Topological Mapping with Multiple Visual Manifolds

Abstract

Abstract-We address the problem of building topological maps in visual space for robot navigation. The nodes of our topological maps consist of clusters along manifolds, and we propose an unsupervised learning algorithm that automatically constructs these manifolds -the user need only specify the desired number of clusters and the minimum number of images per cluster. This spectral clustering like framework allows each cluster to optimize a separate set of clustering parameters, and we demonstrate empirically that this flexibility can significantly improve clustering results. We further propose a framework for servoing the robot in our manifold space, which would allow the robot to navigate from any point on one manifold (topological node) to any specified point on a second manifold. Finally, we present experimental results on indoor and outdoor image sequences demonstrating the efficacy of the proposed algorithm.

Learning slip behavior using automatic mechanical supervision

“…This representation, based on both color and texture, has been shown to achieve satisfactory results for classifications of natural terrains [1] 1) Dimensionality reduction: As the proposed representation is very high dimensional, we use a nonlinear dimensionality reduction technique, Isomap in particular [23], to automatically select a smaller number of suitable dimensions to represent the data. Nonlinear dimensionality reduction techniques have been successfully applied to find appropriate patterns in unsupervised fashion for visual or robot sensor data [9], [10]. With a dimensionality reduction tool at hand [23] we are able to apply the proposed in Section II method to more complex visual representations, as the texton-based one [24].…”

Appearance-based mapping using minimalistic sensor models

“…In Grudic and Mulligan (2005), a spectral-clustering-like algorithm is proposed which clusters the images to appropriately describe the topology of the map. In practice, our vision system could be replaced by any other kind of sensor modality.…”

Discovering natural kinds of robot sensory experiences in unstructured environments

“…A semi-supervised approach to discovering clusters in vision data is introduced in Grudic and Mulligan (2005). By allowing each cluster to self-optimize its parameters, they are able to discover clusters that more accurately correspond to the predefined ones, as well as detect outlying points that do not belong to any cluster.…”

“…Such "robot-centric" maps (Grudic and Mulligan, 2005) require that the robot accurately recognize when it is in certain types of space. By combining our classification with odometric or ground truth data, rough topological maps can be derived.…”