Neural Network Learning Of Robot Dynamic Uncertainties And Observer-Based External Disturbance Estimation For Impedance Control

Estimation of dynamic uncertainties is a critical and fundamental problem when designing a control system for a robot. During robot-environment interaction, in addition to the internal dynamic model uncertainties, the external environment-exerted force will also enter the dynamics. For robot impedance control, an exact dynamic model of the robot is needed but usually not available. It has been shown that integrating an impedance controller with a disturbance observer can achieve accurate impedance control. However, it works only for robots in free motion but not robot-environment interaction. Although a disturbance observer is able to accurately estimate the dynamic uncertainties, the estimation is lumped uncertainties that contain all uncertainty sources including both the internal and the external disturbances. Without separating these two parts, the method of combining an impedance controller and an observer will result in the human-applied force being canceled instead of interacting with the robot. To solve this problem in this paper, we propose a framework for learning the internal disturbances and separating the external disturbances by integrating three entities: an impedance controller, a neural network (NN) model, and a disturbance observer. In the framework, the impedance controller provides compliant robot behavior, while the observer captures the lumped uncertainties, and the NN learns to separate the external disturbances. Simulation results of an application scenario with an obstructive virtual fixture demonstrate the effectiveness of the proposed framework.