On-line learning of robot inverse kinematic transformations

Because of the learning and adaptive features, the computational, normally feedforward (static) neural networks have been used in robotics, particularly to obtain solutions to inverse kinematics problems. The procedure, in general, employs two modes of operation. The first mode is to train the network off-line, while the second mode achieves the tracking to a desired position within the task-space based on the trained data. However, the objective of this paper is to propose an online learning and adaptive scheme using a dynamic neural network. It is demonstrated in this paper that the proposed scheme, taking the desired Cartesian coordinates as the inputs, determines the robot joint angles and makes the robot reach the desired position, thereby achieving learning and performing actions together. The computer simulations are presented by approximating the human leg as a two-linked robot to demonstrate the effectiveness of the proposed scheme.