Learning Robot Behaviours by Extracting Fuzzy Rules from Demonstrated Actions

In this paper we describe a supervised robot learning method which enables a mobile robot to acquire the ability to follow walls and negotiate confined spaces by having these behaviours demonstrated with example actions. We achieve this by demonstrating the desired motion with a remote control while accumulating training data from the robot’s sensors and teacher’s instructions. To speed up learning and make the training data more comprehensive, additional training patterns are added to the training data by translating the demonstrated exemplars so that training data applicable to locations near the demonstrated paths are also obtained. Once sufficient training data is collected, the robot’s fuzzy rule base is generated with a fuzzy rule extraction algorithm which is tolerant to the noise and uncertainties associated with robot training data. Results of simulated and real robot experiments are provided which demonstrate the effectiveness of this approach to robot learning.