AUTONOMOUS ROBOT LEARNING MODEL BASED ON VISUAL INTERPRETATION OF SPATIAL STRUCTURES

Abstract

The main concept of the presented research is an autonomous robot learning model for which a novel ARTgrid neural network architecture for the classification of spatial structures is used.The motivation scenario includes incremental unsupervised learning which is mainly based on discrete spatial structure changes recognized by the robot vision system.The learning policy problem is presented as a classification problem for which the adaptive resonance theory (ART) concept is implemented.The methodology and architecture of the autonomous robot learning model with preliminary results are presented.A computer simulation was performed with four input sets containing 22, 45, 73, and 111 random spatial structures.The ARTgrid shows a fairly high (>85%) match score when applied with already learned patterns after the first learning cycle, and a score of >95% after the second cycle.Regarding the category proliferation, the results are compared with a more predictive modified cluster centre seeking algorithm.