Generalization of neural network for manipulator inverse dynamics model learning

Wenhui Huang, Yunhan Lin, Chen Jie, Liu Mingxin, Huasong Min

Abstract: The inverse dynamics model of manipulators learned from recurrent neural networks demonstrates higher precision than those obtained through analytical modeling methods. Variations in end-effector loads and previously unseen trajectory points can lead to inaccurate torque estimations in dynamic models of manipulators. This paper integrates innovative feature expansion, feature enhancement, and regularization into an end-to-end inverse dynamics model learning framework. The proposed model employs a bidirectional long short-term memory (BiLSTM) network, augmented by a spatial attention mechanism with Convolutional Neural Networks (CNN) and a Max-Pooling method, which enhances the extraction of latent spatial features, and a multi-scale parallel temporal attention mechanism, which captures the dynamic changes of objects in the temporal dimension. A novel motion residual vector is designed to expand features, and a motion residual module is proposed to assist the network in perceiving changes in end-effector loads. To prevent overfitting, novel spatial attention standard deviation regularization are implemented. Experimental results across different trajectories and end-effector loads validate the generalization capability of the proposed method. The proposed method is compared with five methods, experimental results across different trajectories and end-effector loads validate the generalization capability of the proposed method. It surpasses state-of-the-art methods, achieving the highest overall accuracy. In cross-validation experiments, the validation loss remains stable as the training loss decreases, demonstrating the proposed approach’s strong generalization performance in dynamics model learning.