Self-tuning control with neural network for robot manipulator

This paper presents an approach of the self-tuning control with neural network for robot manipulator in an object balancing task. A 3DOF robot arm (Novint Falcon 3D haptic) is used to hold a flat plate balancing a cylindrical object put on. Since a neural network algorithm were presented earlier in [1], [2] in order to learn and control the posture of the robot, we now employ the visual feedback into neural network to enable the robot arm learn to move its end effector. A webcam is used to determine position of a cylinder object rolling on a flat plate that the robot is holing. The images are processed to the object position by neural network. The output of the neural network is the height of the robot's end-effector that the robot has to lift the plate. The neural network must learn and self-calibrate by some repeating trail movements until the virtual feedback enable the robot arm adopts recalibrating parameters to stabilize the rolling task. The results of experiments show the learning procedure of the neural network is succeed in self-tuning control.