A two-stage Bayesian framework for rapid dynamics identification in industrial robots

Accurate dynamic models that update in real-time are vital for industrial robots to adapt flexibly and robustly to evolving environments and tasks, enabling precise model-based control, motion planning and disturbance estimation. However, the identification of a dynamic robot model involves multiple challenges. First, existing friction models cannot accurately characterize torque variance during prolonged robot operations. Second, traditional offline identification methods such as the least squares (LS) method, suffer from over-fitting when the inertia parameters are unknown a priori. Furthermore, existing online identification approaches, such as sliding LS, fail to meet the accuracy and real-time requirements of industrial robots. To address these challenges, this article proposes a two-stage Bayesian framework for rapid offline identification and online updating of industrial robot dynamics. In the offline stage, we developed a recursive sparse Bayesian learning (RSBL) method to select the dominant parameters and discover a thermal-related nonlinear dynamic friction model from data, which was then used to find a sparser and simpler inertia model. The obtained friction and inertia models were used as prior knowledge for online updating. The RSBL method significantly reduces the computational complexity and runtime, while it also accelerates online model updating. Experimental results for a 6-DOF medium-load robot validate dual offline improvements in prediction accuracy and model sparsity, along with higher online prediction accuracy and shorter computational time.