Learning energy efficient walking based on ballistics

This paper presents a method to realize energy efficient walking of a biped robot with a layered controller. The lower layer controller is a state machine which consists of four states: (1) constant torque is applied to hip and knee joints of the swing leg. (2) no torque is applied so that the swing leg can move in a ballistic manner. (3) a PD controller is used so that the desired posture can be realized at the heel contact, which enables a biped robot to walk stably. (4) as the support leg, hip and knee joints are servo-controlled to go back and the torque to support upper leg is applied. With this lower layer controller, parameters that enable robot to walk as energy efficiently as human walking can be explored by the upper layer controller without caring to avoid falling down.