A new approach to monitoring nonperiodical robot operations by 3-D computer vision is proposed. In nonperiodical operations, a robot is considered to perform a variable sequence of primitive operations. The proposed monitoring system is synchronized to the robot control system to check whether or not the robot performs each primitive operation as the control system commands it to do. If not, the monitoring system will inform the control system to stop the robot. One surface point of the end effector of the robot is selected in advance as the feature for robot operation monitoring using a number of fixed and well-calibrated CCD cameras. At each sampling time instant, the 3-D data of the feature point computed by a pair of calibrated cameras is matched with the corresponding model data created in the learning stage. The Mahalanobis distance is employed to measure the statistical offset after matching the input 3-D data with the stored model. If the distance is smaller than a certain predefined threshold value, then the robot is decided to be in a proper status; otherwise it is in an improper status. Experimental results with a fast monitoring speed show the feasibility of the proposed approach.