Inverse kinematics identification of a spherical robot based on BP neural networks

This paper proposed a method of neural networks to deal with the identification of the inverse kinematics of a spherical robot BHQ-1. The proposed method solves the problems of model error introduced by the generalized inverse method. It can compensate the external perturbation in the actual environment by applying an on-line learning technique, which improves the precision of the inverse kinematics model. Neural networks can approximate arbitrary order nonlinear systems and the robustness of neural networks has been proved, which shows that the deduced inverse system can be applied to actual control of spherical robot. At last, some test data has been used to validate the performance of the off-line trained model and the simulation results show that the inverse model is accurate and stable.