Difficulties lie in the study and application of fruit-picking robots. In particular, large random positioning errors can occur due to disturbance, which is difficult to compensate for using control methods. Existing end-effectors cannot conduct fault tolerance for these errors and are not widely applied. Therefore, a limited universal and fault-tolerant end-effector was designed to address binocular vision-positioning errors. The theory underlying the design of the new mechanism was also proposed based on institutions and vision positioning, in which random positioning errors are regarded as systematic “fault errors†of the end-effector, enabling the mechanism to complete operations within the error range. The relationship between the gripper and the cutter was modeled, and a mathematical model of error tolerance was established. Moreover, a new limited universal end-effector was designed. The binocular vision-positioning errors (including original and random positioning errors) were analyzed, and static and dynamic positioning experiments were conducted using a mechanism and a vision-positioning experimental platform based on binocular vision. The maximum positioning errors obtained were 60.1 mm in the z-direction and 17.2 mm in the x-direction, which were within the fault-tolerance range. Moreover, indoor and outdoor picking experiments were conducted for litchi and citrus using the picking manipulator based on binocular vision. The picking success rates were over 84% and 78% in indoor and outdoor tests, respectively. Finally, the favorable fault-tolerant capacity of the end-effector was validated with a comparison experiment that showed that the limited universal picking manipulator based on binocular vision could achieve litchi and citrus pickings within an acceptable error range. The results verified the applicability of the fault-tolerant design.