Abstract
This paper presents a learning approach to acquire holding-up manipulation by a humanoid robot. Information of force sensors attached at the surface of the arms are mapped to a lower-dimensional space and utilized to find a better configuration for the manipulation task. The proposed motion generation framework was verified by simulations with a humanoid robot. It was shown that the robot could judge failure of its manipulation using force sensor information, and succeed to modify its configuration so as to realize holding-up manipulation.
