Traversability Classification using Unsupervised Online Visual Learning for Outdoor Robot Navigation

1
Traversability Classification using Unsupervised
On-line Visual Learning for Outdoor Robot
Navigation

• Dongshin Kim
• Jie Sun
• Sang Min Oh
• James M. Rehg
• Aaron F. Bobick
• Robotics_at_GT, Georgia Institute of Technology

2
Traversability as an Affordance

• The traversability of a terrain region is not a
  simple function of its height above the ground
• Traversability is an affordance determined
  jointly by the robot and its environment.
• Solution Learn a classifier for traversability
  based on the robots driving experience

3
Autonomous On-Line Traversability Learning

• Our approach Automatic data collection,
  labeling, and online classification
• Unknown outdoor environment
• Using visual feature as the key expression of the
  terrain
• Egocentric local map

4
Autonomous On-Line Traversability Learning

• Collecting positive examples

• Online clustering and
• Collecting more positive
• Examples

5
Autonomous On-Line Traversability Learning

• Collecting positive examples

• Online clustering and
• Collecting more positive
• Examples

• Robot hits the tree
• Online clustering and
• Collecting positive and
• Negative data

6
Robot Overview

• our robot (sponsored by DARPA LAGR project)
• Training signal
• GPS, Bumper, and Slip sensor
• Learn models of the camera imagery

7
Experimental Result
Start
Goal
8
Experimental Result
Goal
Start
9
Video
10
Contributions

• A novel on-line algorithm for learning the
  traversability of unknown outdoor terrain
• A method to automatically acquire labeled
  training examples without human supervision
• Experimental demonstration of successful
  learning of complex terrain classes and the
  resulting improved ability to finish navigation
  tasks in challenging outdoor environments

11
Related Work

• C. Wellington and A. Stentz. Online adaptive
  rough-terrain navigation in vegetation In IEEE
  Intl. Conf. on Robotics and Automation (ICRA),
  2004.
• Boris Sofman, Ellie Lin, J. Andrew Bagnell,
  Nicolas Vandapel, and Anthony Stentz Improving
  Robot Navigation Through Self-Supervised Online
  Learning Proceedings of Robotics Science and
  Systems, August, 2006.

12
Questions