Research On Multi-Robot Cooperative Localization Based On Relative Observation In Unknown Environment

Robotics has already been widely developed and used in many fields. Urged by investigation and application, research of multi-robot system has received more and more attention, and it is gradually becoming an active field full of challenge. In some tasks, such as military operation, aviation, services and RoboCup, cooperation of multi-robot is extremely important. In order to cooperate, robots must know their location. So it is an important and key problem to improve the capabilities and accuracy of cooperative localization.This work focuses on the basic and important problem in multiple robots research: cooperative localization. It addresses the problem of cooperative localization of a mobile robot team based on the relative observations and motions in unknown environments. A few issues are investigated, including the complete model structure of cooperative localization, selection of measurements, optimal motion strategies, and application and improvement of Particle Filter.First, the complete model of cooperative localization of multi-robot is discussed. Based on the model, the Extended Kalman Filter is used to fuse the data of motion and the relative measurements to localize every member of the mobile robot team. The filter equations are deduced and the characteristics of the observations are analyzed thoroughly. Contraposing the unconverged cases of localization in some conditions based on only relative bearings, we mend the filter process so as to improve the practicability and reliability. Simulations and real experiments on the Nubot platform have been done to prove the validity of the method.Second, we consider the problem of the increasing computational burden as the size of the team is becoming large. We discuss how to select the better measurements among all of those ones obtained by the group. We present a method to select those measurements which yield the most information gain in estimating robots location. The seleted measurements are used to update the whole group pose estimation and the covariance matrix. It ensures the necessary localization accuracy and meantime reduces the computational burden, so as to improve the reliability and real-time of localization. We compare the localization accuracy and the computation time by using different number of measurements. Simulation results show that the proposed method can effectively improve the efficiency in dealing with multi-robot localization, especially when the group is large. The distributed filter computation is discussed. A few communicating filters cooperate to process the localization computation by exchange of information. This decentralized structure is helpful to reduce the complexity of computation and increase the flexibility of the system.Then, we also discuss the problem of optimal trajectories of multiple robots. The localization covariance is taken as the cost function. Error distribution of localization based on relative distance in different situations is analyzed thoroughly. Under some constraints we study the optimal motion strategies of the multiple robots so that the localization uncertainty can be minimized. Simulation experiments prove that the optimal trajectories can improve the accuracy of cooperative localization than the general motion strategies.The cooperative localization of heterogeneous robots without their initial positions is investigated. The implementations of Particle Filter (PF) and combination of EKF and PF in cooperative localization are studied. We present a method to combine EKF with PF in dealing with multi-robot absolute localization. It makes use of the efficiency and real-time of EKF, and the robust and adaptability of PF to improve the localization performance. Simulations and real data experiment acquired with Nubot platform have been done to verify the method. We propose an improved Unscented Kalman Filter (IUKF). Its performance is more robust than UKF in dealing with noisy measurements. We propose the use of the IUKF for proposal generation to combine new observation with prior probability distribution. Based on it, the improved Unscented Particle Filter (IUPF) is presented. Through simulation and real data experiment, The IUPF is proved to have better performance than UPF. The parallel structure of Gaussian Particle Filter is presented to process the computation in order to get real time implementation. This parallel structure is verified by a group of five computers system.Finally, we summarize the general work of this thesis and give a short outlook on possible future research.