Optimization of Tendon-Driven Robot Joint Stiffness using GA-based Learning

Abstract

A tendon-driven robot mechanism allows us to control the joint stiffness independantly of the joint torques. Generally, the optimal joint stiffness for desired tasks cannot be found easily. So we adopt a Radial Basis Function Network (RBFN) to describe the trajectory of the stiffness matrix, and modify the parameters using Genetic Algorithm (GA) to find the optimal trajectories. As typical tasks that have impulsive disturbance forces, we choose ball hitting and receiving tasks. The optimal joint stiffness for the hitting task gives the ball the fastest initial speed after the hit, on the other hand, the one for the receiving gives the slowest speed. We use a 2 DOF robotic arm driven with 6 tendons, because we can adjust all of elements of the joint stiffness matrix independently of the joint torques. After modifying conventional GA to fit it for real robot experiments, we make two kinds of experiments given above and show the results.