Arrangement Map for Task Planning and Localization for an Autonomous Robot in a Large-Scale Environment Paulo Pinheiro School of Electrical and Computer Engineering (FEEC) University of Campinas Campinas,Brazil 13083-852 Email: pinheiro@ic.unicamp.br Eleri Cardozo School of Electrical and Computer Engineering (FEEC) University of Campinas Campinas,Brazil 13083-852 Email: eleri@dca.fee.unicamp.br Jacques Wainer Institute of Computing (IC) University of Campinas Campinas,Brazil 13083-852 Email: wainer@ic.unicamp.br Eric Rohmer School of Electrical and Computer Engineering (FEEC) s University of Campinas Campinas,Brazil 13083-852 Email: rohmer@dca.fee.unicamp.br Abstract—This paper presents a planning approach for solving the global localization problem for cleaning robots. The approach is based on architectural design features of the building such as walls and doors to help the robot on finding the best route to go. Lighter POMDP plans are generated only for representative rooms of the environment, decreasing size of the set of possible states. The plans are created offline only once and used indef- initely regardless of mission and missions are combined online through vectorization. The plan only will requires as an input, the environment map and the robots capabilities regarding actions and observations. We demonstrate the single level approach and the map decomposition with experiments on both V-REP Simulator and the Pioneer 3DX robot. This approach allows the robot to perform both the localization and tasking in a large-scale environment while keeping the accuracy. I. I NTRODUCTION Localization for mobile robots is one of the most explored areas in robotics due its importance for the autonomous mobile robots. In cases where the robot’s pose may be uncertain or unknown, the autonomy is directly related to the ability of the robot to locate itself and planning for actions is one of tools can be used to improve this skill. In this paper, our concerns is to solve the localization problem using planning. If the robot’s pose is unknown, but the robot is in a partial or well known environment, a sequence of planned actions can lead the robot to maximize the useful observations and therefore to find out its pose more quickly. These plans are functions made of the probability distributions of the robot’s pose, actions and observations, and only depend on the environment and the robot’s features, thus they can be precomputed. We use a Partially Observed Markov Decision Process (POMDP) to model the plans [1]. Also we are considering that the robot can act towards its goals at the same time it is acting to locate itself. We allow for the localization plan (which was precomputed) to be mixed at execution time with the goal plan, executing both task and localization. This results in reaching its goal faster than if one would follow the standard practice of first localizing itself and then reaching its goal. Let’s suppose that the robot’s main task is move out to an specific point on the environment and the robot does not know its location. In our work, the planner defines a set of actions that is a trade-off between the localization optimal path (precomputed by the POMDP planner) and the tasking path in order to decrease the total number steps performed as shown in Fig. 1. (a) (b) (c) (d) Fig. 1: (a) the Pioneer 3DX robot used. (b) localization optimal path. (c) task optimal path. (d) simultaneously localization and tasking path shorter than the combination of (b) and (c).