Spatially-adaptive learning rates for online incremental SLAM

Abstract

Several recent algorithms have formulated theSLAM problem in terms of non-linear pose graph optimization.These algorithms are attractive because they offer lower computationaland memory costs than the traditional Extended KalmanFilter (EKF), while simultaneously avoiding the linearizationerror problems that affect EKFs.In this paper, we present a new non-linear SLAM algorithmthat allows incremental optimization of pose graphs, i.e., allowsnew poses and constraints to be added without requiring the solutionto be recomputed from scratch. Our approach builds uponan existing batch algorithm that combines stochastic gradientdescent and an incremental state representation. We develop anincremental algorithm by adding a spatially-adaptive learningrate, and a technique for reducing computational requirementsby restricting optimization to only the most volatile portions ofthe graph. We demonstrate our algorithms on real datasets, andcompare against other online algorithms.