Complex robot training tasks through bootstrapping system identification

Many sensor-motor competences in mobile robotics applications exhibit complex, non-linear characteristics. Previous research has shown that polynomial NARMAX models can learn such complex tasks. However as the complexity of the task under investigation increases, representing the whole relationship in one single model using only raw sensory inputs would lead to large models. Training such models is extremely difficult, and, furthermore, obtained models often exhibit poor performances. This paper presents a bootsrapping method of generating complex robot training tasks using simple NARMAX models. We model the desired task by combining predefined low level sensor motor controllers. The viability of the proposed method is demonstrated by teaching a Scitos G5 autonomous robot to achieve complex route learning tasks in the real world robotics experiments.