2015 Volume 27 Issue 5 Pages 701-710
Recently, robots with many degrees of freedom to achieve various tasks have been developed, and learning based control for these robots have attracted considerable attention. However, for these robots, there are two significant problems: real-time learning and a lack of generalization ability. In this paper, to solve these problems, we propose to design a body for abstracting the necessary small state-action space from a huge state-action space by utilizing properties of the real world, like dynamics, mechanical constraints, and so on. As an effective example, we consider grasping and carrying tasks and develop a redundant manipulator inspired by an octopus. We show that by designing the manipulator to utilize properties of the real world, the state-action space can be abstracted, and the real-time learning and lack of generalization ability problems can be solved.