Accuracy of the Inverse Kinematics of a Planar Redundant Manipulator Solved by an MLP Neural Network

The inverse kinematics of manipulators can be solved by approximating the data obtained using forward kinematics through the use of a multi-layer perceptron neural network, provided the input and output data are swapped before training. However, with redundant manipulators, reaching the same point is possible with an infinite number of joint angle settings, rendering the existence of an inverse function impossible. Data for training the neural network must be prepared so that only one combination of angles is used for each point reached. Achieving this may involve supplementing the obstacle avoidance function, even though the forward kinematics of the manipulator lacks a clear, analytically accessible solution. The paper addresses how to find a solution to forward kinematics that optimally fulfills the function describing the obstacle avoidance problem, ensuring uniqueness and continuity for use in neural network training. The acquired data were used to train a neural network with one hidden layer, and the accuracy of the resulting network was verified on a simple trajectory.