Back
to indexSebastian Thrun and Arno Bucken
Summary
Topological maps have the advantage that they are small in space, and efficient to process, but difficult to localize a robot's position on. Topological maps consist of nodes that identify notable regions in space. Grid-based maps are quadratic in processing time and space, but are simple to develop and maintain consistency with. The author's approach combines these two models into a cohesive structure, maximizing efficiency and reducing disk space. However, sensor integration is complex, and the process to compute the probability of grid occupation is intensive. In addition, the bayesian back prop methods are slow and require training data, which I'm not willing to rely on for outdoor environments.
Methods
Sensors are interpreted over time to construct a map, and each cell is populated with the probability of occupancy, based on applying Bayes' rule. Computing the prior probabilities in this step requires summation over previous estimates, see formula in the paper. A trained backpropigation network is applied to compute the probabilities. Then the topological maps are constructed on top of the grid maps, by identifying the critical points (points on the voronoi diagram that minimize clearance locally), and drawing critical lines between them. The data structure is condensed into the important information, and paths are planned through the course.
Keywords
topological maps, grid based maps, neural nets, back prop
Rating
6
Bibtex Entry
@inproceedings{ thrun96integrating,
author = "Sebastian Thrun and Arno Bucken",
title = "Integrating Grid-Based and Topological Maps for Mobile Robot Navigation",
booktitle = "{AAAI}/{IAAI}, Vol. 2",
pages = "944-950",
year = "1996",
url = "citeseer.nj.nec.com/thrun96integrating.html"
}