Bayesian Landmark Learning for Mobile Robot Localization

  • Authors:

  • Sebastian Thrun

  • Affiliations:

  • Computer Science Department and Robotics Institute, Carnegie Mellon University, Pittsburgh, PA 15213-3891. URL: http://www.cs.cmu.edu/∼thrun

  • Venue:
  • Machine Learning
  • Year:
  • 1998

Quantified Score

Hi-index 0.00

Visualization

Abstract

To operate successfully in indoor environments, mobilerobots must be able to localize themselves. Most currentlocalization algorithms lack flexibility, autonomy, and oftenoptimality, since they rely on a human to determine what aspectsof the sensor data to use in localization (e.g., what landmarksto use). This paper describes a learning algorithm, called BaLL,that enables mobile robots to learn what features/landmarks arebest suited for localization, and also to train artificial neuralnetworks for extracting them from the sensor data. A rigorousBayesian analysis of probabilistic localization is presented,which produces a rational argument for evaluating features, forselecting them optimally, and for training the networks thatapproximate the optimal solution. In a systematic experimentalstudy, BaLL outperforms two other recent approaches to mobilerobot localization.