A guided learning strategy for vision based navigation of 4-legged robots

Abstract

This paper presents a new approach to building low level navigation behaviors for 4-legged robots through vision based demonstration learning. The main novelty of the approach is that rather than observing other entities and adapting their kinematics to the robot constraints, a supervisor controls the robot to achieve the desired behavior through a proper interface. The guided actions and the relevant input parameters are related via Case Base Reasoning, so that the robot can retrieve them later to work in an unsupervised way. This intuitive acquisition of reactive behaviors allows bottom-up construction of more complex emergent behaviors and avoids low level kinematics analysis and possible associated errors. The system has been successfully tested using a Sony Aibo robot. Experiments have proven that the robot is capable of adopting a variety of reliable behaviors depending on its relative position in relation to a ball through different trainings. Also, being reactive, the system is resistant against punctual errors and occlusions.