Learning Motion Skills from Expert Demonstrations and Own Experience using Gaussian Process Regression

While today's industrial robots achieve high performance in a wide range of tasks, the programming and adaptation of such robots to new tasks is still time-consuming and far from convenient. This shortcoming is a major obstacle for the deployment of robots in small to medium-size companies and also for service applications. Robot learning by imitation has been identified as an important paradigm to tackle this practicability issue. However, such approaches are only applicable in controlled static scenarios. To perform less constrained tasks, the robot needs the ability to generalize its skills and to adapt to changing conditions. It is thus a natural idea to formulate the problem as learning from experience and to incorporate demonstrations by an expert in the learning process. We present a new framework for learning of motion skills that seamlessly integrates demonstration learning from an expert teacher and further optimization of the demonstrated skill by own experience. Our method employs Gaussian Process regression to generalize a measure of performance to new skills that are close to already learned skills in the state-action space. We evaluate our approach for the task of grasping an object.