Breedbot: an evolutionary robotics application in digital content
The Authors
Orazio Miglino, Department of Relational Sciences “G.Iacono”, University of Naples “Federico II”, Naples, Italy and Laboratory of Autonomous Robotics and Artificial Life, Institute of Cognitive Sciences and Technologies, National Research Council, Rome, Italy
Onofrio Gigliotta, Laboratory of Autonomous Robotics and Artificial Life, Institute of Cognitive Sciences and Technologies, National Research Council, Rome, Italy and Department of Psychology, University of Palermo, Palermo, Italy
Michela Ponticorvo, Department of Relational Sciences “G.Iacono”, University of Naples “Federico II”, Naples, Italy
Stefano Nolfi, Laboratory of Autonomous Robotics and Artificial Life, Institute of Cognitive Sciences and Technologies, National Research Council, Rome, Italy
Abstract
Purpose – This paper aims to describe an integrated hardware/software system based on evolutionary robotics and its application in an edutainment context.
Design/methodology/approach – The system is based on a wide variety of artificial life techniques (artificial neural networks, genetic algorithms, user-guided evolutionary design and evolutionary robotics). A user without any computer programming skill can determine the robot's behavior in two different ways: artificial breeding or artificial evolution. Breedbot has been used as a didactic tool in teaching evolutionary biology and as a “futuristic” toy by several science centers. The digital side of Breedbot can be downloaded on the web site: www.isl.unina.it/breedbot
Findings – The results in this pilot study suggest that using Breedbot in an educational context can be useful to improve learning in biology.
Research limitations/implications – As this is a pilot study, one limitation is the small sample considered. The issue will be investigated further with a wider population and subject-matter, which will also improve the Breedbot system.
Practical implications – These results suggest that tools like Breedbot could be introduced into biology curricula at schools.
Originality/value – The paper describes an original application in digital content and shows the importance of using such a tool in an Edutainment context. It is therefore interesting for teachers, vocational trainers and anyone involved in educational activities.
Article Type:
Research paper
Keyword(s):
Robotics; Evolution; Education; Entertainment.
Journal:
The Electronic Library
Volume:
26
Number:
3
Year:
2008
pp:
363-373
Copyright ©
Emerald Group Publishing Limited
ISSN:
0264-0473
Introduction
Nowadays, new-generation robots are widely used in Edutainment. There is a niche market that is fed by big and small industries which produce and sell didactic robotics kits. Between the many products available, a good example of this category are the Lego MindStorm kit, in the new version NXT(http://mindstorms.lego.com/) that is conceived mainly for schools and the small robot Surveyor (http://www.surveyor.com/) that is oriented to university and research centre. Together with these kits, there are some robots that can be imagined as artificial pet companions: famous examples of this product line are Furby robots produced by Tiger Electronics (www.hasbro.com/furby/) and Paro (http://paro.jp/english/), a robot that is used also for cognitive rehabilitation. The main part of these robots presents a hardware structure that is sometimes very complex (formed by sensory apparatus with cameras, infrared sensors and ultra-sounds etc; by actuators as wheels, tracks, mechanical limbs and by powerful on-board computers), but this complexity is not counterbalanced on the behavioral side, as the behaviors they show, on the contrary, are quite rudimentary. Substantially we can observe in present edutainment robots a remarkable distance between their “bodies” sophistication (hardware) and poverty of their “minds” (Artificial Intelligence software underlying behaviors production).
The growth of this niche sector depends just on the individuation of methodologies and approaches that can reduce the gap between bodies and minds. In this work we describe our prototype by which we try to apply artificial life techniques to robotics for Edutainment. In particular, it seems to us very promising to use the physical robot as a bridge between digital worlds created by artificial life and physical environments in which we and, obviously, the robots act. As it is known artificial life (Langton, 1995) is a discipline that tries to simulate in a digital world (software) the functional and structural features of biological systems. This discipline uses various techniques, models and methodologies such as genetic algorithms, evolutionary computation, artificial neural networks, etc. that, in the framework of the A-life approach to technology, are used to build up digital objects by breeding, training or evolving them.
Many applications are proposed in the context of User Guided Evolutionary Design (Bentley, 1999) where a breeding digital environment allows users to “evolve” novel artificial objects. The mechanism is simple: on a computer screen a set of variants of the object users are seeking to evolve is proposed between which users can choose a subset of variants. The system then produces a new set of variants. The selection/production process proceeds until users have obtained the desired objects. There are applications in the field of evolution of artistic images (Sims, 1991), training human face profiles for crime investigations, objects design for furniture firms, etc. Most of those systems produce software images for static objects (compare for artistic images Artificial Painter: Pagliarini and Lund, 2000; Lund et al., 1995). Moreover, within the broader field of artificial life, researchers are developing models and techniques to evolve actual robots (Nolfi and Floreano, 2000).
Our system, Breedbot, is built up by integrating the above cited A-life-like methodologies in order to permit children to evolve interesting mobile robots behaviors in an digital environment that can then be downloaded and tested on real robots. We attempted to try to avoid any form of “information processing methodology” (i.e.: programming) for children using hi-tech (i.e.: robots). This way the A-life techniques that are used are, to a certain extent, invisible for children who use it. Breedbot simulates the evolution of populations of digital agents (simulated robots) with elementary navigational capabilities. Based on artificial life and evolutionary robotics techniques, the digital environments allows the user to select agents with efficient behavior. This is done through a user- guided genetic algorithm (UGGA) that permits to select robots that are controlled by a neural network (i.e. the model of the brain of an artificial pet). For the user-guided genetic algorithm the basic idea consists in showing the ongoing process (robots behaviors) on a computer screen (digital environment) and allowing the user to judge the robots global behavior, as a teacher or a breeder does in everyday life. In the following sections we describe the Breedbot system in detail. In section 4 we present some preliminary applications while section 5 is devoted to identify Breedbot limits and future directions for its improvement.
Breedbot: an environment for breeding robots
Breedbot is an integrated hardware/software system that allows users with no technical or computer experience to breed a small population of robots. We will now describe Breedbot hardware and software in detail.
The actual robot (hardware)
The robot is shown in Figure 1. It is built using motors, infra-red sensors, bricks and an on-board computer from the Lego Mindstorms© kit. It is rectangular in shape. Its base measures 16*15 cm and it is 10 cm high. To move, it uses two large drive wheels, each controlled by its own small electric motor. Two wheels provide stability. All the wheels are fixed so there is no steering mechanism. The on-board computer (a Lego Mindstorms RCX) and the electric power supply are located on top of the motors. The sensor system – three Mindsensor infrared sensors – is placed above the on-board computer. The first sensor is mounted half way along the robot's short side and points in its direction of motion. The other two are fixed half way along the long sides. Each sensor produces a signal with a strength inversely proportional to its distance from an obstacle. The sensors can detect obstacles up to a maximum range of 15 cm (Figure 2).
Breeder tools (software)
On software side, Breedbot uses a digital environment to simulate a process of artificial evolution that allows users with no technical or computer experience to breed a small population of robots. At the beginning of each simulation, the computer screen shows a first generation of robots in action. After a certain time, some of the robots are selected to produce offspring. Users can let the system select the “best robots” or make the decision themselves. If the system makes the decision, it rates the robots by their ability to explore the environment, and selects those with the highest scores. Human users, on the other hand, simply choose the robots they think have performed best. Once the selection procedure is over, the system creates clones of the selected robots. During this process it introduces random mutations into their control systems. The robots created in this way constitute a new generation. This selection/cloning/mutation cycle can be iterated until the “breeder” finds a particularly capable robot. At this point the brain of the simulated robot can be uploaded to a real robot and the user can see how it performs in the real world. Figure 1 shows a robot which has just received a “brain” developed with Breedbot. Figure 3 shows Breedbot's graphical interface.
On the left side there are the nine robots while they explore a rectangular arena with walls. On the right side there is a display that indicates the current generation, radio buttons to choose between human and machine selection, STOP and RESET buttons and a graph, updated after each generation, that shows the mean fitness obtained in the exploration task along generations. The left hand side of the screen displays the behavior of the nine simulated robots in the arena, which is surrounded by walls. The right hand side provides information about the state of the system (the number of the current generation, a graph showing changes in the mean fitness of the population) along with a number of commands allowing the user to stop the system and to choose between human and artificial selection. The pull-down menu in the top left corner contains system utilities (to change the geometry of the environment, save configurations etc.). Breedbot is designed to be easy to use for breeders of small mobile robots. Breeders can use the system's graphical interface to organize their own experiments in artificial evolution and if they want, they can select the individuals which will be allowed to reproduce. They can stop the program at any time, choose what they consider to be a well-adapted robot, and use the infrared port to upload its control system (its artificial neural network) to a real Lego MindStorms robot (see Figure 4).
Artificial life techniques in Breedbot
Up to now, we have described the system without considering what is going on “inside”, an aspect that is not directly perceivable by users. In this section we will open this black box to describe the Artificial Life techniques that are employed in the system.
The on-board computer on the hardware side and the Breedbot simulator on the software side, implement an artificial neural network that is the system that controls the robot. The network, a simple perceptron, receives sensory stimuli from the infra-red sensors, processes the data and activates the robot's motors. The system's neural architecture consists of a layer of input neurons and a layer of output neurons (see Figure 2). The input neurons receive stimuli from the sensors and transmit these signals, through one-directional links (“connections”) to the output neurons. Each connection is associated with a transfer value (its “weight”). This way, the signal arriving to the output neurons is filtered by the weights of the connections from neurons in the input layer. The input layer is made up of three sensor neurons, two proprioceptor neurons and two bias neurons. Each infrared sensor is associated with a single sensor neuron which receives its signal and activates the rest of the network. The two proprioceptor neurons have recurrent connections from the motor neurons (see Figure 2). Thus the state of these neurons at time t+1 reflects the state of the motor neurons at time t. Finally, the bias neurons are always “on” (they always have an activation of 1). These neurons, which do not receive any kind of signal from the external environment, play an essential role, ensuring that the robot is always able to move, even when receives no input from the sensors. The output layer consists of two motor neurons: these neurons determine the robot's behavior at any given moment. Each motor neuron controls an electric motor. Its output is regulated by a threshold activation function. If the sum of the inputs to the neuron is equal to or higher than the threshold the neuron produces an output of “1”. For values below the threshold, the output is 0. When a motor receives a “1” it turns clockwise for 2 seconds. When it receives a “0” it does nothing. In this way, the robot has three possible behaviors: it can move forward for 3 cm (when both motors are on); it can turn 10 degrees to the left (when the right motor is on and the left motor is off); it can turn 10 degrees to the right (when the right motor is off and the left motor is on).
The software simulator replicates the physical characteristics of the robot and the training arena. Using the simulator we can conduct artificial evolution experiments with a population of 9 simulated robots. This aspect is based on genetic algorithms and evolutionary robotics. In terms of size, sensors, motors, and neural architecture, each individual in the population is identical to all the others. Only the weights of the connections in their control systems distinguish them. These are stored in their “genotypes”. When breeding begins (the first generation of robots), the weights of the connections are extracted randomly from a uniform distribution in the range −1, 1. For a certain time, the robots are allowed to move freely. Then the “breeder” (either the system or a human being) selects three robots for reproduction. Each robot's genotype (the values of its connections) is cloned three times, producing three offspring. But the clones are not perfect. During the copying process 3 per cent of the weights “mutate”. The choice of which weights modify and the new value of the modified weight are random.
Breedbot edutainment application
The preliminary application of Breedbot has regarded two main ambits, both concerning Edutainment. First of all Breedbot has been used in various exhibition for scientific divulgation in Italy, such as the Festival of Science in Genoa and in Rome and Futuro Remoto in Naples. In these demonstrations, Breedbot has been used by many children that could use both the software and the hardware of Breedbot prototype. Just relying on qualitative observations, we noticed that the joint use of the digital content and the physical robot led a better understanding of the themes that were meant to introduce.
For this reason we have run a pilot experiment (Miglino et al., 2004) in a school to verify if using software and hardware together could produce better results in learning. It is, in fact, well-known that computer simulation can be a powerful teaching tool, because they allow the learner to manipulate the most important variables involved in a process and observe what are the results (Milheim, 1992; Pappo, 1998; Towne, 1995; Jonassen, 2006; Pan et al., 2006; Singh, 2003). It is not very much investigated if the use of simulation (digital content) and robots can enhance learning, so we addressed this issue running an experiment in Italy, that we describe in the next section.
Introducing evolutionary biology through Breedbot
In school curricula there is a wide attention to Darwinian theory. This is justified, on one side, by the importance of the theory itself, and, on the other side, on the importance this theory holds in opposing the anti-scientific creationist theories on the origin of Universe. A strong obstacle to this educational mission consists in the impossibility to run actual experiments on evolution, as the time-scale involved in biological evolution makes it impossible to apply the traditional methods of experimentation and observation that are widely used in other scientific disciplines. The simulative method can help us in this enterprise: an appropriate software including all relevant variables, may allow an user to manipulate these variable (for example number of individuals in a population, sexual or asexual reproduction, genotype to phenotype mapping etc.), thus providing as useful tool to teachers. With Breedbot, thanks to the underlying Evolutionary Robotics methods, teachers and students can also “soil” their hands with the physical matters related on evolution, exploring, for example, how relevant is a particular physical structure to solve a problem. This way we aimed at verifying if transferring the digital content (about Biology) to a physical environment (through Breedbot) might be useful for instruction about Biology.
Materials and methods
We have designed, together with biology teachers, a lesson supported by Breedbot. We have concentrated our didactic program on the concepts of creationism vs evolutionism, Lamarckian vs Darwinian evolution, selection mechanisms, artificial vs natural selection, the role of the environment in the genotypic-phenotypic mappings. In particular, we have concentrated our efforts on making an incremental approach with some experiments that showed basic principles of evolutionism. In order to have some data about the efficiency of our approach we have compared the learners' performances in two cases: one in which the users interact with a traditional multimedia hypertext (the control group), and one in which the users interact with the Breedbot (the experimental group).
Subjects
Two groups, each consisting of 22 Italian high school students aged 14-15 years old, were observed in the acquisition of the Darwinian evolutionary theory. Both groups attended a normal lesson about evolutionary biology given by their teacher. After the lesson, a group used a multimedia hypertext (the control group); the other group (the experimental group) used a suite of artificial life software. In order to quantify the acquisition of the basic notions about evolutionary biology, we asked the subjects to fill a multiple-choice questionnaire. The questionnaire was distributed before and after the lesson, and after the software session.
The questionnaire
The questionnaire was developed together with biology teachers in order to measure the students' competence in basic concepts of evolutionary biology. The questionnaire has 14 items with four different answers, of which only one is correct.
The lesson
The lesson lasted half an hour and was a standard unit of a biology course of Italian high-school curriculum. The teacher first illustrated non-scientific theories about the origin of life (creationism and fixism) and the Lamarckian and Darwinian evolutionary theories. In this framework, the students were introduced to the concept of genotype, phenotype, selection mechanisms, and the rule of environment in the genotype-phenotype mappings.
The hypertext
We used a commercial hypertext edited in Italy (Montateti, 1998) that has a wide distribution in the schools. The hypertext is dedicated to illustrate the Darwinian evolutionary theory with detailed text description, maps, pictures and some animations. The argument is divided in several sections. Some of them explain the historical background and the socio-cultural context where Darwinian thought grew up while the others treat the basic notions about the Darwinian vision of species origin and human beings origin. More in-depth, there are seven sections:
- “Darwin's Life”. In this section the hypertext shows a map of Great Britain where several active links take the user to discover the most important places, facts and dates of Charles Darwin's life.
- “Characters”. The hypertext shows texts and photos of a consistent number of people, friends and competitors, that were very present in Darwin's life.
- “Evolutionism”. The hypertext reports about all the scientific theories related to the modern concept of evolution existing before Darwin (i.e. from Lamarck to the early 1900s).
- “The Journey”. An accurate map full of texts and photos describe the famous journey of the “Beagle” reporting, step by step, all Darwin's observations along the way.
- “The Beagle”. A virtual tour inside the ship that Darwin used for his research journey.
- “The Origin of the Species”. The hypertext presents to the user the concepts of evolution and natural selection trying to explains it with the help texts, graphs and images. Successively, the Darwinian and Lamarckian theories are compared.
- “The Origins of the Human”. In this section, thanks to a graphical animation, all the path, the genealogic tree, from simple organism to human being is shown.
The experimental group could use the Breedbot system, both software and hardware.
Results
The subject performance during the learning process is shown in Figure 5. Both groups start with the same level of knowledge, as can be seen in Figure 5. An ANOVA test comparing the two groups' results of the questionnaire before the lesion showed no significant difference. After the software/Breedbot sessions both groups improve their competence on the theme: there is a significant difference confronting the results before and after the lesson (F 1,21=4.68, P <0.05 for the “Hypertext” condition, and F1,21=4.89, P<0.05 for the “Artificial life” condition). There is also a significant difference between the two groups (control versus experimental) at the end of the learning process. The “Breedbot” user higher marks than the “Hypertext” users (F 1,21=5.78, P<0.05).
The results of this pilot study indicated that the use of Breedbot system, both software and hardware, improves learning (in the present case about a curricular subject such as Biology).
This comparative study shows that the students increase their knowledge about evolutionary theory by using Breedbot system, and also that both tools (hypertext and Breedbot) produce a statistically significant increment on learning performances (i.e.: scores in a questionnaire), and this effect is stronger in the case of Breedbot users.
Conclusions and future directions
The preliminary results presented above suggest that the use of an integrated software-hardware system as Breedbot can be very useful in Edutainment. What we would like to stress is that its use allows to focus on the gap-link between digital and real world. As the process runs in simulation and can be transferred in reality, it makes it possible for users to ask about the limits and potentials both in simulation and in reality. This is, of course, just a suggestion: the issue will be faced up in more detail in the next investigating efforts. Another issue we aim at tackling regards the extension of the Breedbot system. We now intend to improve the system to handle with different robots. In particular it will be particularly interesting to try the system with the Surveyor robot that is quite different in size and shape and is provided with different actuators. The robot built with Lego Mindstorms Kit has two wheels, while Surveyor is a crawler, that means that it has tracks to move. This modifies completely the robot interaction with the external environment, thus underlying, once again, the importance of transferring digital agents in real world.
Figure 1A picture of the robot built with the Lego Mindstorms kit
Figure 2A schematic representation of the robot and of its control system
Figure 3A snapshot representing Breeder Tools (software)
Figure 4The transfer of the control system (an artificial neural network) from the digital environment to the real Lego MindStorms robot, through the infrared port
Figure 5The graphs show the average performance of the control group (solid line) and the experimental group (dotted line) across time
References
Bentley, P. (1999), Evolutionary Design by Computers, Morgan Kaufmann Academic Press, San Francisco, CA, .
Jonassen, D.H. (2006), Modeling with Technology: Mindtools for Conceptual Change, Merrill/Prentice-Hall, Columbus, OH, .
Langton, C.G. (1995), Artificial Life: An Overview, The MIT Press/A Bradford Book, Cambridge, MA, .
Lund, H.H., Pagliarini, L., Miglino, O. (1995), “The artificial painter”, Proceedings of ICANN 1995 International Conference on Artificial Neural Network, 5-13/6/95, Granada, Spain, 5-13 June 1995, .
Miglino, O., Rubinacci, F., Pagliarini, L., Lund, H.H. (2004), "Using artificial life to teach evolutionary biology", Cognitive Processing: International Quarterly of Cognitive Science, Vol. 5 pp.123-9.
Milheim, W.D. (1992), Computer-Based Simulations in Education and Training: A Selected Bibliography, Educational Technology Publications, Englewood Cliffs, NJ, Selected Bibliography Series No. 8, .
Montateti, G. (1998), Charles Darwin, Editori Riuniti, Milano, .
Nolfi, S., Floreano, D. (2000), Evolutionary Robotics. The Biology, Intelligence and Technology of Self-Organizing Machines, MIT Press, Cambridge, MA, .
Pagliarini, L., Lund, H.H. (2000), "Art robots and evolution as a tool for creativity", in Bentley, P., Corne, D. (Eds),Creative Evolutionary Systems, Morgan Kaufmann, San Francisco, CA, .
(2006), in Pan, Z., Aylett, R., Diener, H., Jin, X., Göbel, S., Li, L. (Eds),Technologies for E-Learning and Digital Entertainment, 1st International Conference, Edutainment 2006, Hangzhou, China, 2006, Proceedings. Lecture Notes in Computer Science 3942, Springer, New York, NY, .
Pappo, H.A. (1998), Simulations for Skills Training, Educational Technology Publications, Englewood Cliffs, NJ, .
Sims, K. (1991), "Artificial evolution for computer graphics", Computer Graphics, Vol. 25 No.4, pp.319-28.
Singh, H. (2003), "Building effective blended learning programs", Educational Technology, Vol. 43 No.6, pp.51-4.
Towne, D.M. (1995), Learning and Instruction in Simulation Environments, Educational Technology Publications, Englewood Cliffs, NJ, .
About the authors
Orazio Miglino is Full Professor at the University of Naples “Federico II” and holds a research position at the Institute of Sciences and Technology of Cognition (previously Institute of Psychology) of the National Council of Research of Rome. His research activity is mainly concentrated on Cognitive Science and Artificial Life. In particular, his interest is oriented toward construction of formal model based on Neural Networks and mobile robots that simulate cognitive and adaptive processes of natural beings, such as orientation and spatial navigation.
Onofrio Gigliotta is PhD student in Psychology at University of Palermo and research assistant at LARAL (Laboratory of Artificial Life and Robotics) at the Institute of Sciences and Technology of Cognition of the National Council of Research of Rome. His research is focused on evolutionary robotics, language evolution, cognitive psychology, artificial life, neural networks, genetic algorithms, agent based simulations, embodied communication, social simulations, and edutainment.
Michela Ponticorvo studied Psychology and holds a PhD in “Psychology of Programming and Artificial Intelligence” and has a scholarship at the Department of Relational Sciences, University of Naples “Federico II”. Her research interest is mainly in the interdisciplinary study of brain, behavior and evolution through computational models ranging from Artificial Intelligence (Artificial Life, Evolutionary Robotics, Artificial Neural Networks, Genetic Algorithms) to Psychology: sensory-motor co-ordination, and spatial orientation. She is the corresponding author and can be contacted at: michela.ponticorvo@unina.it
Stefano Nolfi is the head of LARAL (Laboratory of Artificial Life and Robotics) at the Institute of Sciences and Technology of Cognition of the National Council of Research of Rome where he coordinates the European Integrated Project ECAgents: Embodied and Communicating Agents. He has been a fellow of: Centre for Research in Language, University of California, San Diego, USA; Laboratory of Microcomputing, Swiss Federal Institute of Technology (EPFL), Lausanne, Switzerland; SONY Computer Science Laboratory, Tokyo, Japan; Institute of Advanced Studies of Berlin, Germany; University of New South Wales, Canberra, Australia. His research interests are in the field of neuroethological studies of adaptive behavior in natural and artificial agents and include: Evolutionary robotics, Artificial life, Complex systems, Neural networks, Genetic algorithms.