In this work, we consider the problem of nonlinear system identification using data to learn multiple and often coupled parameters that allow a simulator to more accurately model a physical system or mechanism and close the so-called reality gap for more accurate robot control. Our approach uses iterative residual tuning (IRT), a recently developed derivative-free system identification technique that utilizes neural networks and visual observation to estimate parameter differences between a proposed model and a target model. We develop several modifications to the basic IRT approach and apply it to the system identification of a five-parameter model of a marble rolling in a robot-controlled labyrinth game mechanism. We validate our technique both in simulation—where we outperform two baselines—and on a real system, where we achieve marble tracking error of 4% after just five optimization iterations.
Journal of Mechanisms and Robotics Open Issues