Sensorless force estimation for industrial robots using disturbance observer and neural learning of friction approximation

Contact force estimation enables robots to physically interact with unknown environments and to work with human operators in a shared workspace. Most heavy-duty industrial robots without built-in force/torque sensors rely on the inverse dynamics for the sensorless force estimation. However, this scheme suffers from the serious model uncertainty induced by the nonnegligible noise in the estimation process. This paper proposes a sensorless scheme to estimate the unknown contact force induced by the physical interaction with robots. The model-based identification scheme is initially used to obtain dynamic parameters. Then, neural learning of friction approximation is designed to enhance estimation performance for robotic systems subject with the model uncertainty. The external force exerted on the robot is estimated by a disturbance observer which models the external disturbance. A momentum observer is modified to develop a disturbance Kalman filter-based approach for estimating the contact force. The neural network-based model uncertainty and measurement noise level are analysed to guarantee the robustness of the Kalman filter-based force observer. The proposed scheme is verified by the measurement data from a heavy-duty industrial robot with 6 degrees of freedom (KUKA AUGLIS six). The experimental results are used to demonstrate the estimation performance of the proposed approach by the comparison with the existing schemes.