An improved FastSLAM framework using soft computing

Abstract

FastSLAM is a framework for simultaneous localization and mapping (SLAM) using a Rao-Blackwellized particle filter. However, FastSLAM degenerates over time. This degeneracy is due to the fact that a particle set estimating the pose of the robot loses its diversity. One of the main reasons for losing particle diversity in FastSLAM is sample impoverishment. In this case, most of the particle weights are insignificant. Another problem of FastSLAM relates to the design of an extended Kalman filter (EKF) for the landmark position's estimation. The performance of the EKF and the quality of the estimation depend heavily on correct a priori knowledge of the process and measurement noise covariance matrices (Q_t and R_t ), which are, in most applications, unknown. Incorrect a priori knowledge of Q_t and R_t may seriously degrade the performance of the Kalman filter. This paper presents a modified FastSLAM framework by soft computing. In our proposed method, an adaptive neuro-fuzzy extended Kalman filter is used for landmark feature estimation. The free parameters of the adaptive neuro-fuzzy inference system (ANFIS) are trained using the steepest gradient descent (SD) to minimize the differences of the actual value of the covariance of the residual from its theoretical value as much possible. A novel multiswarm particle filter is then presented to overcome the impoverishment of FastSLAM. The multiswarm particle filter moves samples toward the region of the state space in which the likelihood is significant, without allowing them to go far from the region of significant proposal distribution. The simulation results show the effectiveness of the proposed algorithm.