Abstract: Abstract: In order to improve the automation degree in agriculture, reduce the workload of farmers, enhance labor efficiency and product quality and ensure that the fruit can be harvested in real time, we designed a robot system which can imitate human picking behavior and pick fruits in real time. In this paper, we first analyzed the physiological characteristics of human upper limb and built a motion model according to human upper limb movement. In the model we selected the key joints and bones as the acquisition object, while the acquisition information includes its location, angle, speed and some other information. In addition, based on the bionics characteristics, we built a motion model of the robot manipulator. The system can gather the skeleton information of the human body using the depth camera Kinect device when the human picking movement occurs. After that the system calculates the position deviation of the joint and the rotation angle of the skeleton by calculating the space vector. At the same time, the system calculates the robot arm for each joint position and rotation angle information. According to the information of instructions, the system can translate the mathematical information into many instructions and send them to the robot manipulator. The robot manipulator will follow the instruction and fulfill the task of picking in real time when instructions have been received correctly. The action system includes 3 main sub-modules: information acquisition module, instruction transition module and instruction execution module. The information acquisition module uses a depth camera Kinect device and acquires the desired three-dimensional (3D) physiological data from the human behavior. The degree of mechanical freedom of human arm is 7 while five-axis robot limb is enough to arrive anywhere human limb can arrive. So the paper built a model with 5 degrees of mechanical freedom to imitate the human picking behavior. The 3D information includes 3 kinds of physiological parameters: position, velocity and angle information. The position information can be obtained through Kinect SDK (software development kit). Based on the position data the module can calculate the vector data automatically. Additionally, through Kinect SDK the module can get 30 frames of position data per second which can support velocity calculation when the data are denoised and smoothed by a filter. The instruction transition module will compare the motion data with the data already saved in the database to find out the most similar movement instruction. The instruction transition module consists of a reverse calculation sub-module which can calculate all the location, mechanical rotating angle and speed information of the robot arm joint according to the destination position and motion model data. The mechanical data in addition to instruction will be sent to the instruction execution module. The instruction execution module in fact is a robot arm. Its mechanical arm shape is very similar to that of human arm. There are mechanical joints corresponding to human upper limb joints on the mechanical limb. The 3 linkage mechanisms can simulate the movement of human upper limbs, which can help the robot finish the picking task. In this paper, we analyzed the mechanism of human upper limb, built a mathematical model based on the bionics principle of robot, built a model of manipulator with 5 degrees of freedom, and studied the actual coordinate movement of the mathematical model of the mechanical arm in the process of fruit picking. And on the basis, we put forward a mathematical method, which is relatively simple and effective. In order to verify the validity of the model, we performed a simulation test, and the results of the test were in the acceptable range, which verified the validity of the model.