Robotics: Science and Systems III

Spatially-Adaptive Learning Rates for Online Incremental SLAM

Edwin Olson, John Leonard, and Seth Teller

Abstract: Several recent algorithms have formulated the SLAM problem in terms of non-linear pose graph optimization. These algorithms are attractive because they offer lower computational and memory costs than the traditional Extended Kalman Filter (EKF), while simultaneously avoiding the linearization error problems that affect EKFs. In this paper, we present a new non-linear SLAM algorithm that allows incremental optimization of pose graphs, i.e., allows new poses and constraints to be added without requiring the solution to be recomputed from scratch. Our approach builds upon an existing batch algorithm that combines stochastic gradient descent and an incremental state representation. We develop an incremental algorithm by adding a spatially-adaptive learning rate, and a technique for reducing computational requirements by restricting optimization to only the most volatile portions of the graph. We demonstrate our algorithms on real datasets, and compare against other online algorithms.



    AUTHOR    = {E. Olson and J. Leonard and S. Teller},
    TITLE     = {Spatially-Adaptive Learning Rates for Online Incremental SLAM},
    BOOKTITLE = {Proceedings of Robotics: Science and Systems},
    YEAR      = {2007},
    ADDRESS   = {Atlanta, GA, USA},
    MONTH     = {June},
    DOI       = {10.15607/RSS.2007.III.010}