We target the difficult problem of catching in-flight objects with uneven shapes. This requires the solution of three complex problems: predicting accurately the trajectory of fast-moving objects, predicting the feasible catching configuration, and planning the arm motion, all within milliseconds. We follow a programming-by-demonstration approach in order to learn models of the object and the arm dynamics from throwing examples. We propose a new methodology for finding a feasible catching configuration in a probabilistic manner. We leverage the strength of dynamical systems for encoding motion from several demonstrations. This enables fast and online adaptation of the arm motion in the presence of sensor uncertainty. We validate the approach in simulation with the iCub humanoid robot and in real-world experiment with the KUKA LWR 4+ (7-DOF arm robot) for catching a hammer, a tennis racket, an empty bottle, a partially filled bottle and a cardboard box.
