Image recognition is one of them, either with camera on the robot (CV combined with mapping), or cameras in rooms monitoring location of the robot.
Another option would be to use dead reckoning technique, but without any external references error will build up. Maybe RF tags would help with that.