Because of residency time constraints and activity time variation of cluster tools, it is very difficult to operate them. This paper addresses their real-time control issues. A Petri net (PN) model is developed to model them. This model describes the robot activity sequence with robot waits included. Via this model, to operate a cluster tool is to determine robot wait times and a two-level operation architecture is proposed, which includes an off-line periodic scheduler and real-time controller. By this approach, the system can adapt to random activity time variation. A cluster tool derived as a not-always- schedulable system by the existing methods is shown always-schedulable by using the proposed innovative approach.