Pose recovery for a mobile manipulator using a particle filter (Invited Paper) Andrea Gasparri, Stefano Panzieri, Federica Pascucci, and Giovanni Ulivi Dip. di Informatica e Automazione Universit` a degli Studi Roma Tre Via della Vasca Navale 79, 00146, Roma, Italy Email: {gasparri,panzieri,pascucci,ulivi}@dia.uniroma3.it Abstract— A multirobot system composed by a five DoF manip- ulator (SCORTEC-ER 1) and a mobile base (ATRV-Jr by iRobot) is considered. The system is equipped with a uniform software architecture able to control the motion of both components. Some exteroceptive sensors like sonars and laser rangefinders are available on the mobile platform, some others are mounted on the end-effector of the manipulator: a colour camera, able to locate objects to operate with; an ultrasound range finder, near the camera, that will be used to get distance measures. In this paper we consider the problem of relocalising the mobile manipulator in a known environment supposing that no base movement are allowed to avoid hazard operations. As only the arm can be safely panned to explore the surroundings, the proposed procedure is based exclusively on the ultrasound rangefinder. Using such measures, and making the hypothesis that the arm configuration in known, a Monte Carlo filter is designed to estimate the base pose. Such multiple hypotheses filter is appropriate in this kind of problems where, in the best case, it is possible to retrieve a finite number of possible poses compatible with available measures. I. I NTRODUCTION Sensor fusion encompasses several approaches to the com- bination of data obtained from several sensors and, sometimes, further relevant knowledge provided by other sources. Sensor fusion is aimed to the improvement of the overall accuracy and, more in general, to draw more specific conclusions than those obtained by single sensors. This idea is hardly original as in natural world all the animals use all their sense together to improve survival opportunities. In literature, sensor fusion is brought to its maximum extent only in connection with battlefield intelligence and target acquisition, where it is necessary to identify friends and foes and to assess the threat level [1], [2]. Nonmilitary applications of sensor fusion range from medicine to environmental mon- itoring and include advanced robotics, in particular when the environment is not (fully) structured. In this case, the success of the operations requires the use of several sensors in proper positions to get a good knowledge of the nearby environment, of the robot relative position and of possible targets. Frequent applications in mobile robotics include map build- ing [3], robot localization with respect to an a priori known map, detection of particular sites during the motion (doors, power supply points, object to collect or avoid). In the last years, techniques to perform simultaneously map building and localisation (SLAM) have been developed [4], [5]. This paper addresses the global localisation problem, where the initial pose has to be determined from scratch [6]. Most of the approaches proposed in literature translates the global localisation problem into an estimation problem for systems with uncertainty. In this context probability theory seems to provide an efficient mathematical framework.The key idea is to recursively maintain a probability distribution (Belief) over all poses (state space points) in the environment. Several prob- abilistic global techniques have been proposed, introducing different kind of discretisation of the state space. Grid based methods [7] discretise the state space into equally spaces grid of points. Such samples are used as a basis to approximate the belief by a piece-wise constant function. The advantage of these approaches come from the capability to represent more complex distribution, e.g., multi-modal distribution, however they suffer from great computational and memory load [8]. A more promising approach is based on sequential Monte Carlo integration methods [9], and several research groups apply such methods to solve localisation [3]. A set of samples (particles) is used to represent the probability density function encoding the robot knowledge of the pose [10]. This reduce the high computational overhead coming with the use of grid based discretisation of the state space. For the samples set to be a good approximation of the belief, it is important that the particles gives support where the probability density function differs from zero. Here a pose recovery algorithm based on a particle filter is proposed. The algorithm is designed for SCAT (SCORTEC- ER 1 ATRV-Jr), a multirobot platform integrated by the research group of Roma Tre, during the first year of PICTURE program, as explained in sec. II. The novelty of the approach consists in the use of a evolutional algorithm to select and propagate ( i.e., draw) the particle. To understand the proper- ties of the proposed approach, Markov localisation is reviewed in sec. III and then the pose recovery algorithm is presented in sec. IV. Preliminary experiments show application of the method to the case of still vehicle and moving arm (sec. IV- E).