Optimal selection of cane usage with humanoid robot

Humanoid robots are required to improve its walking ability (stability and efficiency) in order to adapt to various environments. Extension of the embodiment by a cane is effective as well as human like elderly. In this paper, we propose a scheme that enables the robot to perform cane-supported walking and select the optimal cane usage based on its condition so as to maximize the performance of cane support. For cane-supported walking, we propose cyclic and leg-like usages, which correspond to two-point and three-point gaits. Our strategy for selecting the optimal cane usages, which have respective utilities, utilizes falling factors in Selection Algorithm for Locomotion (Us-SAL). We express affinities between the cane usages and the falling factors as Q-values, which are learned by Q-learning, and select the optimal usage based on the Q-values. Us-SAL was validated by an experiment; the robot selected the optimal usage for its condition, and walked efficiently without falling.