Inverse Dynamic Model using GRU Networks Learning

Abstract

Considering the designing the controller of a lower limb rehabilitation robot, an accurate inverse dynamic model is usually needed to calculate the joint torque, which is the key to trajectory execution. However, the physical modeling method will have various uncertainties, resulting in inaccurate estimating and measuring physical parameters. Therefore, this paper proposes to use a Gated Recurrent Unit (GRU) to learn the inverse dynamic model. First of all, the gait data and plantar force data of healthy people are obtained synchronously using the NOKOV 3D optical motion capture system and the Bertec 3D force measurement platform. Secondly, using MATLAB, the angle, angular velocity, angular acceleration and joint torque of the hip and knee joints at different moments can be calculated when people walk normally with the data derived from the system. Finally, the calculated joint angles, angular velocities, and angular accelerations are used as inputs to the GRU network, and joint torques are used as outputs to learn the inverse dynamics model of the exoskeleton robot. The evaluation index of the training model is the Root Mean Square Error (RMSE), which eventually settles at about 0.15. The results show that the model has a good learning ability and has made some certain contribution to the development of the exoskeleton robot control system.