They use rubber gripper to spin each side and known algorithms to solve. The hard part is the speed of the visual recognition.
Perhaps they do only use visual feedback for determining the state, which would be a better demo, but I doubt anyone would be able know from that article!