Building probabilistic motion models for SLAM

This paper describes the use of two sequential machine learning techniques for building a mobile robot´s motion model, which is an essential component in SLAM algorithms. First, the recursive least squares (RLS) algorithm is used to build a static motion model. RLS can learn motion model parameters as soon as new data (the robot´s poses from odometry readings and ground-truth poses) arrive. As a result, it can detect when the model has converged and stop training. Second, the bi-loop recursive least squares (BiRLS) algorithm is applied to learn motion model parameters when the robot operates on a dynamically changing ground surface. BiRLS can keep track of the changes in time-varying parameters better than RLS without additional data and is better suited for building a dynamic motion model. Experimental results using data from a physical robot demonstrate the effectiveness of the two sequential learning algorithms in the FastSLAM 2.0 framework.