Neural hysteresis friction modeling for industrial robot dynamics identification

Industrial robot dynamics lay the foundation for high-precision and high-speed control, and accurate identification of dynamic parameters is essential for precise dynamic calculations.The choice of friction models is a critical component in the identification of industrial robot dynamics.Traditional static friction models struggle to capture the hysteresis effects caused by robot joint elasticity and clearances, leading to large torque prediction errors when the joint velocity crosses zero.Due to the presence of hysteresis effects, the joint velocity crosses zero in the forward direction, and the reverse direction will have different friction patterns.Although the hysteresis effects can be modeled as an ordinary differential equation (ODE), it is difficult to determine the ODE structure that achieves both generalization and accuracy to describe the hysteresis effects of the friction model.To address this issue, we propose the neural hysteresis friction (NHF), which uses neural ODE to model the hysteresis effects in a data-driven manner, thereby mitigating the current inadequacies in the study of dynamic friction characteristics.The experiments on a real 6-axis industrial robot demonstrate that our proposed method can accurately model the friction dynamics during directional switching and outperform other modeling methods.Velocity tracking control experiments show that NHF can effectively reduce tracking errors when the velocity crosses zero.