CS&E Seminar: Giorgio Grisetti - 10 June 2010 - 12.00 @ Aula Magna

Speaker: Giorgio Grisetti, Autonomous Intelligent Systems Lab, Albert Ludwigs Universitaet Freiburg (Germany)

Title: Graph-Based SLAM using Stochastic Gradient Descent and Hierarchical Least-Squares

Date: 10 June 2010, 12:00 noon

Location: Aula Magna at DIS

Abstract. Constructing a map while moving through an unknown environment is an essential skill for a mobile robot that has to accomplish complex tasks over prolonged periods of time. Typical applications include search and rescue, domestic assistance, autonomous vehicles, automated delivery, and many more. This problem is known in the literature as simultaneous localization and mapping (SLAM). Due to the development of efficient optimization methods, graphbased SLAM approaches gained an increasing popularity in recent years, and are now considered by many to be the best methods available. In this talk, we will introduce graph-based SLAM and we will present two state-of-the-art approaches of graph optimization for SLAM. The first method is based on stochastic gradient descent and exploits the structure of the problem through an appropriate parametrization. It exhibits an excellent robustness to poor initial guesses and it is able to quickly find approximate solutions. The second approach is a hierarchical extension to well known least-squares optimization techniques. It allows the on-line estimation of the optimal solution for the sparsified problem. This solution is employed to construct an improved initial guess for the original problem, thus speeding up its convergence.

Giorgio Grisetti is post-doc at the autonomous intelligent systems lab of the Albert Ludwigs University in Freiburg since 2006. He received is master degree in 2001 and his PhD degree in 2006 in computer engineering from "La Sapienza" university of Rome. He is interested in developing complex autonomous systems with mobile robots. He is specially active in the areas of mobile robot navigation, environment modeling, and state estimation. He is author of over 40 publications on international peer-reviewed conferences and journals and of open source software packages addressing different aspects of the SLAM problem (www.openslam.org).


Web page of the CS&E Seminars series: http://www.dis.uniroma1.it/~seminf/