Learning of Object Manipulation with Stick/Slip Mode Switching

This paper presents a learning approach to acquisition of manipulation control for carrying an object accompanied with stick/slip mode switchings. The proposed learning architecture has hierarchical structure. The upper layer determines a global trajectory and a sequence of mode switchings by reinforcement learning while giving the lower layer commands about small motions keeping (sliding or sticking) modes. The lower layer realizes a desired small regional motion using the estimated boundary information about mode switches by Support vector machine (SVM). Mode transition data are collected off-line for SVM learning. The proposed learning framework is applied to an object moving task with two-DOF manipulator, where the desired configuration of the object is specified so that stick/slip mode switchings are inevitable to accomplish the given task. Simulation showed that the proposed learning architecture realized an object control with stick/slip mode switching.