Learning motor skills: from algorithms to robot experiments

摘要

Abstract In this thesis, we discuss approaches that allow robots to learn motor skills. Motor skills can often be represented by motor primitives, which encode elemental motions. To date, there have been a number of successful applications of learning motor primitives employing imitation learning. However, many interesting motor learning problems are high-dimensional reinforcement learning problems which are often beyond the reach of current reinforcement learning methods. This thesis contributes to the state of the art in reinforcement learning applied to robotics both in terms of novel algorithms and applications. We show how motor primitives can be employed to learn motor skills on three different levels. All proposed approaches have been extensively validated with tasks such as Ball-in-a-Cup, darts, table tennis, ball throwing, or ball bouncing both in simulation, and on real robots.