Learning to Be a Good Tour-Guide Robot J. Javier Rainer, Ramon Galán Intelligent Control Group. Universidad Politécnica de Madrid C/ José Gutiérrez de Abascal 2. 28006, Madrid, SPAIN e-mail: javier.rainer@upm.es ; ramon.galan@upm.es Abstract—Thanks to the numerous attempts that are being made to develop autonomous robots, increasingly intelligent and cognitive skills are allowed. This paper proposes an automatic presentation generator for a robot guide, which is considered one more cognitive skill. The presentations are made up of groups of paragraphs. The selection of the best paragraphs is based on a semantic understanding of the characteristics of the paragraphs, on the restrictions defined for the presentation and by the quality criteria appropriate for a public presentation. This work is part of the ROBONAUTA project of the Intelligent Control Research Group at the Universidad Politécnica de Madrid to create "awareness" in a robot guide. The software developed in the project has been verified on the tour-guide robot Urbano. The most important aspect of this proposal is that the design uses learning as the means to optimize the quality of the presentations. To achieve this goal, the system has to perform the optimized decision making, in different phases. The modeling of the quality index of the presentation is made using fuzzy logic and it represents the beliefs of the robot about what is good, bad, or indifferent about a presentation. This fuzzy system is used to select the most appropriate group of paragraphs for a presentation. The beliefs of the robot continue to evolving in order to coincide with the opinions of the public. It uses a genetic algorithm for the evolution of the rules. With this tool, the tour guide-robot shows the presentation, which satisfies the objectives and restrictions, and automatically it identifies the best paragraphs in order to find the most suitable set of contents for every public profile. Keywords- Cognitive systems; learning; autonomous robot; fuzzy systems; decision making. I. INTRODUCTION Autonomous robots are intelligent machines capable of performing tasks in the world by themselves, without explicit human control over their actions [1]. Within the development of multiple applications for a mobile robot, probably one of the first real world applications of indoor service robots has been mobile robots serving as tour guides in museums or exhibitions. We have developed our own interactive mobile robot called Urbano specially designed to be a tour guide in exhibitions [2]. The basic characteristics of Urbano are described in Section 3. Our goal has been to create an automatic presentation generator that allows the flexible and dynamic display of information depending on the distinct kinds of audiences and other parameters that characterize the presentation. The automatic selection of contents for composing sophisticated presentations is a non-trivial task. If the aims of the presentation, preference and interest for a particular subject have to be taken into account, it becomes even more complex. The knowledge is based on an ontology of domain- specific concept words. Ontologies have been known in computer science as consensual models of domains of discourse, usually implemented as formal definitions of the relevant conceptual entities [3]. Some of the most recent works about automatic generation are: [4,5,6,7,8]. These works propose different architectures and methodologies than those presented here. What is original is the introduction of fuzzy logic in the pruning of the resulting tree and in the quality index of the presentations. This paper is focused on the automatic generation of presentations by a robotic system. Aspects concerning the grounding of the symbols used by the robot, e.g., paragraphs in the presentation or the learning of new concepts – have been circumvented for problems of space. In [9], the authors explore the problem of learning and the symbol-grounding problem, and propose a systemic and integrative approach both problems. II. AUTOMATIC TEXT GENERATION. The history of natural language processing is characterized by the influence of artificial intelligent [10,11]. A natural language generator generally has access to a wide knowledge field from which it must select information to present to the users in various formats. Generating text is, therefore, a decision-making problem with multiple restrictions: knowledge, available linguistic tools, the objectives of the user to whom the text will be directed, the situation, and past discourse. It deals with identifying the factors involved in this process and determining the factors that intervene in this process and its consequences [11,12,13]. An additional contribution of our work is to provide a dynamic framework that allows us to take into account each possible scenario. On the other hand, there can be no single general-purpose presentation format for all users, because each user differs in all aspects of interests and expertise levels, and in the devices used to visualize the requested information. These aspects are used as restrictions. As a result, to prepare a presentation, the candidate items can be multiple, depending on knowledge server. The items 2010 Ninth International Conference on Machine Learning and Applications 978-0-7695-4300-0/10 $26.00 © 2010 IEEE DOI 10.1109/ICMLA.2010.92 595