Inverse Jacobian control-based neural network model for controlling a 3DOF arm robot

One approach that has proven effective to control robot's movement is inverse Jacobian control, which allows controlling robot motion with high accuracy. The trajectory of the robot arm is directly related to kinematics, namely the speed and rotation of the robot's joints to reach the desired end effector position. In finding the inverse solution from the end-effector speed to the speed of all robot joints, the Jacobian matrix is a tool that represents the calculation of the trajectory. However, this method requires quite complicated mathematical calculations, especially if the Jacobian matrix obtained is singular. This study proposes the use of Multi-Layer Perceptron neural network controller to imitate inverse Jacobian control for a 3DOF arm robot, with the aim of improving the performance of the control system. The training process involves collecting position and velocity data of the robot's end-effector, as well as calculating the Jacobian matrix and its inverse. The neural network is trained to map the position and speed of the end-effector to the required control signals. Simulation results show that the neural network controller can imitate inverse Jacobian control. In addition, the performance obtained on the robotic arm with neural network control has a better mean square error (MSE) when compared to the robotic arm using state feedback control. The average increase in MSE performance is 31.24 %. This approach is expected to make a significant contribution to the development of a more flexible and efficient robot control system. However, the performance on the z axis shows less satisfactory results compared to the x and y axes since effect of gravity forces acting directly in the vertical direction