An integrated system for robot grasping of nonrigid, unknown, and fragile objects in cluttered environments is presented. The system is based on three levels of grasp control. In the low level, reactive control policies are used to control contact forces and locations using tactile, kinematic, and load force feedback. The sensor data is also used to determine grasp and object properties that are used by the higher levels to control the grasp. In the middle level, a grasp representation, called a grasp mode, is used to generically represent the grasp/object configuration. Grasp modes are defined in terms of the Grasp Image Space (GIS), a space of grasps constructed through object exploration and simulation that includes information regarding hand configuration, applied forces, and object properties. The use of grasp modes allows for the abstraction of the grasp. In the high level, the mechanism for grasp mode switching and behaviour-based action is developed. Switching takes place by searching the GIS for a grasp mode that is likely to succeed, given the current knowledge of the grasp/object properties. The system was tested experimentally and in simulation. The experimental tests focused on handling uncertainties in object information and location and on grasping fragile objects without damage. An experiment was performed to handle unknown changes in object friction by switching between multiple control policies. The results showed a 30% grasp success rate without switching and a 100% success rate with switching. Experiments were also performed to grasp objects in cluttered environments despite having no object knowledge. A Naive Bayes Classifier was used to learn switching rules to handle this scenario, and a grasp success rate of 95% was recorded.