Model reconstruction of serial manipulators: a stepwise data-driven approach

Advancements in dynamic modeling methods of robotic manipulator are critical to the effective implementation of model-based control. Traditional approaches rely on rigorous first-principles-based dynamic modeling and precise parameter identification, while this paper explores an alternative through data-driven model reconstruction. To tackle the curse of dimensionality in the model reconstruction of a serial robotic manipulator with multi-degree-of-freedom, a relative activation indicator is proposed. Based on this indicator, the k-means clustering algorithm is utilized to classify the data under different working conditions. Subsequently, we leverage the fundamental prior knowledge to find the dynamical characteristics of each cluster and reconstruct the dynamic model in a stepwise manner using the method of sparse identification of nonlinear dynamics (SINDy). For the library generation of SINDy, the strategy of double-feature-set for serial manipulators with common joint types is proposed. Simulation results show that the stepwise model reconstruction approach not only reduces the size of the library of candidate functions but also decreases the impact of data noise on the reconstruction results. Finally, controllers based on the reconstructed models are deployed on the experimental platform and the experimental results demonstrate the improvement in trajectory tracking performance and the potential of the proposed method in engineering applications.