Neural network based inverse kinematics for robotic manipulator

• Main Author: Zhang, Shuqi
• Other Authors: CHEAH Chien Chern
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Electrical and electronic engineering::Control and instrumentation::Robotics
• Online Access: https://hdl.handle.net/10356/140562
• Institution: Nanyang Technological University

A fundamental function of robotic manipulators is to let the end-effector precisely follow the trajectory indicated by the user in the Cartesian workspace. Thus, the solution to the inverse kinematics problem is critically important in the control of robotic manipulators. Traditional approaches for the inverse kinematics problem are usually dependent on the known geometric model of the robotic manipulators. If the geometric model is imprecise or unknown, it would be very problematic to solve the inverse kinematics problem. Because neural networks have strong approximation ability and learning ability, they can be used to approximate the inverse kinematics. In this dissertation, neural networks are applied in approximating the inverse kinematics of the UR5e manipulator. Two typical types of neural networks are used, including multilayer perception neural network and radial basis function neural network. For the multilayer perception neural network, the suitable network structure, activation functions as well as training algorithms are investigated. The best performance is achieved when there are 30 hidden neurons with sigmoid activation function and Levenberg-Marquardt training algorithm; the MAE is around 0.1mm. As for the radial basis function neural network, the selection of center vectors as well as the spread value is studied. In order to further improve the performance for certain trajectories and reduce the computation, a new two-layer radial basis function neural network is proposed, and better performance is obtained compared to the original single-layer RBFN; the MAE is reduced by more than 70%. Practical experiments were carried out on the UR5e robotic manipulator simulator using the NN solutions and the errors were calculated. The MAE of the MLPN is around 0.1-0.2mm and the MAE of the two-layer RBFN is smaller than 1mm. The computation time is smaller than 0.05s for each point (CPU: Intel i7-8550U). In conclusion, the effectiveness of using neural networks to solve the inverse kinematics problems is verified by the simulations and practical experiments.