Learning the Inverse Model of the Dynamics of a Robot Leg by Auto-imitation

Abstract Walking, running and hopping are based on self-stabilizing oscillatory activity. In contrast, aiming movements serve to direct a limb to a desired location and demand a quite different manner of control which also includes learning the physical parameters of that limb. The paper is concerned with the question how reaching a goal can be integrated into locomotion. A prerequisite of piecing together both types of control is the acquisition of a model of the limb’s inverse dynamics. To test whether auto-imitation, a biologically inspired learning algorithm, can solve this problem, we build a motor driven device with a two-segmented arm. A preliminary study revealed that at least the forearm - with the upper arm fixed - can be made controllable by this method we called ”autoimitatively adaptable inverse control”.

• Type: Conference
• Year: 2007
• Where: AMS
• Authors: Karl-Theodor Kalveram, André Seyfarth
• Keywords: AMS 2007, Communications, Limb’s Inverse Dynamics, Motor Driven Device, Self-stabilizing Oscillatory Activity