We propose Domain-Randomized Instance Set (DRIS), a representation that simultaneously propagates a set of randomized instances to capture uncertainty in system dynamics, enabling more robust policy learning and zero-shot sim-to-real transfer. We demonstrate DRIS on a challenging reactive ball-catching task using a flat plate without mechanical stabilization, showing strong robustness to perceptual and parametric uncertainties.