotic experimentation capabilities aswell as prototyping new robotic architectures. One primaryconcern with neural processing is the extensive nature of com-putation, a crucial concern with real time robotics. To improveon performance and reduce the size and cost of robots, we havedeveloped an embedded distributed robot architecture sup-ported by adaptive middleware managing overall architectureand communication. While most time-consuming tasks cantake advantage of the distributed robotic system by processing

Fig. 5. Top: Internet aerial view of autonomous robot and robot’scamera view of ”blue” prey-like stimulus. Down: NSL frames show-ing results from different visual and neural modules in a basic preyacquisition robot experiment (without barrier).

Fig. 6. Results from prey acquisition experiment for 10cm barrierwith direct detour around barrier.

them remotely, there are a number of issues that arise fromsuch a distributed architecture, such as what happens whencommunication between the robot and computational systemactually fails or becomes extremely slow or unreliable. Therobot could respond in many ways, simply waiting withoutdoing anything until communication is restored, ending itsmission, or performing other more limited tasks that may put itback in action. Additionally the robot could actively search fora location where communication can be reestablished. As partof the process of robot experimentation we have taken modelspreviously simulated under NSL where their correctness isfirst tested. After that, the models are prototyped under theMIRO robot architecture to test their behavior under real worldconditions. The MIRO architecture has proven quite beneficialproviding real-time monitoring capabilities of both externalas well as internal robot behavior. Among the interestingaspects that have emerged from the robot experiments is the

problem of ”losing” the prey once the robot directs itselfaround the barrier. While this can be solved by a ”pan”control on the camera, where the camera can always ”look”into the prey, it raises an interesting number of issues such arecalling prey positions from memory such as with memorysaccade models [34]. Until now we have experimented withsingle robot neuroethological models, such as prey acquisitionand predator avoidance. We are currently working on multi-robot experimentation with self-made robots, where each robotinstantiates its own prey, robot or predator behavior. In general,the MIRO architecture and the Adaptive Robotic Middlewareare currently at the prototype stage and have not yet beencompleted. We expect that once we complete and integratethese two architectures we will be able to incorporate morecomplex neuroethological models as currently done.

ACKNOWLEDGMENTS

This research was supported by the Office of Naval Re-search under MURI Research Contract N00014-02-1-0715 andthe UC MEXUS CONACYT Advanced Network ServicesCollaborative Grant. The authors would like to thank the”Asociacin Mexicana de Cultura, A.C.” and the differentstudents at the CANNES Laboratory at ITAM who participatein this project.

REFERENCES

[1] Arkin, R.C, Behavioral based Robotics. MIT Press, 1998.[2] Sutton, R. and Barto, A., Reinforcement Learning: An Introduction.

MIT Press, 1998.[3] Sukhatme, G.S. and Mataric, M.J., “Embedding Robots Into the Inter-

net,” Communication of the ACM, vol. 43(5) pp 67-73, 2000.[4] Goldberg, K. and Siegwert, R., Beyond Webcams: An Introduction to

Online Robots. MIT Press, 2002.[5] Arbib, M.A, “Levels of Modeling of Mechanisms of Visually Guided

Behavior,” in Behavior Brain Science 10:407-465, 1987.[6] Cervantes-Perez, F., Franco, A., Velazquez, S. and Lara, N., “A Schema

Theoretic Approach to Study the ’Chantitlaxia’ Behavior in the PrayingMantis,” in First Workshop on Neural Architectures and Distributed AI:From Schema Assemblages to Neural Networks, 1993.

[7] Beer, R. D., Intelligence as Adaptive Behavior: An Experiment inComputational Neuroethology. Academic Press, 1990.

[8] Cliff, D., Neural Networks for Visual Tracking in an Artificial Fly,Towards a Practice of Autonomous Systems: First European Conferenceon Artificial Life, Editors, F.J., Varela and P. Bourgine. MIT Press,1992.

[9] Arbib, M.A., The Metaphorical Brain 2. Wiley, 1989.[10] Cervantes-Perez, F, Lara, R. and Arbib, M.A., “A neural model of

interactions subserving prey-predator discrimination and size preferencein anuran amphibia,” Theoretical Biology, vol. 113, 117-152, 1985.

[11] Cobas, A. and Arbib, M.A., “Prey-catching and Predator-avoidance inFrog and Toad: Defining the Schemas,” Theoretical Biology, vol. 157,271-304, 1992.

[12] Arbib, M.A., Schema Theory, in the Encyclopedia of Artificial Intelli-gence, Editor Stuart Shapiro, 2:1427-1443, 2nd ed. Wiley, 1992.

[13] Weitzenfeld, A., “ASL: Hierarchy, Composition, Heterogeneity, andMulti-Granularity in Concurrent Object-Oriented Programming,” inWorkshop on Neural Architectures and Distributed AI: From SchemaAssemblages to Neural Networks, 1993.

[14] Arkin, R.C., Ali, K., Weitzenfeld, A. and Cervantes-Perez, F., “BehaviorModels of the Praying Mantis as a Basis for Robotic Behavior,” Roboticsand Autonomous Systems, vol. 32 (1) pp. 39-60, 2000.

[15] Arkin, R.C., Cervantes-Perez, F. and Weitzenfeld, A., “EcologicalRobotics: A Schema-Theoretic Approach,” Intelligent Robots: Sensing,Modelling and Planning, eds. R.C.Bolles, H.Bunke, and H.Noltemeier,vol. World Scientific 377-393, 1997.

[16] Cervantes-Perez, F., Herrera, A. and Garca, M., “Modulatory effectson prey-recognition in amphibia: a theoretical ’experimental study’,”Neuroscience: from neural networks to artificial intelligence, Editors P.Rudoman, M.A. Arbib, F. Cervantes-Perez, and R. Romo, vol. (4):426-449, 1993.

[17] Corbacho, F. and Arbib M., “Learning to Detour,” in Adaptive Behavior,3(4):419-468, 1995.

[18] Weitzenfeld, A., Arbib, M. and Alexander, A., The Neural SimulationLanguage: A System for Brain Modeling. MIT Press, 2002.

[19] Fagg, A., King, I, Lewis, A., Liaw, J. and Weitzenfeld, A., “A Testbedfor Sensorimotor Integration,” in IJCNN, 1:86-91, 1992.

[20] Weitzenfeld, A., Cervantes, F. and Sigala, R., “NSL/ASL: Simulationof Neural based Visuomotor Systems,” in International Joint Conferenceon Neural Networks, 2001.

[21] Weitzenfeld A., “A Multi-level Approach to Biologically InspiredRobotic Systems,” in 10th International Conference on Artificial NeuralNetworks and Intelligent Systems, 2000.

[22] Weitzenfeld, A. and Arbib, M., “A Concurrent Object-Oriented Frame-work for the Simulation of Neural Networks,” in Workshop on Object-Based Concurrent Programming (ECOOP/OOPSLA), 1991.

[23] Teeters, J.L. and Arbib, M.A., “A model of the anuran retina relatinginterneurons to ganglion cell responses,” Biological Cybernetics, vol.(64):197-207, 1991.

[24] Weitzenfeld A. and Gutirrez-Nolasco, S., “ASL/NSL: A Multi-levelComputational Model for Distributed Neural Simulation,” in SCSC 2000Summer Computer Simulation Conference, 2000.

[25] Balch, T. and Arkin, R.C., “Communication in Reactive MultiagentRobotic Systems,” in Autonomous Robots, 1994.

[26] Gutierrez-Nolasco, S. and Venkatasubramanian, N., “A ComposableReflective Communication Framework,” in IFIP/ACM Workshop onReflective Middleware, 2000.

[27] ——, “A Reflective Middleware Framework for Communication inDynamic Environments,” in International Symposium on DistributedObjects and Applications, 2002.

[28] Venkatasubramanian N., Deshpande M., Mohapatra S., Gutierrez-Nolasco S. and Wickramasuriya J., “Design and Implementation of aComposable Reflective Middleware Framework,” in IEEE InternationalConference on Distributed Computer Systems, 2001.

[29] Costa F.M. and Blair G.S., “Integrating Meta-information Managementand Reflection in Middleware,” in International Symposium on Dis-tributed Objects and Applications, 2000.

[30] Kon F., Roman M., Liu P., Mao J., Yamane T., Magalhaes L.C.and Campbell R.H., “Monitoring, Security and Dynamic Configurationwith the Dynamic TAO Reflective ORB,” in IFIP/ACM InternationalConference on Distributed Systems Platforms and Open DistributedProcessing, 2000.

[31] Loyall J., Schantz R., Zinky J. and Bakken D., “Specifying andMeasuring Quality of Service in Distributed Object Systems,” in FirstInternational Symposium on Object-Oriented Real-Time DistributingComputing, 1998.

[32] Pal P.P., Loyall J., Schantz R., Zinky J., Shapiro R. and MegquierJ., “Using QDL to specify QoS aware Distributed (QuO) ApplicationConfiguration,” in Third IEEE International Symposium on Object-Oriented Real-Time Distributed Computing, 2000.

[33] Corbacho, F. and Weitzenfeld A., “Learning to Detour,” The NeuralSimulation Language, A System for Brain Modeling, vol. MIT Press,2002.

[34] Dominey, P. and Arbib, M.A., “A cortico-subcortical model for gener-ation of spatially accurate sequential saccades,” Cerebral Cortex, vol.1(2):153-175, 1992.