Manipulation Learning on Humanoid Robots

Purpose of Review The ability to autonomously manipulate the physical world is the key capability needed to fulfill the potential of cognitive robots. Humanoid robots, which offer very rich sensorimotor capabilities, have made giant leaps in their manipulation capabilities in recent years. Due to their similarity to humans, the progress can be partially attributed to the learning by demonstration paradigm. Supplemented by the autonomous learning methods to refine the demonstrated manipulation actions, humanoid robots can effectively learn new manipulation skills. In this paper we present continuous effort by our research group to advance the manipulation capabilities of humanoid robots and bring them to autonomously act in an unstructured world. Recent Findings The paper details progress in the area of humanoid robot learning, ranging from trajectory imitation, motion adaptation in order to maintain feasibility and stability, and learning of dynamics to statistical generalization of actions, autonomous learning, and end-to-end vision-to-action learning that exploits deep neural networks. Summary With the focus on manipulation, the presented research provides the means to overcome the complexity behind the problem of engineering manipulation skills on robots, especially humanoid robots where programming by demonstration is most effective.