DMCA
Learning to Drive Over Rough Terrain
| Citations: | 1 - 0 self |
BibTeX
@MISC{Sammut_learningto,
author = {Claude Sammut and Waleed Kadous and Raymond Sheh},
title = {Learning to Drive Over Rough Terrain},
year = {}
}
Share
 |
 |
 |
 |
||
OpenURL
Abstract
This paper addresses the problem of learning to drive a robot over rough terrain as might be found after a disaster such as an earthquake. At present, rescue robots are tele-operated, either under radio control or through a tether. Tethers limit the range of the robot and can become tangled and under radio control, the robot may go out of range and lose communication. Therefore, some degree of autonomy is very desirable. Most previous attempts at autonomous traversal of rough terrain have adopted the classical approach of path planning based on models of the vehicle and terrain. However, it is extremely difficult to construct a model of the interaction of a tracked robot with loose rubble and debris. Nevertheless, human operators are able to drive robots over this kind of terrain. Behavioural cloning is a natural technique to use on such a case. We capture traces of an operator's skill as he drives the robot over a field of rubble. These traces are used by a learning program to create a set of control rules that model the operator's skill. There are two features that distinguish this work from previous research in behavioural cloning. The first is that the state of the system at the time an operator's action is taken, is characterised by the contents of a range image from a 3D camera. Thus, the input to the learning program is extremely complex, requiring a relational representation. The second unique aspect of this work is a comparison of two distinct approaches to behavioural cloning. In one, we assume that the operator only has a limited ability articulate some aspects of his skill, so we rely heavily on induction from the training data. In the second approach, we provide the operator with tools to give strong guidance to the learning system.
Keyphrases
drive rough terrain behavioural cloning radio control rough terrain lose communication second approach learning system relational representation previous attempt range image path planning control rule limited ability second unique aspect natural technique training data rescue robot strong guidance human operator classical approach autonomous traversal distinct approach learning program previous research tracked robot
