Probabilistic robot localization and situated feature focusing

摘要

Robot localization, i.e., the task of recognizing the current position of the robot from sensor inputs is an essential problem for autonomous mobile robots. We discuss the localization problem through probabilistic models, information theoretic criteria, and statistical learning. When we use some variety of sensors or high dimensional inputs like image pixels, decreasing first their dimensionality, or extract features, is necessary for making the data tractable. We show popular feature extraction methods for localization and some properties of them. After feature extraction, we can construct position estimation probabilistic models by regression. By probabilistic modeling, the information theoretic meaning of a feature extraction method becomes clearer. We introduce a mutual information-based criterion to evaluate the feature set, and compare this criterion with Kullback Leibler divergence and the average Bayesian localization error. In general, the evaluation result of the feature extraction depends strongly on the particular region of the environment. A feature performing well in a local region may not be good for the other local region. For an entire environment, an appropriate feature should be selected according to the corresponding situation. We call this idea situated feature focusing that select feature extraction modules and local regression models. This approach can be realised by Bayesian networks to estimate possibility of current situation and the mixture of experts which is the combination of various feature extractions.