On-line learning of robot arm impedance using neural networks

Impedance control is an effective control method for a manipulator that is in contact with its environment. Nevertheless, the characteristics of force and motion control are determined by impedance parameters of the end-effector of the manipulator, which must be designed according to the given task. This report presents a method that uses neural networks to regulate impedance parameters of the manipulator's end-effector while identifying environmental characteristics through on-line learning. Four kinds of neural networks are used: three for the position, velocity and force control of the end-effector, and one for the identification of environments. First, the neural networks for the position and velocity control are trained during free movements. Then, the neural networks for the force control and identification of environments are trained during contact movements. Computer simulations show that the method can regulate stiffness, viscosity and inertia parameters of the end-effector and identify unknown properties of the environments through on-line learning.