Curiosity-driven learning of traversability affordance on a mobile robot

Ugur E., Dogar M. R., Cakmak M., ŞAHİN E.

6th IEEE International Conference on Development Learning, London, Canada, 11 - 13 July 2007, pp.128-133 identifier identifier

  • Publication Type: Conference Paper / Full Text
  • City: London
  • Country: Canada
  • Page Numbers: pp.128-133
  • Middle East Technical University Affiliated: Yes


The concept of affordances, as proposed by J.J. Gibson, refers to the relationship between the organism and its environment and has become popular in autonomous robot control. The learning of affordances in autonomous robots, however, typically requires a large set of training data obtained from the interactions of the robot with its environment. Therefore, the learning process is not only time-consuming, and costly but is also risky since some of the interactions may inflict damage on the robot. In this paper, we study the learning of traversability affordance on a mobile robot and investigate how the number of interactions required can be minimized with minimial degradation on the learning process. Specifically, we propose a two step learning process which consists of bootstrapping and curiosity-based learning phases. In the bootstrapping phase, a small set of initial interaction data are used to find the relevant perceptual features for the affordance, and a Support Vector Machine (SVM) classifier is trained. In the curiosity-driven learning phase, a curiosity band around the decision hyperplane of the SVM is used to decide whether a given interaction opportunity is worth exploring or not. Specifically, if the output of the SVM for a given percept lies within curiosity band, indicating that the classifier is not so certain about the hypothesized effect of the interaction, the robot goes ahead with the interaction, and skips if not. Our studies within a physics-based robot simulator show that the robot can achieve better learning with the proposed curiosity-driven learning method for a fixed number of interactions. The results also show that, for optimum performance, there exists a minimum number of initial interactions to be used for bootstrapping. Finally, the trained classifier with the proposed learning method was also successfully tested on the real robot.