Whole-body Joint Friction Modeling in a Torque-controlled Humanoid Walking Robot

Abstract

Previous research on friction estimation in robotic joints has been primarily focused on the dominant contribution of the drive unit friction originating in particular from the harmonic drive gear. With torque sensors mounted on the joint output side, the frictional effects of the drive units can be compensated by the torque controller to a large extend. However, in an assembled robot the friction in mechanical parts on the output side of the torque sensor still affects the actuation performance and leads to permanent control errors. This work aims at developing a whole-body friction estimation method that relies solely on internal position measurements of the robot and can capture the friction state on the output side of the joint torque sensor. Therefore, by comparison of a model-based acceleration estimate with the measured joint acceleration, obtained by numerical differentiation of the joint positions, a joint friction estimate can be determined. A simulation model is designed to evaluate the proposed method. Simulation results confirm that reliable friction torque estimates can be obtained given a sufficiently accurate dynamic model of the robot. It is shown that real-world imperfection like sensor noise, vibrations induced by the closed-loop control and inertia modeling errors limit the performance of the friction estimation in the experiment.