April 9, 2009 - Steven M. LaValle

Title: Combinatorial Filters: Handling Sensing Uncertainty by Avoiding Big Models

Abstract:
Over the past several years, Bayesian filtering techniques have become mainstream tools in robotics research that handles uncertainty. Variations include the classical Kalman filter and recent particle filters, all of which are routinely used for robot localization, navigation, and map building. In this talk, I will introduce a new class of filters, called combinatorial filters, that are distinctive in several ways: 1) they simplify modeling burdens by avoiding probabilities, 2) they are designed for processing information from the weakest sensors possible, and 3) they avoid unnecessary state estimation. In many ways, they are the direct analog to Bayesian filters, but handle substantial amounts of uncertainty by refusing to model it. The emphasis is on detecting and maintaining tiny pieces of information that are critical to solving robotic tasks, such as navigation, map building, target tracking, and pursuit-evasion.