164 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 29, NO. 2, APRIL 1999 Learning Sensor-Based Navigation of a Real Mobile Robot in Unknown Worlds Rui Ara´ ujo and An´ ıbal T. de Almeida Abstract—In this paper, we address the problem of navigating an autonomous mobile robot in an unknown indoor environment. The parti-game multiresolution learning approach [22] is applied for simultaneous and cooperative construction of a world model, and learning to navigate through an obstacle-free path from a starting position to a known goal region. The paper introduces a new approach, based on the application of the fuzzy ART neural architecture [7], for on-line map building from actual sensor data. This method is then integrated, as a complement, on the parti-game world model, allowing the system to make a more efficient use of collected sensor information. Then, a predictive on-line trajectory filtering method, is introduced in the learning approach. Instead of having a mechanical device moving to search the world, the idea is to have the system analyzing trajectories in a predictive mode, by taking advantage of the improved world model. The real robot will only move to try trajectories that have been predicted to be successful, allowing lower exploration costs. This results in an overall improved new method for goal- oriented navigation. It is assumed that the robot knows its own current world location—a simple dead-reckoning method is used for localization in our experiments. It is also assumed that the robot is able to perform sensor-based obstacle detection (not avoidance) and straight-line motions. Results of experiments with a real Nomad 200 mobile robot will be presented, demonstrating the effectiveness of the discussed methods. Index Terms—Learning systems, mobile robot navigation. I. INTRODUCTION R OBOT programming and control architectures must be equipped to face unstructured environments, which may be partially or totally unknown at programming time. Two of the most important tasks for autonomous navigation of a mobile robot are path planning, and building a world model. In this paper, we assume that there is no predefined world model, initially available to a mobile robot, and the robot path finding problem is considered. In this problem, a robot path avoiding collisions with obstacles must be generated, so that the system moves from a starting location to a goal region in the world. A general learning approach is considered where the robot integrates the abilities to concurrently construct a world model and learn a path to the goal. A learning system brings the benefit of being based on a concept of self programming, in which control of a complex system in principle does not need extensive analysis, modeling, programming, assistance, or teaching by human experts. Instead, the system acquires its competencies. Manuscript received February 10, 1997; revised April 5, 1998 and August 15, 1998. This paper was recommended by Associate Editor R. A. Hess. The authors are with the Institute for Systems and Robotics (ISR), and the Electrical Engineering Department, University of Coimbra, P´ olo II, P-3030 Coimbra, Portugal (e-mail: rui@isr.uc.pt; adealmeida@isr.uc.pt). Publisher Item Identifier S 1083-4419(99)02302-X. A widespread approach for mobile robot navigation is based on the occupancy grid representation of the environment [23]. Occupancy grids represent the world as a two-dimensional array of evenly spaced cells, with each cell holding a value which represents the confidence in whether it is occupied space or free space. Although grid-based models are easy to build and maintain, they impose a constant resolution structure onto the environment without any selectivity concerning the nature and clutter of the world. A very localized feature of the world may impose a very high (constant-) resolution grid over the entire state-space. This implies high data requirements, and induces excessive detail on world modeling and updating, on reasoning (high computational costs), and on the paths that result from such a model. Also, the difficulties on the direct application of grid-based models on localization have been pointed out in [28]. An alternative for overcoming the space and time complexities of grid-based methods is to use a vari- able resolution state-space partition (e.g., [17] and [31]). Local resolution is usually only high enough to capture the important local detail of the world. This enables a lower number of cells (space) and thus lower search effort (time). Another alternative to the costs of grid-based models is to use a set of geometric primitives for representing objects in the world (e.g., [16] and [28]). Geometric primitive representations, have been difficult to build, but are significantly more compact, less complex, and fully applicable to high- and low-level motion planning (e.g., Section IV) and localization approaches (e.g., [28]). With higher dimensions the geometric model data requirements be- come exponentially smaller than the requirements of constant- resolution cellular models. There are other approaches that do not require a map at all, but also do not provide a resulting world model. For example [4] proposes a reactive approach that is based on a fuzzy system that learns to coordinate dif- ferent control strategies or behaviors, that must be predefined and programmed. In [11], a preprogrammed reactive system with set of basic behaviors, guides an associated reinforcement learning system. This can be seen as an automatic teacher that enables convergence improvement on a system which otherwise would have greater learning difficulties. Reactive systems are difficult to design and program. To overcome this problem, in [9] a methodology is proposed for behavior engineering that incorporates reinforcement and evolutionary learning. Additionally, pure reactive systems may generate inefficient trajectories since they choose the next action as a function of the current sensory readings, and the robot perception range is limited. This perception range limitation is known as the hidden state problem (e.g., [8] and [20]). To address this problem, hybrid approaches combine reaction, 1083–4419/99$10.00  1999 IEEE