The University of Edinburgh -
Division of Informatics
Forrest Hill & 80 South Bridge

Research Paper #694

Title:A Dynamic Net for Robot Control
Authors:Hallam,B; Hallam,JC; Hayes,GM
Date:May 1994
Presented:In Published in book, "Neural Systems for Robotics", ed. Omid Omidvar and Patrick van der Smagt.
Keywords:learning, mobile robot, behavioural model
Abstract:This paper describes ongoing work to assess the appropriateness of using Halperin's Neuro-Connector model, originally devised to explain fish behaviour, to control a robot. The model is described in mathematical detail and suggestions for improving computational efficiency are made. The operation of the model and the information necessary for initialisation are also discussed. Preliminary experiments on a mobile robot platform are described and some simulation results presented. The Neuro-Connector model is computationally expensive and requires careful initialisation, so is not suitable for standard robot tasks. However, it can tolerate a designer's imprecise understanding of sensory states, refining the stimuli used to trigger each behaviour until only specific and reliable ones are considered, and can tolerate changes in the environment and in its own sensor system. It can be trained as much as programmed and can be instructed whilst operational. This is feasible because it learns in only four stimulus presentations. Therefore this model warrants further investigation with particular reference to controlling sensor-rich robots in complex and changing environments, especially where the task requires fine sensory discriminations to be made.

[Search These Pages] [DAI Home Page] [Comment]