Fault-Tolerant Certainty Grid W. Elmenreich Vienna University of Technology Institut f¨ ur Technische Informatik Treitlstraße 1-3/182-1 wil@vmars.tuwien.ac.at Abstract World modelling for mobile autonomous robot is usually a process that uses sensor data as input and provides a model of the robot’s environment as output. In this paper we investigate on sensor fusion methods for robustness and fault tolerance. We evaluate three methods according to their performance, memory con- sumption, and required sensor configurations. The algorithms have been implemented in a four- wheeled autonomous mobile robot that uses a set of three infrared sensors to build the world model. We present a performance analysis of the new algorithms based on simulation and experimental data. 1 Introduction Autonomous mobile robots belong to a class of ap- plications whose operability depends on sensor data. Sensor, as physical devices converting a physical prop- erty into a measurement that can be interpreted by a computer system, are affected by several sources of error, like sensor deprivation, limited spatial or tem- poral coverage, imprecision, cross-sensitivity, and un- certainty. Thus, dependable applications may never depend on a single sensor. In order to overcome these problems, the inputs from several sensors are com- bined to form a dependable representation of the en- vironment, the world model. For mobile robots, usually the world model is repre- sented as a two-dimensional or, especially for non-flat outdoor environments, a three-dimensional map of the robots surrounding. Typical approaches to generate such a map are the grid-like division of the environ- ment, like the occupancy grid approach of Elfes [2]. Based on this grid, the navigation and path planning decides on the robot’s actions. Erroneous or ambigu- ous sensor readings are detected and solved by pro- cessing redundant sensor information. Redundant in- formation can be evaluated at different levels of ab- straction. In general, if the evaluation is performed at sensor level, the system complexity can be kept low at the cost of hardware expenses. If evaluation of re- dundant information is made at application level, the performance is better, since at application level has more knowledge about the reasonableness of a par- ticular result. However, system complexity increases since normal processing functions become intertwined with error-detection and fault-tolerance functions [7]. It is the objective of this paper to investigate on sensor fusion methods for robustness and fault toler- ance for a grid-like representation of a mobile robot’s world model. The remainder of the paper is organized as follows: Section 2 describes the system architecture of a sen- sor grid application and examines the possibilities and benefits of applying sensor fusion at particular levels in this model. Section 3 presents certainty grid al- gorithms that can handle faulty measurements. Sec- tion 4 presents the results from the evaluation, while Section 5 discusses the results. Section 6 concludes the paper. 2 Architectural Considerations An autonomous robotic system contains at least a set of sensors and actuators and a control application. Sensors and actuators are the interface to the process environment and belong to the transducer level of the robotic system, while the control application belongs to the control level. In general, a system is composed out of compo- nents, whereas components are subsystems like sen- sors, actuators, processing nodes, and communication channels. As a matter of fact, every single component of a system will eventually fail [1]. Requirements for highly dependable systems can only be met, if these failures are taken into account. Some systems, fail in a manner that they still pro- vide a service, however at a degraded level. For ex- ample, a sensor may be affected by cross-sensitivity and fail to render a measurement with the specified accuracy – however, the measurement contains still information that can be exploited. In the following sections we focus on sensor fusion 1576 Proceedings of ICAR 2003 The 11th International Conference on Advanced Robotics Coimbra, Portugal, June 30 - July 3, 2003