Biologically Inspired Mobile Robot Control. Francesco Mondada Edoardo Franzi * 1. Introduction and goal of the project

Transcription

1 Biologically Inspired Mobile Robot Control Algorithms Francesco Mondada Edoardo Franzi * Abstract. To get a better understanding of the application of articial neural networks to the robotics eld, the rst part of this 3-year project supported by the NFP - PNR23 project was devoted to the study of some applications of neural networks to motor control, sensor fusion and image processing and to the development of sensor interfaces and neural network dedicated accelerators. The second part of the project, described in this paper, takes advantage from the experience gained in the rst part and studies the application of articial neural networks to a particular sub-eld of robotics: mobile robotics. Most of the problems of the domain seem to be particularly adapted to neural network solutions. To illustrate this, we present in the rst part of the article an example implementation of a very simple behaviour in a real robot using several design methodologies. The robot used for these experiments is described. In the second part of the document we present some more complex experiments based on a distinctive capability of neural networks, that is the ability to learn from examples. In the last two sections we present the beginning and the direction of our future work: The goal of this project is to study the neural networks approach applied to a group of mobile robots. In section 5. a rst experiment involving ve robots is presented, while the following section outlines the future strategy. Laboratoire de microinformatique (LAMI), Departement d'informatique, Swiss Federal Institute of Technology (EPFL), 1015 Lausanne, Switzerland, mondada@di.epfl.ch & franzi@di.epfl.ch 1. Introduction and goal of the project There is a wide range of industrial applications of autonomous mobile robots, including robots for automatic oor-cleaning in buildings and shellcleaning on ships, mobile surveillance systems, robots transporting parts in a factory without the need for xed installations, fruit collection and harvesting systems. These mobile robot applications are out of the reach of current technology. Nevertheless, today's techniques have large success in other classical domains of robotics as, for instance, in the automatic assembly. This indicates that in the mobile robotics eld the problems to solve have a higher level of complexity of those present in the classical robotics. Actually, the mobile robot must deal with the unpredictable real world, very dierent from the articial and deterministic working place of a traditional robot, e.g., in a car factory. Several researchers extimate that the classical computer science approach used today in robotics cannot solve the problems of the mobile robotics and have proposed new approaches. One of the most authoritative researcher going in this direction is R. A. Brooks who proposed the subsumption architecture [1]. Other researchers propose new computational approaches like the fuzzy logic [2] or the articial neural networks [3] [4]. The concept of this last approach is fascinating: It consists in designing new computational structures using all the available information on the very ecient world of biology [5] [6]. The resulting computational structures are very dierent from those of the computer science world. The most interesting feature of this approach is that biological neural networks have been designed to be adapted to the real world. This suggest the biological approach as the most appropriate to solve the problems existing in the mobile robotics eld. Our work goes in this direction and it will explore the possibilities of this eld. The redundancy, the parallelism and the resulting robustness of the computational structure are

2 Francesco Mondada and Edoardo Franzi Additional turrets Vision Grippers manipulator Inter-robot communications Parallel extension bus CPU board RAM EPROM A D CPU Multi-microcontroller extension link k1.eps mm Serial (RS232) MC MHz 32-bit microcontroller 1/2 MByte of RAM and EPROM 6 x analog inputs of 10-bits of resolution Asynchronous serial link Synchronous multi-microcontroller link Basic configuration Local sensory-motion bus Control Accus. Bridges H A PWM Sensory-motion board Motor 2 x motors 2 x incremental sensors (600 imp./turn) 8 x I.R. proximity sensors 4 x NiCd accumulators Figure 1: Khepera architecture and extensions. the interesting points of the biological approach. These mechanisms are not only exploited inside the brain structure, but also at the level of groups of animals, giving very interesting results in collective behaviour of insects, for instance. To gain a better understanding of these mechanisms, we will use the knowledge present at the two levels and will develop, in parallel with the study of neural networks control algorithms, experiments on collective behaviour. The sole way to validate an algorithm that has to deal with the real world is to test it on a real robot. For this reason, our work is based on experiences with real robots. In order to be able to perform these experiments with groups of robots and to have a good control on the robot architecture at the hardware and software level, we have developed our own experimentation mobile robot, Khepera, presented in the next section. 2. Khepera, a real mobile robot Thanks to the extensive knowledge acquired with the rst part of our PNR23 project [7], it was possible to design an original miniature robot. The requirements to implement collective behaviour algorithms such as neural processing capability, sensor interfacing and representation, and the need for extension capabilities have led us to design the architecture of the miniature robot Khepera [8]. No similar product exists on the market. 2.1 Hardware The Khepera architecture (gure 1) is built around two main boards: the CPU one and the basic sensory-motion one. Application-specic extension turrets for vision, inter-robot communications or equipped with grippers can be directly controlled via the Khepera extension busses. Khepera can be powered by an external supply when connected to a visualisation software tool for a long time; however, on-board accumulators provide Khepera with thirty minutes of autonomous work. Maybe one of the most interesting features of Khepera is the possibility of connecting extensions on two dierent busses. One parallel bus is available to connect simple experimentation turrets. A more sophisticated interface scheme with turrets uses a small local network; this allows the connection of intelligent turrets (equipped with a local microcontroller) and the migration of conventional or neural pre-processing software layers closer to the sensors and actuators. This topology makes it possible to implement distributed biological controls, such as arm movement coordination or feature extraction and pre-processing in the vision, as observed in a large number of insects. Of course, this multi-microcontroller approach allows the main microcontroller of Khepera to execute

3 Biologically Inspired Mobile Robot Control Algorithms only high level algorithms; therefore attaining a simpler programming paradigm Available main conguration. The new generation of Motorola microcontrollers and in particular the MC68331 makes it possible to build very powerful systems suitable for the miniature neural control. Khepera takes advantage from all the microcontroller features to manage its vital functionality. The basic conguration of Khepera is realised around two boards: the CPU and the sensory-motion one. The CPU board is a complete 32 bit machine including a 16 MHz microcontroller, system and user memory, analogue inputs, extension busses and a serial link allowing a connection to dierent host machines (terminals, visualisation software tools, etc.). The sensorymotion board includes two DC motors coupled with incremental sensors, eight analogue infra-red (IR) proximity sensors and an on-board power supply scheme. k13.eps mm Figure 2: Khepera robot with gripper and vision modules Available additional turrets. To make experiments involving environment recognition, object detection, object capture and recognition possible, two intelligent turrets have been studied and built: the stereoscopic vision and the gripper one. The stereoscopic vision gives one-dimensional images useful for robot navigation (frequential ltering used in obstacle detection and avoidance was described in [9]). The gripper turret makes it possible for Khepera to interact with objects in its environment. Dierent classes of objects can be detected by the gripper sensors (classes characterised by the size and the resistivity of the objects) Additional turrets under study. Robots having collective behaviour need means to perform inter-robot communications and interrobot localisation. These functionalities are under study at the moment of writing. 2.2 Software Managing all the Khepera resources is a complex task. The large number of asynchronous events to control and the necessity to share some critical interfaces have led to the development of a complete low-level software organised as a collection of primitives [10]. At the moment of writing, experiments with Khepera are performed in two different ways: in the stand-alone conguration and in connection with visualisation software tools. In stand-alone applications all the Khepera resources are directly controlled by the low-level collection of stack oriented call systems (directly accessible by a C compiler). Visualisation software tools control Khepera through the standard RS232 link by a generic high level protocol. Indeed, performing experiments by this way is very simple, because we can make a complete abstraction from Khepera hardware implementation. 3. Robustness and simplicity in biology: a comparative example of control algorithms Braitenberg [6] shows clearly and on strong biological grounds that behaviours which seem complex can be generated by a very simple control structure. We have implemented some ideas of Braitenberg in a real mobile robot and made an experiment proposed in [6]. The experience showed that an observer gets the impression of a complex behaviour even when simple control algorithms are implemented. The biological organisms, through the long evolutionary process, have selected very ecient ways to solve the problems that they encounter every day. In the mobile robotics eld we are faced with very similar problems. We hope to nd better solutions mimicking the computational structures of animals. To illustrate this approach we will make a comparative example of design of a mobile robot control algorithm for an obstacle avoidance and dark attraction behaviour. We will illustrate this design following three current philosophies: subsumption,

4 Francesco Mondada and Edoardo Franzi robot light sensors dark direction dark attraction motor speed robot IR sensors distances avoid obstacles k2.eps mm motor speed go forward motor speed s s motor control robot Figure 3: Structure of the subsumption control algorithm. fuzzy logic and Braitenberg's methodology (or biologically inspired methodology). A fourth solution will be given to show some limitations of the engineering design methodology. All these experiments use as input the data coming from the six infra-red (IR) proximity sensors. These sensors are composed by two parts: an IR emitter and a receiver. The emitter and the receiver are independent, so that it is possible to use the receiver to measure the reected light (with the emitter active) or to measure the light in the evironment (without emission). The reected light measurement can give some information about the obstacles. In fact, this measure is not only a function of the distance of an object in front of the emitter but also of the environment light and of the object nature. 3.1 Subsumption architecture This architecture has been proposed by Brooks [1] and has been successfully implemented on several robots of the MIT and of other institutes. The interesting and innovative aspect of this approach is the parallel processing of the several layers of the control algorithm with a characteristic interconnection scheme: The higher control modules interact with the lower ones by suppressing or inhibiting the connection present in this control module. The global task of our example has been here decomposed in two sub-tasks: to avoid obstacles and to go in the opposite direction of the light source. We have programmed these basic functions in several modules that have been interconnected as suggested by Brooks ideas. The basic functionality present in the lower block is to go forward. The motor speed message to go forward given from the \go forward" module to the \motor control" module is suppressed by the \avoid obstacle" module, taking information from the IR proximity sensors. To have a good distance measurement, the IR reected light measurements have been corrected taking into account the ambient light. The motor speed message sent by the \avoid obstacle" module appears when an obstacle is detected by the IR proximity sensors; the message turns the robot in order to avoid it. When the obstacle disappears from the sensors, the corresponding suppression signal also disappears and the robot continues to go forward. The \dark attraction" module takes the information from the ambient light measurement and acts suppressing the same communication link, but before the \avoid obstacle" module. The motor speed signal coming from the \dark attraction" module turns the robot in the direction of the dark and suppresses the go forward signal only for a short time at a given frequency. Our experience shows that the subsumption architecture makes it possible a good coordination between the several modules of the control structure. The real time concept is present at the level of the architecture and any unpredictable situation can appear. The problem rests within the modules. These little programs can be programmed in a very classical way and are the delicate point of the control algorithm. The design can be robust, but this robustness must be programmed and that is not obvious. Another critical point is that of modularity: if only a few levels exist, it is not dicult to decompose the global functionality into little modules and to intercon-

5 Biologically Inspired Mobile Robot Control Algorithms ambient light sensors reflected light sensors fuzzyfication dark light sensor value near far fuzzy inference engine k3.eps mm if... then... defuzzyfication stop back forward motor speed motors sensor value Figure 4: Fuzzy logic diagram. nect them, but for more complex tasks the communications between the modules must be studied properly. 3.2 Fuzzy logic Fuzzy logic is a mathematical branch that manipulates vague concepts and uncertain systems [11] [12]. A recent wave of commercial fuzzy products, most of them from Japan, has popularized this approach. Most of the applications are in the eld of control. The basic idea is that conventional mathematical tools are not suited to deal with ill-dened and uncertain systems, especially with those dicult to model. Fuzzy inference systems can deal with this type of problems and can model the qualitative aspects of human knowledge by means of fuzzy if-then rules. They have to capture the imprecision of the reasoning process without employing precise quantitative analysis. Moreover, they are viewed as a step toward a reconciliation between conventional precise mathematical control and human-like decision making. The fuzzy logic treatment can be decomposed into 3 principal stages: fuzzycation, fuzzy rules computation and defuzzycation. If we take, in our example, the same inputs considered for the subsumption control algorithm (six ambient light and six reected light measurements) and the same outputs (two motor speeds), we can describe the system with the diagram in gure 4. In this example a minimal number of membership functions is taken: two for the inputs and three for the outputs. For this number of inputs, outputs and membership functions a complete fuzzy inference engine needs rules! It is clear that, with this number of rules, the design of a control algorithm is impossible. For this reason it is necessary to reduce the number of inputs (the number of outputs can not be reduced), and to simplify the fuzzy inference motorengine. In our experiment we have chosen to deal with ve inputs: distance in front of the robot, left and right distances and left and right ambient lights. For every input we dened three membership functions, therefore classiying the distances and the ambient lights as big, average or small. Every motor speed is represented with ve linguistic variables: backward fast, backward slow, stop, forward slow, forward fast. Preliminary experiments have been done in order to nd a minimum number of rules needed to avoid obstacles. The minimum number observed is three, but we obtained a better behaviour with 14 rules. The nal behaviour, with obstacle avoidance and dark attraction has been implemented with 20 rules. The advantage of this approach is the easy way to describe a robot behaviour with linguistic rules. Another important aspect is the possibility to control non-linear systems. The critical problem is that the number of rules increases very fast with the number of inputs and outputs, and it is very dicult to make appropriate denitions of parameters for membership functions. Moreover, in some applications a very important factor is the big computational time needed for fuzzy algorithms. For these reasons, the fuzzy logic approach is a very interesting solution for the control of simple non-linear systems, that have a control strategy which we can easily transform into fuzzy

6 Francesco Mondada and Edoardo Franzi a) b) - + k4.eps mm + - Figure 5: Examples of wiring scheme for a) an avoidance behaviour, b) an attraction behaviour. linguistic rules. All the successful Japanese products based on fuzzy logic have only a few inputs and outputs but take advantage of the good sensors. For the control of mobile robots the fuzzy logic data treatment can be better used in submodules, like the \dark attraction" module of the subsumption-based algorithm presented before. 3.3 Braitenberg approach The third example is based on the description of the vehicles two and three of [6]. In these vehicles the sensors are directly wired to the motors. Every wire has a multiplication factor, that can be positive (excitatory) or negative (inhibitory). Depending on the kind and on the structure of the connection, the vehicle can display avoidance or attraction behaviours. Braitenberg describes in the second part of his book the biological inspiration of these wiring schemes. Using these models and the sensors of our robot, it is easy to obtain a robot that avoids obstacles and a robot that follows the dark (avoids the light). To give to the robot a basic motivation to go forward, an oset is added to the motor speed. To build the global behaviour of our example, it is sucient to combine the two networks adding the two outputs at the motor level. This composed network will produce the desired behaviour. This very simple solution based on biological observations has some very interesting advantages: the simplicity and the robustness to noise or to failure. Indeed, in the case of Khepera, equipped with 6 IR proximity sensors, the failure of one sensor can be tolerated without important behavioural changes. On the other hand, this solution is adapted to this problem and can not be easily extended. Indeed, the combination of the two networks can be a rst source of unpredictable results, depending on the importance of every sub-network. The design methodology is replaced by the biological motivation and the parameters of the algorithm are determined by trial and error. In some cases the resulting algorithm can run without a mathematical proof. 3.4 Alternative solution During the tests of the Braitengerg vehicles a better and simpler solution to the obstacle avoidance and dark attraction problem has been found. At the beginning of this section we say that the IR measurements depend on the ambient light. This non-linear property has the eect that the obstacles placed in the light seem to be closer than those placed in the dark. The standard engineering design methodology would suggests to correct this phenomenon to obtain light independent distance measurements for the obstacle avoidance sub-algorithm. The solution found during the test consists in building a Braitenberg vehicle using the reected light measurement without correction to avoid the obstacles. Doing that, the dark attraction is automatically accounted for in the algorithm by the property of the sensors of increasing the rejection of enlightened objects. The oor reecting a part of the emitted light, this tendency appears also without real obstacles. These four examples, and in particular the last one, show that some very simple and ecient solutions are not easy to nd with current design methodologies: Some very interesting non-linear properties are considered as a defect and corrected. In the biological world the solutions are the result of a long evolutionary process and try to take advantage from every property of sensors and actuators. For this reason, biological structures approach an optimum for a given problem. At the moment, the sole use of this knowledge is to understand the computational structures, copy and apply them to our robots, trying to adapt

7 Biologically Inspired Mobile Robot Control Algorithms collision detectors US Unconditioned Stimulus IR sensors or cameras k5.eps mm Motor actions avoid right avoid left go forward (default) prewired reflexes Hebbian rule Conditioned Stimulus CS Figure 6: Structure of the DAC control structure for obstacle avoidance. them to our technology. 4. Learning algorithms Many research projects have already been made in this eld and interesting algorithms have already been developed and tested by simulation. This rather theoretical investigation is very important for a better understanding of the fundamental principles of an algorithm. But, after this preliminary study, it is necessary to implement a control algorithm on a real robot. Today only few groups are successful at this level, for several reasons. One is the large amount of knowledge necessary to reach this point: theory of neural nets, programming languages, real-time control mechanisms, mechanics, electronics and optics. In the time allowed to this project it was impossible to solve all these problems. This situation and the features of the Khepera robot have triggered a fruitful cooperation with the group of Prof. Rolf Pfeifer of the University of Zurich. 4.1 The Distributed Adaptive Control The control architecture we evaluate is developed according to the design methodology of distributed adaptive control ([4], in this paper a comparison with other approaches is also made). This is derived from a distributed self-organising model of the behavioural phenomenon of classical conditioning [13] [14]. We will base our example on an agent that can learn to avoid obstacles. The basic setup of the agent is given by its value scheme [15] which can be seen as genetically predened. The value scheme denes the properties of the sensors and eectors, the morphology of the system, the structure of the control architecture, and the mechanisms for changing the properties of this structure. Moreover, it denes some basic re- exes, such as reversing and turning to the left when the system collides to the right. These re- exes consist of pre-wired relationships between primitive sensors of the system and its actions. They provide the system with a coarse adaptation to the environment. The way in which they are activated by the system-environment interaction will initially completely determine the actions the agent will execute. To enable the system to adapt to the exact properties of its interaction with the environment, the system is equipped with a distal sensor, that is with a device sensing the environment without physical contact. The integration of this sensor to the actions of the system will lead to a ne tuned adaptation of the system-environment interaction. This integration process can be seen as the development of a specic categorisation of the interactions (see [16] for a further analysis of the resulting categories). It is important to note that the control architecture dened by the value scheme must be seen as a structure with specic spatial properties. Sensors. The agent has two types of sensors: The distal sensors and the proximity ones. In the

8 Francesco Mondada and Edoardo Franzi rst experiment the former category of sensors is based on the IR signals and in the second experiment on the signal originated by the two cameras. The proximity sensors, or collision sensors, are de- ned by the saturation of the IR sensors. Actions. The agent can perform 4 actions: \turn left", \turn right", \reverse", and \advance". The avoidance actions consist of a \reverse" action followed by a turn. The default action of the system is to \advance". Control. The control architecture consists of three groups of units. The 6 units of the rst group (CS - Conditioned Stimulus) receive their input from one of the IR sensors. The continuous activation, s i, of each unit i is determined by the normalised value of the intensity of the IR light re- ected by the obstacles and detected by IR sensor r i. The units of the second group (US - Unconditioned Stimulus), which also consists of 6 units, receive their input from the collision sensors. As in the case of the CS, there is a one-to-one mapping between units of US and collision sensors. If one collision sensor is triggered, which happens when an IR sensor saturates, the corresponding unit of the US will receive 1 at its input. Together with the input from a specic primitive sensor, the units of this group also receive inputs from the CS. This input is modulated by the weights of the projections from the CS to the US. The input or local eld, h i, to the units of the US is dened as: h i = c i + NX j=1 K ij s j (1) where c i denotes the input from the proximity sensor, s j the activation of the units in the CS, and K ij the weight of the projections of the CS to the US. These weights are updated according to: K ij = 1 N ( s is j? sk ij ) (2) where N denes the number of units in the CS, the learning rate, the decay rate, and s the average activation in the group US. The quantity s introduces an active decay, such that it will only take place when other connections increase in strength. The actions are coded by a set of command units in group M which consists of three elements. The relations between the US and M are pre-wired. A collision to the left will automatically trigger a \reverse-and-turn-right" action (and symmetrically for collisions to the right). One unit in the US can trigger a unit in M. The group US can be partitioned in two clusters of units, each dependent on the unit in M to which they project: e.g. all units of US that are connected to collision sensors located to the left of the center of the agent project to the \reverseturn-right" command unit. Since the activation of all the units in the system can only have positive values, the learning takes place in the form of development of excitatory connections between the CS and the US. Moreover, the connections between US and M are excitatory. 4.2 Learn to avoid obstacles with IR sensors Robot size k10.eps mm Figure 7: Conguration of the obstacles in the rst experiment. The conguration of the obstacles used in this rst experiment is shown in gure 7. This environment (measuring cm) has been build with little pieces of wood on the table near to the computer collecting and showing the results of the experiment. The robot was connected to the computer with a cable carrying communication signals and power supply. The learning process is performed as a sequence of steps. At every step the robot reads the sensor values, updates the activation of the network, executes an action, and adapts the weights of the network. Only the default action, i.e., going forward, is continuously performed until interrupted by an event on the sensors. As a quantication of the behaviour, gure 8 depicts the amount of collisions accumulated over the time for three dierent experiments, each lasting about 10 minutes. These learning curves show an equivalent development as reported in earlier

9 Biologically Inspired Mobile Robot Control Algorithms k8.eps mm Figure 8: Total amount of collisions of the robot after a given number of steps. work on this model [4]: The robot learns the association between collision and the corresponding augmentation of the value of the proximity sensor. The robot, however, when compared with the model tested in simulation, displays a higher variability in the specics of the actions that build the trajectory; e.g., while the simulated agent will always exactly turn of the specied angle (9 ) the robot is not (and could not be) so accurate. 4.3 Experiments with a simple visual system One of the characteristics of the DAC control architecture is its independence from the exact properties of the sensors. To explore this property we have performed experiments where the distal sensors, or IR sensors, were replaced by a onedimensional visual system. The one-dimensional vision module has been designed to make it possible experiments with more sophisticated sensors [9]. It consists of two one-dimensional cameras. Every camera delivers a linear image of 64 1 pixels with a resolution of 10 bits per pixel. To obtain independency from the intensity of ambient light, a special sensor adjusts the sensitivity of the cameras. This module is added to the basic robot using the extension connectors. k14.eps mm Figure 9: Typical experimentation environment. ( c Photo H. R. Bramaz, Adliswil) Layout of the visual system. The visual system is designed to detect and dierentiate several spatial frequencies present on the horizontal vision line of the two eyes. The system consists of several layers of units. We consider the array of pixels as the input layer of the system. Every input is connected to the units of a frequencyspecic layer by connections having a particular \Mexican hat" form. This expresses the properties of a specic on-center-o-surround structure (gure 10a). These connections perform a band pass lter around a central frequency. The value of this base frequency is given by the width of the \Mexican hat" function. Four of these frequency specic layers are connected in parallel to the input layer (gure 10b). Every layer is excited by a dierent range of frequencies. The average of the

10 Francesco Mondada and Edoardo Franzi k6.eps mm Value of the connection k7.eps mm image pixels hf mhf mlf lf a) b) Figure 10: a) Connection between the input layer and the ltered layer. b) Global structure of the vision system with an example of the activation of the output units. absolute activity of the layer is proportional to the presence of these specic spatial frequencies on the one-dimensional image. One output unit per layer collects the average value. The four output units characterise the spectrum of the image, representing the presence in the image of four categories of frequencies, from high (hf) to low frequency (lf). Within one visual subsystem, camera and neural net processing, the spatial information of the input is lost: The visual system is only sensitive to certain ranges of frequencies. For instance, shifts of the stimulus on this array cannot be detected. By combining two cameras, however, the input to the system is also classied in terms of its spatial organization by detecting whether it is the left or the right camera. The network architecture is repeated for every camera. Therefore, the output of the visual system is coded by eight units. These eight outputs are used as the CS in the control architecture. The environment consists of white surfaces with regular black vertical lines. This gives the walls a particular horizontal spatial frequency Results. The general results are similar to those obtained in the previous experiments made with the IR sensors and with simulation: the system successfully integrates its visual system into its actions. To further understand the learning process we will focus on the characteristics of the weight matrix, K, implementing this integration process. The interconnectivity between the CS and the USC is represented in gure 11. The collision sensors placed at the front-left of the robot are labelled according to their position relative to the direction of displacement: 5, 45, and 85. The middle high frequency (mhf) of the left camera is associated most strongly with the USC, in this case with the unit connected to the collision sensor located at 45. Given the properties of the robot and the environment, the robot mainly has collisions at 45 or more. The angle under which the left camera now detects the stimuli is modulated towards the higher frequencies. The angle under which the right camera detects the stimuli related to the general type of collisions implies a modulation towards the lower frequencies. 4.4 Discussion Several other learning algorithms have been tested, using new architectures [17] [18] and multinetworks control systems [19]. Despite the dier-

11 Biologically Inspired Mobile Robot Control Algorithms k11.eps mm Figure 11: Interconnection matrix between the CS eld (vision system output neurones) and the left half part of the USC eld. ences in the approaches, the results for obstacle avoidance are similar. The only important dierences are in the possible extension of the structure, the necessary pretreatement of the sensory inputs and the kind of information needed by the network to perform the learning. All these aspects are important, but to solve more complex tasks the extensibility of the structure needs further study. 5. Collective behaviour The goal of this project is to study the neural networks approach in the robotics eld applied to a group of mobile robots. The experiments presented before are a preliminary study to nd good control structures that can be applied on a \social" robot. Some experiments involving ve real mobile robots have already been made to test very simple collective behaviours. The performed experiment consists in an emergent and adaptive behaviour within a group of robots. The goal is to sort objects. This method is inspired by the way ant colonies sort their brood [20]. In these experiments the individual behaviour of the robots does not describe in detail the global goal of the group. The interesting aspect of this approach is that the global behaviour observed at the group level is more complex that the one programmed on one robot. The algorithm is the following: The robots move around randomly and without communication. They can only perceive objects just in front of them. They use only three rules: First, if there is any obstacle, the robot moves randomly. Next, if the robot perceives a wall it avoids it like the Braitenberg vehicles. Finally, if it detects an object and if it does not carry an object already, it picks it up. If it carries an object, it puts it down next to the object just met. The sorting action at the group level arise in this simple mechanism: when several objects are aligned the robot has problems to perceive each object because clusters appear like walls. Therefore, the probability to pick up an object in a cluster is lower than the probability to pick up an isolated object. Moreover, interactions between the robots appear in the experiment. They pass objects to each other and they organise the choice of their work territory on their own. As a rst conclusion, the theoretical models of collective behaviours can be successfully employed with real robot to perform decentralised organisation implying simplicity, robustness, exibility. We show that explicit communications, hierarchical control and complex knowledge about the robot's environment are not necessary to produce sorting or clustering tasks. Future work will attempt the improvement of perception mechanisms and the precise study of robot group capacities. 6. Future work and research strategy Since the beginning of the Khepera development, several universities have been interested in using this robot for their experimentation. This has been made possible by the collaboration with the manufacturing company Forelec SA, Le Locle. At

12 Francesco Mondada and Edoardo Franzi the moment seven universities are experimenting on Khepera: AI Laboratory, University of Zurich (Prof. R. Pfeifer). Institut fur Programmstrukturen und Datenorganisation, University of Karlsruhe (Prof. G. Goos). Institute of Psychology, Consiglio Nazionale delle Ricerche, Roma (Prof. D. Parisi). School of Cognitive and Computing Sciences, University of Sussex (Dr. Inman Harvey). Laboratoire d'etudes et Recherche en Informatique, N^mes (Dr. Claude Touzet). Laboratoire de biochimie, Ecole Nationale Superieure, Paris (Prof. J. A. Meyer). Ecole Nationale Superieure de l'electronique et ses Applications, Paris (Prof. J.-P. Cocquerez). With all these institutions we are in good contact and collaboration is facilitated by the common experimentation platform. The work described in this paper is partly the result of this activity. In the future we will intensify these contacts and use the common experience to nd together better design methodologies and computational structures. Acknowledgements We would like to thank Andre Guignard for the important work in the design of Khepera and Paul Verschure, Claude Touzet, Jelena Godjevac, Philippe Gaussier, Stephane Zrehen, Paolo Ienne and the group of the LAMI for the help in testing the algorithms and in the development of the experimentation tools. This work has been supported by the Swiss National Research Foundation (project PNR23) and by the EPFL. Both are cordially thanked. REFERENCES [1] R. A. Brooks. A robust layered control system for a mobile robot. IEEE Robotics and Automation, RA-2:14{23, March [2] J. Heller. Kollisionsvermeidung mit fuzzylogic. Elektronik, 3:89{91, [3] U. Nehmzov and T. Smithers. Using motor actions for location recognition. In F. J. Varela and P. Bourgine, editors, Proceedings of the First European Conference on Arti- cial Life, pages 96{104, Paris, MIT Press. [4] P. F. M. J. Verschure, B. J. A. Koese, and R. Pfeifer. Distributed adaptive control: The self-organization of structured behavior. Robotics and Autonomous Agents, 9:181{196, [5] N. Franceschini, J.-M. Pichon, and C. Blanes. Real time visuomotor control: From ies to robots. In Proceedings of the Fifth International Conference on Advanced Robotics, pages 91{95, Pisa, Italy, June [6] V. Braitenberg. Vehicles. Experiments in Synthetic Psychology. MIT Press, Cambridge, [7] E. Franzi. Systeme multiprocesseur pour la commande de robots. Rapport R91.50, LAMI - EPFL, Lausanne, [8] F. Mondada, E. Franzi, and P. Ienne. Mobile robot miniaturization: A tool for investigation in control algorithms. In Proceedings of the Third International Symposium on Experimental Robotics, Kyoto, Japan, Accepted paper. [9] F. Mondada and P. F. M. J. Verschure. Modeling system-environment interaction: The complementary roles of simulations and real world artifacts. In Proceedings of the Second European Conference on Articial Life, Brussels, [10] E. Franzi. Low level BIOS of minirobot Khepera. Rapport R93.28, LAMI - EPFL, Lausanne, [11] J. Godjevac. State of the art in the neuro fuzzy eld. Rapport R93.25, LAMI - EPFL, Lausanne, [12] H.-J. Zimmermann. Fuzzy Sets Theory - and Its Applications. Kluwer Academic Publishers, [13] P. F. M. J. Verschure and A. C. C. Coolen. Adaptive elds: Distributed representations of classically conditioned associations. Network, 2:189{206, 1991.

13 Biologically Inspired Mobile Robot Control Algorithms [14] I. P. Pavlov. Conditioned Reexes. Oxford University Press, London, [15] G. M. Edelman. The Remembered Present: A Biological Theory of Consciousness. Basic Books, New York, [16] P. F. M. J. Verschure and R. Pfeifer. Categorization, representations, and the dynamics of system-environment interaction: A case study in autonomous systems. In J. A. Meyer, H. Roitblat, and S. Wilson, editors, From Animals to Animats: Proceedings of the Second International Conference on Simulation of Adaptive Behavior, [17] P. Gaussier and S. Zrehen. Emergence of behaviors on a mobile robot: Learning with neural networks. In Learning days in Jerusalem, Jerusalem, [18] P. Gaussier and S. Zrehen. A novel topological map for cognitive applications, Paper submitted to NIPS. [19] C. Touzet. Apprentissage par renforcement neuronal. In Proceedings of the Second European System Science Congress, Prague, Accepted paper. [20] J. C. Deneubourg, S. Goss, N. Franks, A. Sendova, A. Franks, C. Detrin, and L. Chatier. The dynamics of collective sorting: Robot-like ant and ant-like robot. In J. A. Mayer and S. W. Wilson, editors, Simulation of Adaptive Behavior: From Animals to Animats, pages 356{365. MIT Press, 1991.