Optimized UKF-based perception of a repetitive dynamic event

Living beings are rarely passive observers. Instead, they are characterized by a "Perception-Action Coupling", i.e. they perceive in order to move and they move in order to perceive, which constitutes an essential loop for learning. Drawing inspiration from nature, this work proposes a novel method for minimizing the uncertainty of the information gathered when observing a Repetitive Dynamic Event (RDE), considering a sensor that can change its position and orientation. The method considers an Unscented Kalman Filter (UKF) for fusing the sensor’s measurements with the evolution of the current dynamical model of the observed RDE. The proposed method aims at finding the optimum UKF parameters and pose of the sensor, for maximizing the quality of the estimation. Simulation and experimental evaluations show the improvement achieved by the proposed method, as compared to using a non-optimized UKF, in terms of the quality of acquired information of a dynamic event. The experiments are conducted using a UR10e robotic manipulator with an eye-in-hand ZED 2 RGB-D camera.