A study on forward kinematics positioning modeling of a six-axis robotic arm based on neural network

As the number of axes and the precision requirements of industrial robotic arms continue to increase, traditional methods often face limitations in addressing non-geometric errors. Neural networks can handle highly nonlinear and complex systems, making them well-suited for this problem. This study proposes a custom neural network to build a forward kinematics model that inputs six joint rotation parameters and generates a 4×4 homogeneous transformation matrix, encompassing the end-effector’s position and orientation. The six joint rotation parameters include rotation and translation matrices along the XYZ axes. Using the ABB IRB1300 six-axis industrial robotic arm as a model, the custom neural network was trained on 100,000 randomly generated datasets of joint rotation angles, end-effector positions, and orientations. Experimental results on 10,000 test datasets show that the proposed method achieves an end-effector position error of 0.029 mm, with orientation errors of 0.064° in Roll, 0.057° in Pitch, and 0.093° in Yaw.