S.H.G. ten Hagen

Towards global consistent pose estimation from images

We propose a method for making globally consistent pose estimates using vision. Lu and Milios (1997) described an approach where the links between poses were estimated using range scanner data. When using point correspondences the length of the links cannot be estimated and the approach of Lu and Milios has to be modified. First, we use the nonlinear orientation part of the pose differences to obtain a reference trajectory. Then the reference trajectory is used to scale and orientate the linear spatial part of the pose differences, such that the positions can be estimated as well. We show the results of an experiment of navigating a robot equipped with an omnidirectional camera on the corridor.We propose a method for making globally consistent pose estimates using vision. Lu and Milios (1997) described an approach where the links between poses were estimated using range scanner data. When using point correspondences the length of the links cannot be estimated and the approach of Lu and Milios has to be modified. First, we use the nonlinear orientation part of the pose differences to obtain a reference trajectory. Then the reference trajectory is used to scale and orientate the linear spatial part of the pose differences, such that the positions can be estimated as well. We show the results of an experiment of navigating a robot equipped with an omnidirectional camera on the corridor.

Discovering stages in web navigation

Users of web sites often do not know exactly what they are looking for or what the site has to offer. During navigation they use the information found so far to formulate their information needs and refine their search. In these cases users need to pass through a series of pages before they can use the information that will eventually answer their question. Recommender systems aimed at leading users to target pages directly do not provide optimal assistance to these users. In this paper we propose a method to automatically divide web navigation into a number of stages. A recommender can use these stages to recommend pages which do not only match the topic of a users search, but also the current stage of the navigation process. As these recommendations are more tailored toward the users current situation, they can provide better assistance than recommendations made by traditional recommender systems.

Good features to map

In this paper visual homing is combined with an appearance model to obtain a navigation method that does not require estimates of the pose. The robots current view is used to select a target from a collection of panoramic images that forms the map. This requires selecting features to find correspondences and in this paper some feature selection methods are evaluated to see which one is best suited for navigation.In this paper visual homing is combined with an appearance model to obtain a navigation method that does not require estimates of the pose. The robots current view is used to select a target from a collection of panoramic images that forms the map. This requires selecting features to find correspondences and in this paper some feature selection methods are evaluated to see which one is best suited for navigation.

Machine Learning and Reinforcement Learning

In this chapter we outlined a method for the development of adaptive systems using methods from Machine Learning and Reinforcement Learning and demonstrated this with a variety of adaptive systems. The main questions are the translation of the problem into a learning problem and the assessment of the feasibility of adaptive systems for the problem at hand.

Pseudo-parametric Q-learning using feedforward neural networks

In this paper we focus on Q-learning in domains with continuous state and action spaces. We discuss how Q learning relates to System Identification (SI) methods for Linear Quadratic Regulation (LQR) and show how the methods compare on linear systems. We also study the use of a feedforward network as a nonlinear function approximator for the Q function and introduce the the concept of Pseudo-Parametric Q-Learning (PPQL). In the PPQL framework the feedforward network is implemented such, that the results can be interpreted in term of LQR conditions. Experiments show that it performs well, but does not necessarily converge to a stable solution. The LQR interpretation indicates the origin of that problem.