External Force Estimation for Nonlinear Systems Using Recurrent Neural Network

This paper proposes an estimation method of external force to be used for force control that allows manipulators to flexibly interact with humans. In the proposed method, a recurrent neural network (RNN) is used to acquire nonlinear inverse dynamics of manipulator. By using the trained RNN to estimate nominal joint torque, external force (disturbance) can be calculated as a difference between the nominal joint torque and actual joint torque that is measured. By simulation based on a two-link manipulator affected by gravity, it has been confirmed that the RNN can be well-trained to acquire the inverse dynamics and to reproduce the joint torque from the joint angle, the joint angular velocity, and the joint angular acceleration. Furthermore simulation and experiment were conducted to estimate external force from the acquired inverse dynamics based on the trained RNN. The results showed that the proposed method can be feasible.