A Hybrid Active Global Localisation Algorithm for Mobile Robots Andrea Gasparri, Stefano Panzieri, Federica Pascucci, and Giovanni Ulivi Abstract— Localisation is one of the most important tasks to be accomplished in order to realize the complete autonomy of a mobile robot. In this paper, a new strategy for global localisation is proposed. Applying this method a robot is able to safely initialise its position or relocalise itself in case of recovery of pose tracking failure. The algorithm presented adopts an hybrid approach. First a particle filter is used to generate hypotheses on the possible pose supposing that no movements are allowed to avoid collisions. Thereafter safe trajectories are planned and executed to reduce the remaining ambiguities while the hypotheses are monitored and validated by a set of parallel Extended Kalman Filters. The novelty of this approach stands on the ability to generate the pose hypotheses without any feature- based knowledge. As a consequence, a landmark-based description of the environment is no longer required for the algorithm execution. I. I NTRODUCTION When mobile robots operate in real world environment they require reliable localisation systems to accomplish their missions. Therefore a localisation module is always included in mobile robot control architecture. Due to the difficulty in obtaining a reliable pose estima- tion, localisation has been a high active field of research in the last two decades. Indeed, mobile robots operate in environments that have not specifically engineered for them and localisation algorithms have to be able both to cope with the uncertainty of robot+environment system and to integrate data from different kinds of sensors. There are two basic instances of the localisation problem: when starting a task determine the robot pose in absence of an initial estimate (global localisation) and during the motion, maintain a precise estimate of the pose by keeping track of the robot movements (pose tracking). Most of the algorithms presented in literature are focused on the pose tracking, whereas less attention is dedicated to global localisation. Several research groups address localisa- tion supposing that a rough knowledge of the initial pose is supplied to the robot[1], [2]. A widespread approach to face such instance is based on stochastic estimation theory: the pose estimation problem is translated in terms of probability density estimation. In this framework the pose tracking is solved using a uni-modal probability density, estimated by a common tool, the Extended Kalman Filter (EKF) [3]. The probabilistic approach can be adopted also to deal with global localisation, however in this case uni-modal probability density can not be used as multiple hypotheses All the authors are with the Dip. di Informatica e Automazione, Uni- versit` a “Roma Tre”, Via della Vasca Navale, 79, 00146 Roma, Italy (www.dia.uniroma3.it) needs to be handled [4]. As no knowledge about the initial configuration is provided, a general approach to discover the current pose consists in comparing the information extracted from sensor readings with an a priori map in order to form hypotheses. When absolute sensors (i.e., GPS, active beacons) are not available, due to the perceptual aliasing, several candidates are retrieved. In this context multi- modal probability densities have to be adopted and EKF have to be replaced with more sophisticated filters, such as grid-based or particle filters (Markov localisation [5], [6]). Several solutions based on these classes of filters have been proposed. In [7] mobile localisation technique is pre- sented which uses multiple Gaussian hypothesis to represent the probability distribution of the robots location in the en- vironment. Moreover, in [8] an active global localisation for mobile robot using multiple hypotheses tracking is proposed. Markov localisation and EKF are combined in [9]. In all these works the key idea is to maintain multiple hypotheses on the robot pose during the navigation, without any interest on the trajectory executed by the robot. This paper addresses global localisation problem when artificial beacons or GPS are not available. The proposed algorithm takes advantage of the multi-modal approach to generate the set of the most likely hypotheses when the robot is still, while during the robot motion the hypotheses are propagated and eliminated by a set of parallel Extended Kalman Filter [10], [11]. In addition, a strategy for planning safe trajectory is also implemented in order to minimise the risk of collision during the navigation. The paper is organised as follows. Section II presents the theoretical background, section III describes the case study investigated, section IV details the proposed algorithm, and, finally, some simulation results and conclusion are reported in section V and VI respectively. II. PROBABILISTIC FRAMEWORK The localisation problem, i.e., the problem of estimation the robot’s pose given noisy measurements, can be described as a stochastic estimation problem. In such framework localisation can be formulated in terms of estimating the probability density over the state space of the robot poses. In literature such probability density is called Belief and is defined as Bel(x k )= p(x k | U k ,Z k ),x ∈ Ξ. (1) i.e., the probability to have the robot at location x k at time k, given all the history of the proprioceptive (U k ) and