Dynamic Model Learning for Robotic Manipulators using BiLSTM Networks

Abstract

To obtain accurate dynamic models of manipulators, a model learning method based on bidirectional long short-term memory (BiLSTM) networks is proposed in this paper. The future states of manipulators are important for learning the inverse dynamic model of manipulators, which was ignored in past research. In this paper, the BiLSTM network was chosen for learning the inverse dynamic model of manipulators. The future desired position and velocity of manipulators are as the input for backward propagation. The actual position and velocity of manipulators are as the input for forwarding propagation. Furthermore, to improve the accuracy of the model, each joint of the manipulator is trained separately. Finally, comparison experiments are conducted using back-propagation (BP), long short-term memory (LSTM), gated recurrent unit (GRU), and the proposed method on the UR5 manipulator. The experimental results show that the proposed method can achieve higher accuracy and faster convergence of the inverse dynamic model. It indicates that the BiLSTM network are more conducive to learn the inverse dynamic model of manipulators.