Robot Security and Failure Detection Using Bayesian Fusion ⋆ F. Aznar, M. Pujol, and R. Rizo Department of Computer Science and Artificial Intelligence, University of Alicante {fidel, mar, rizo}@dccia.ua.es Abstract. This paper shows a Bayesian framework for fuse informa- tion. Using this framework we present a robotic system, based on two processing units. The system is used for the development of a task, done by an autonomous agent, arranged in an environment with uncertainty. This agent interacts with the world and is able to detect, only using its sensor readings, any failure of its sensorial system. Even it can continue working properly while discarding the readings obtained by the erroneous sensor/s. A security unit is also provided to make the system even more robust. The Bayesian Units brings up a formalism where implicitly, us- ing probabilities, we work with uncertainly. Some experimental data are provided to validate the correctness of this approach. Keywords: Bayesian Fusion, Failure Detection, Reasoning Under Uncertainty, Autonomous Agents, Robotics. 1 Introduction When an autonomous agent is launched into the real world there are several problems it has to face. The agent could have a model of the environment rep- resenting the real universe where it will interact. Nevertheless, it is necessary to bear in mind that any model of a real phenomenon will always be incomplete due to the permanent existence of unknown, hidden variables that will influence the phenomenon. The effect of these variables is malicious since they will cause the model and the phenomenon to have different behavioural patterns. In this paper a new fusion model, the Bayesian units, is shown. This model is used to specify the system architecture and determine how the information is fused. The proposed system is composed by two units. The first is used to model the environment for an autonomous agent as the Bayesian Maps [4] formalism. The second is used to send the commands to the robot while verifying its security. The autonomous agent will develop a generic task working with uncertainly. Also, a method of obtaining sensor reliability in real time using an abstraction of various Bayesians maps will be defined. Next, the described models will be applied to a real robot. Finally, conclusions and future lines of investigation to be followed will be highlighted. ⋆ This work has been financed by the Generalitat Valenciana project GV04B685. S. Bandini and S. Manzoni (Eds.): AI*IA 2005, LNAI 3673, pp. 518–521, 2005. c  Springer-Verlag Berlin Heidelberg 2005