Enhanced online parameter estimation of unknown objects via sparse identification of unmodeled dynamics

Accurate parameter estimation of unknown objects is crucial for the precise and safe manipulation of robotic systems in applications such as object positioning, assembly, and collaborative manipulation tasks involving humans or multiple robot agents. However, measurement data obtained from sensors often contain uncertainties, making accurate parameter estimation challenging. In this paper, a systematic methodology for unmodeled dynamics identification that represents uncertainties in measured sensor data is proposed for accurate online parameter estimation of task objects. The sparse identification of nonlinear dynamics (SINDy) technique, a recent machine learning approach, is employed to identify unmodeled dynamics. First, in the learning process, an unmodeled dynamic equation can be obtained by establishing residual data, which are obtained by subtracting the dynamics of prior known objects from measured sensor data and by designing proper candidates that successfully capture uncertain behavior. Second, in the online parameter estimation process for an unknown object, estimation results that are not contaminated by uncertainties can be obtained by incorporating the identified unmodeled dynamic equation into the nominal object equation. To verify the robustness and estimation accuracy of the proposed methodology, experiments were conducted using various objects. The experiments demonstrate that the proposed method improves the estimation accuracy by reducing errors by 15.71 %, on average, compared to the estimation accuracy of conventional methods that only consider nominal object dynamics.