In the future development of orchard industry, intelligent orchard is an important development trend. In order to realize the intelligent orchard, intelligent fruit picking is one of the key bottlenecks. In order to achieve the goal of intelligent fruit picking, a citrus picking robot system suitable for hilly dwarf cultivation of fruit trees was built. Aiming at the uneven ground between ridges in hilly orchards and the terrain inclination angle of 0°~20°, an adaptive leveling platform was designed to keep the base level of the manipulator. The visual system used the depth camera to obtain the point cloud image to establish the three-dimensional model of the fruit tree, and realize the acquisition of the position information of the fruit to be picked. In order to avoid damage to the fruit during the picking process, an integrated end-effector for shearing and clamping was designed, which protected the fruit from damage while picking the fruit by shearing and clamping. The main interference factors in the orchard in the natural environment were light and wind. The light and wind were graded, and 10 foot control experiments were set up. The results showed that under low light or normal light conditions, the average fruit positioning accuracy was 82.5%, the end-effector clamping success rate was 87.5%, and the average fruit picking time was 12.3 s/piece.The average fruit positioning accuracy under high light conditions was 72%, the success rate of end-effector clamping was 80%, and the average time of fruit picking was 12.5 s/piece. The research result can provide a reference for the study of fruit picking in hilly terrain.