Ball Balancing Along a Trajectory

Learning complex behavior in real robot scenarios is a challenging problem that requires solving several tasks simultaneously. Traditional reinforcement learning techniques, however, often favor experiments with a single challenge. Although such experiments are well suited for examining the performance and robustness of learning algorithms they may omit difficulties that may occur in combined tasks. In this work, we address the problem of a combined task that requires learning both controller gains and a trajectory. Initially, we make use of a state-of-the-art policy search method to learn each behavior, respectively. Based on the results we finally learn a behavior for solving the combined challenge.