Variational Inference Method-Based Dynamic Identification of Industrial Robots Considering Joint Flexibility Without Additional Load or Position Measurement

Abstract

Abstract. Accurate dynamic modeling is crucial for achieving high-performance control of industrial robots with joint flexibility. Typical identification methods for joint stiffness require laser trackers or additional joint angle encoders mounted on the robot's motor side, leading to high identification costs, and poor generalization performance. This paper proposes a novel identification method for dynamic parameters, joint stiffness, and damping parameters based on variational inference (VI) and weighted least squares (WLS), employing VI method for joint stiffness estimation in nonlinear state-space models of flexible joint robots and relying on iteratively WLS for identifying dynamic parameters. The proposed method enables the identification of both the robot's dynamic parameters and joint stiffness using a base force/torque sensor and standard motor-side variables, without the need for additional position measurement sensors (e.g., dual encoders) or added loads. Furthermore, excitation trajectories are carefully designed to balance the precision of both dynamic parameters and joint stiffness identification in all workspace, thereby improving generalization performance. Finally, several simulations and experiments are conducted on different robots to validate the effectiveness of the proposed algorithms.