An accurate model creation and path-planning of the environment has been found to be an important precondition for various robotics tasks such as search and rescue, manipulation and obstacle avoidance. Over the past few years, the topic of learning model representations has received considerable attention e.g. building city models, reconstruction of ancient art and building other environment models. Majority of the environmental modelling research relied upon single sensor like cameras or lasers, to acquire features data. In this paper we present an approach which fuses vision and range information together to learn an accurate and better environment approximation.